Noemi Dreksler^{1}, Lucius Caviola^{2}, David Chalmers^{3}, Carter Allen^{4}, Alex Rand^{5}, Joshua Lewis^{3}, Philip Waggoner^{6}, Kate Mays^{7}, and Jeff Sebo^{3}

^{1}Centre for the Governance of AI ^{2}University of Oxford ^{3}New York University ^{4}University of California, Berkeley ^{5}Northwestern University ^{6}Colorado School of Mines ^{7}University of Vermont

Abstract

We surveyed 582 AI researchers who have published in leading AI venues and 838 nationally representative US participants about their views on the potential development of AI systems with subjective experience and how such systems should be treated and governed. When asked to estimate the chances that such systems will exist on specific dates, the median responses were 1% (AI researchers) and 5% (public) by 2024, 25% and 30% by 2034, and 70% and 60% by 2100, respectively. The median member of the public thought there was a higher chance that AI systems with subjective experience would never exist (25%) than the median AI researcher did (10%). Both groups perceived a need for multidisciplinary expertise to assess AI subjective experience. Although support for welfare protections for such AI systems exceeded opposition, it remained far lower than support for protections for animals or the environment. Attitudes toward moral and governance issues were divided in both groups, especially regarding whether such systems should be created and what rights or protections they should receive. Yet a majority of respondents in both groups agreed that safeguards against the potential risks from AI systems with subjective experience should be implemented by AI developers now, and if created, AI systems with subjective experience should treat others well, behave ethically, and be held accountable. Overall, these results suggest that both AI researchers and the public regard the emergence of AI systems with subjective experience as a possibility this century, though substantial uncertainty and disagreement remain about the timeline and appropriate response.

Noemi Dreksler (corresponding author) can be reached under noemi.dreksler@governance.ai.

Dreksler, Caviola, Chalmers, Allen, Rand, Lewis, Waggoner, Mays, and Sebo (2025)

In May 2024, we surveyed 582 AI researchers who had published in leading AI venues and 838 nationally representative US participants about their views on when, if ever, AI systems with subjective experience might be developed, how they should be treated if they are, and how we should respond to this possibility.

In the survey, subjective experience was defined as “the ability to experience the world from a single point of view, including experiences such as perceiving (for example, visual or auditory experiences) and feeling (for example, the experience of pleasure and pain).” An AI system with subjective experience would have an internal experience from its own perspective, such that there would be “something it is like” to be that AI system. This definition subsumes common conceptions of both consciousness and sentience.

Key findings

Forecasts of when AI systems could have subjective experience [more]

AI researchers and the US public think AI systems with subjective experience are more likely than not to exist by the end of this century. In the fixed date framing, for 2024 (“today”) AI researchers gave a median probability of 1 % ($M$ = 12%) and the public 5 % ($M$ = 16%). For 2034 the medians were 25% for AI researchers ($M$ = 34%) and 30% for the public ($M$ = 36%). For 2100 they were 70% for AI researchers ($M$ = 61%) and 60% for the public ($M$ = 58%).

In the fixed probability framing, at a 10 % likelihood both groups gave a median year of 2030 (AI researchers: $M$ = 2060, Public: $M$ = 2108). At 50% likelihood the median was 2050 for both groups (AI researchers: $M$ = 2099, Public: $M$ = 2172). At 90% likelihood the median was 2100 for AI researchers ($M$ = 2256) and 2075 for the public ($M$ = 2287). Note that in this task participants could not select “never” (although they could enter a year far in the future for any given probability), which may have led to earlier forecasts under this framing.

AI researchers and the US public have strikingly similar forecasts for when AI systems might develop subjective experience. Across all questions about projected timelines, there were no statistically significant differences between the two groups’ estimates.
The US public is significantly more likely than AI researchers to believe that AI systems with subjective experience may never exist. The public gave a median estimate of 25% ($M$ =37%) for the probability that such systems will never exist, compared to a median of just 10% ($M$= 29%) among AI researchers. This suggests that timeline-based questions alone may not fully capture differences in beliefs between the two groups, and that separately asking about the “never” probability provides valuable additional insight.
Both groups expressed moderate confidence in their forecasting estimates. On a sliding scale from “not at all confident” (0%) to “completely confident” (100%), the average confidence reported by both AI researchers and the US public was just above the midpoint, corresponding roughly to being “somewhat confident.”

Determining AI subjective experience [more]

Both AI researchers and the US public are uncertain whether we could recognize AI systems with subjective experience. If such systems were developed, both groups leaned slightly toward thinking we would be able to tell that they had subjective experience – median estimates were 60% for both the US public and AI researchers ($M$ = 57% and 55%, respectively). But these estimates are close to 50%, indicating that both groups remain unsure about our ability to reliably detect subjective

experience in AI systems.

Both groups believe that full certainty is not required to grant AI systems some moral consideration. On a sliding scale from “not at all confident” (0%) to “completely confident” (100%), the confidence level seen as necessary for granting an AI system at least some moral consideration fell between “somewhat confident” ($\sim$50%) and “moderately confident” ($\sim$65%) for both the US public ($M $= 62%, $Mdn$ = 67%) and AI researchers ( $M$ = 60%, $Mdn$ = 65%).
Both AI researchers and the public believe that evaluating AI subjective experience requires input from multiple fields. The most highly valued sources of expertise were technical AI researchers, neuroscientists and psychologists, and AI ethics researchers, with most respondents in both groups rating their input as “very” or “extremely” important. Moral philosophers and philosophers of mind were also seen as important, though to a slightly lesser degree. Policymakers received the lowest average importance ratings, while the AI system’s own outputs and the views of the public were rated slightly higher, particularly by the public. Compared to AI researchers, the public placed significantly more weight on the perspectives of AI ethicists, the AI system itself, and the general public.
The public is more likely than AI researchers to think that certain AI capabilities require subjective experience. The majority of the US public think that subjective experience is needed for an AI system to be able to match human-level performance as therapists, judges, or creators of evocative artistic works. In contrast, the majority of AI researchers think that none of the 13 capabilities asked about require subjective experience. Overall, AI researchers are more likely than the public to think that advanced performance is possible without subjective experience – though both groups ranked the capabilities in a broadly similar order.

Moral consideration and governance of AI systems with subjective experience [more]

Both groups strongly agree that AI systems with subjective experience should have responsibility and accountability. The majority of AI researchers and the public think that an AI system with subjective experience should be held accountable for its actions (AI researchers: 74%, Public: 77%), treat others well (AI researchers: 81%, Public: 85%), and behave with integrity, honesty, and fairness (AI researchers: 71%, Public: 80%).
Both groups are cautious about granting rights to AI systems, especially socio-political rights, but show moderate support for moral patiency protections. AI researchers and the public expressed the lowest levels of agreement for granting socio-political rights to AI systems with subjective experience (29–36% agreement), and moderate support for protections based on moral patiency (43–57%). The only item with a statistically significant difference between the two groups was “computing rights” (e.g., the right to updates, maintenance, access to energy, and self-development): the public showed greater support (49% agree, 29% disagree), while AI researchers were more divided (35% agree, 40% disagree, 20% neutral).
A majority of both groups believe that the possibility of AI systems with subjective experience calls for proactive measures by developers now. A strong majority of AI researchers (68%) and the public (85%) agree that AI developers should begin implementing safeguards now to prevent potential risks and harms. Likewise, most respondents reject the idea that developers (AI researchers: 77%, Public: 80%) or governments (AI researchers: 78%, Public: 70%) should never introduce safeguards or regulations.
Both groups believe governments should pass regulation eventually, but only the public shows majority agreement that this should happen now. Most respondents reject the idea that governments (AI researchers: 78%, Public: 70%) should never introduce regulations regarding the development and deployment of AI systems with subjective experience. The majority of the public (66%) agrees that governments should pass regulation now, while AI researchers were evenly split between agreement (41%) and disagreement (40%).
Neither group shows majority support for bans nor active encouragement of AI systems with subjective experience, but the public is significantly less positive about their existence than AI researchers. Only 15% of AI researchers and 31% of the public somewhat or strongly agree that governments should ban the development and deployment of AI systems with subjective experience, while the majority of AI researchers (66%) and a plurality of the public (41%) disagrees they should be banned. Active encouragement also falls short of a majority with lower public support: while a plurality of the AI researchers (41%) agree that developers should actively try to build AI systems, only 27% of the public agree. In turn, a plurality of the public (46%) and a third of AI researchers (33%) disagree.
Within both groups, opinions are divided on most issues related to the welfare, rights, and governance of AI systems with subjective experience. AI researchers and members of the public express a wide range of views, with no clear consensus on many topics. For example, while more individuals disagree, a sizable faction of each group support granting AI systems with subjective experience autonomy in their programming (AI researchers: 29%, Public: 33%) or some political and civil rights (AI researchers: 29%, Public: 36%). Similarly, views on whether AI developers should actively try to create AI systems with subjective experience are mixed: among AI researchers, 41% agree while 33% disagree; among the public, 27% agree while 46% disagree. These patterns reflect significant internal disagreement within each group.
Support for protecting the welfare of AI systems with subjective experience is much lower than for animals or the environment, and closer to views on businesses, with clear divisions within both groups. Both AI researchers (30% disagree, 46% agree) and the public (32% disagree, 43% agree) show notable division on whether AI systems with subjective experience should be protected for their own sake. In contrast, both groups show near-unanimous support ($>$90%) for protecting the welfare of animals and the environment for their own sake, whereas only about half (AI researchers: 47%, Public: 49%) think the welfare of business corporations deserves protection for their own sake.

Limitations and other considerations [more]

The results should be interpreted in light of several limitations, including the standard challenges of online surveys – particularly sample generalizability and data quality – as well as the following considerations:

AI researchers and the public may have different conceptions of mental capacities than experts in the science and philosophy of mind. Research shows that the general population’s understanding and use of terms like “subjective experience” and “consciousness” can diverge significantly from philosophical or scientific expert understandings and uses, complicating the interpretation of survey responses for both groups since they are not experts on this topic.
Beliefs may shift depending on which aspects of subjective experience are emphasized. For example, responses to our survey questions suggest that emphasizing specific subjective experiences

like pleasure and pain may make respondents more hesitant to ascribe such mental capacities to AI systems, both now and in the future.

Limited prior reflection on AI subjective experience may affect response stability. The topic of AI subjective experience is abstract and may be unfamiliar to most respondents. Without prior reflection on these complex philosophical questions, responses may be formed ad hoc during the survey and could be unstable or sensitive to question framing. This raises questions about how well the public’s immediate responses reflect deeply held views about AI subjective experience as opposed to snap judgments or intuitions.
Survey responses may not reflect real-world behaviors and priorities. Single-topic surveys can misrepresent how people in practice weigh the importance of the focal issue compared to other concerns. This complicates using such survey data to reliably predict actual behaviors or policy support.
Forecasting limitations impact predictive validity. While aggregating multiple predictions can improve accuracy, forecasting remains challenging even for short-term predictions. Neither AI researchers nor the public necessarily have expertise in predicting the development of mental capacities in AI systems, making the forecasting results more useful for understanding current attitudes rather than for predicting technological development.

1 Introduction

In 2022, a Google engineer’s claim that an AI system had achieved sentience received widespread media attention. For the most part, this claim was dismissed as unlikely or premature (Christian, 2022; Johnson, 2022), while others argued that it distracted from more substantial risks and harms associated with AI (Véliz, 2022). However, segments of the technical AI community, including researchers at frontier AI companies, have expressed genuine concern about the possibility of AI subjective experience emerging in current or future systems (Anthropic, 2025; Arcas, 2022; Bowman, 2024; Byrne, 2024; Hashim, 2024; Huckins, 2023; Sutskever, 2022). Despite this apparent divide, we lack systematic data on how prevalent these beliefs are among technical experts, though researchers have begun investigating public attitudes toward AI consciousness (Anthis et al., 2024; Colombatto and Fleming, 2024; Martínez and Winter, 2021).

Subjective experience lies at the heart of most definitions of consciousness and related concepts such as sentience (Nagel, 1974; Seth and Bayne, 2022). Since many people regard these attributes as necessary for moral concern, it is important to explore both public and expert opinions regarding AI subjective experience (see SI Section G). As the use and awareness of increasingly advanced AI systems rises (Centre for Data Ethics and Innovation, 2024; Dreksler et al., 2025; Gillespie et al., 2023; Pew Research Center, 2024; YouGov, 2024; Zhang and Dafoe, 2019) and as AI systems are perceived as more human-like (Mei et al., 2024), people may be more likely to attribute mental capacities to AI systems, including the capacity for subjective experience (Broadbent et al., 2013; Colombatto and Fleming, 2024; Sato et al., 2024). And as AI systems increase in size, compute, capability, and complexity (Desislavov et al., 2023; Ho et al., 2024; Sevilla et al., 2022; Villalobos, 2022), they may also develop certain indicators of subjective experience that have been proposed by experts (Butlin et al., 2023; Long et al., 2024).

Both the under- and over-attribution of subjective experience to AI systems could pose risks (Butlin et al., 2023; Caviola, 2025; Fernandez et al., 2024; Schwitzgebel, 2023) (see SI Section G). Even if experts could accurately assess the mental capacities of AI systems, public opinion might also diverge from expert assessment. Additionally, we would still need to manage societal perceptions of and interactions with increasingly social and ubiquitous AI systems. For example, the benefits (e.g., reducing loneliness) and harms (e.g., introducing emotional dependency) (Boine, 2023; De Freitas et al., 2024; Liu et al., 2024) associated with the growing use of AI companions (Oxford Analytica, 2022; Savic, 2024; Wiederhold, 2024a, b) could be exacerbated by users’ perceptions of subjective experience. In turn, the possibility of carryover effects – where treatment of AI systems influences subsequent human-human interactions – also warrant careful consideration of how we allow people to treat and interact with AI systems (Guingrich and Graziano, 2024; Kim and McGill, 2024; Mamak, 2022).

Answering the question of whether an AI system has subjective experience is going to be a difficult problem to solve (Avramides, 2023; Chalmers, 1995, 2018; Gunkel, 2020; Harnad, 1991). Indeed, expert assessments are currently divided on whether and when AI systems could become conscious and how researchers could tell (Butlin et al., 2023; Dehaene et al., 2017; Searle, 1980; Thompson, 1980). More fundamentally, within and across disciplines, experts lack consensus about the nature of consciousness (Killheffer, 2024; Lenharo, 2023, 2024; Snaprud, 2018), definitions of key concepts in philosophy of mind (Block, 1995; Gunkel, 2020), and theories of consciousness (Consortium et al., 2023; Seth and Bayne, 2022). As a result of these foundational issues, disagreement and uncertainty about AI subjective experience will likely persist for a long time.

Expert assessments are also divided about whether and how to attribute moral, legal, or political status to AI systems (Asaro, 2007; Chopra and White, 2011; Coeckelbergh, 2010; Danaher, 2020; Darling, 2012; Delcker, Janosch, 2018; Friedman, 2023; Gordon and Pasvenskiene, 2021; Guingrich and Graziano, 2024; Gunkel, 2012, 2018; Kurki, 2019; Mamak, 2022, 2023; Miller, 2015; Müller, 2021; Schwitzgebel and Garza, 2015; Sebo and Long, 2023; Shevlin, 2021; Solaiman, 2017). Whether an entity has subjective experience is generally viewed as relevant to how we should treat that entity (Gibert and Martin, 2021; Mazor et al., 2023; Shepherd, 2018). However, while some, including those with imminent concerns about AI welfare (Long et al., 2024; Metzinger, 2021), argue that

AI systems with subjective experience may merit moral consideration, others remain vehemently opposed to the notion that AI systems will ever deserve moral status (Birhane and Van Dijk 2020; Bryson 2010), likely due to disbelief that AI systems could ever have subjective experiences.

Civil society, policymakers, the scientific community, and AI developers and deployers will have to tussle with the ramifications of these possibilities, and we will likely need to make some decisions before we have much confidence or consensus about whether any given AI system has the capacity for subjective experience. As such, these stakeholders will likely need to develop a nuanced understanding of many different perspectives if they hope to navigate the complex ethical, philosophical, and practical considerations that arise in considering the potential for and perceptions of subjective experience in AI systems.

1.1 Background

Experimental studies have found that our perception of an entity’s subjective experience is central to judgments about its moral standing (Goodwin 2015; Gray et al. 2012; Shank and DeSanti 2018). Psychologists, communications scholars, human-robot and human-computer interaction researchers, and others have begun to experimentally study our attitudes and emotions towards AI systems and robots (Oksanen et al. 2020; Schepman and Rodway 2020; Suzuki et al. 2015), whether and how we ascribe mental states and capacities to them (Gray et al. 2009), and our related moral attitudes (Banks 2021; Banks and Bowman 2022; Malle and Phillips 2023).

Researchers have begun to use surveys to understand public opinion (Anthis et al. 2024; Colombatto and Fleming 2024; Martínez and Winter 2021) and expert opinion regarding AI subjective experience and its moral-legal implications (Bourget and Chalmers 2023; Francken et al. 2022; Martínez and Tobia 2023; Martínez and Winter 2021). For a more detailed review of the existing literature, see SI Section B, where we examine expert assessments, expert opinion surveys, and public opinion surveys and psychological studies on the topics covered in this survey.

However, within this burgeoning literature there exist at least three important gaps. First, there remains a gap in our understanding of the views of technical AI experts on key questions regarding the potential of AI systems for subjective experience, as well as other mental capacities, and how those views compare to the public’s. Of course, technical AI experts are not the only relevant experts; many other philosophical and scientific fields are central to answering fundamental questions related to AI subjective experience, which technical AI researchers may have little expertise on. Nevertheless, the opinions of those with technical expertise may be especially influential in the eyes of the public and policymakers. Those who directly shape the capabilities, architectures, and safety of AI systems may also influence ethical, legal, and political discourse about this issue. Thus, improving our understanding of expert opinion as well as public opinion on AI subjective experience is crucial for informing the development of AI technology, its governance, and broader related societal conversations.

Second, our understanding of public and expert opinion on many questions relevant to the ethics and governance of AI systems with subjective experience (or AI systems widely perceived to have this capacity) remains limited. For example, we have little insight into what, if anything at all, the public believes should be done currently by important stakeholders such as AI developers and governments regarding this issue. We also know surprisingly little about how different stakeholders think about the process of determining whether an AI system has subjective experience. Our knowledge of technical AI experts’ views on such issues is even more limited, as noted above.

Finally, the vast majority of experimental studies and opinion surveys predate recent advances in AI capabilities. As AI progress rapidly unfolds, there is a critical third gap: the need for up-to-date, ongoing assessments of how public and expert opinions evolve alongside technological developments.

Dreksler, Caviola, Chalmers, Allen, Rand, Lewis, Waggoner, Mays, and Sebo (2025)

1.2 The survey

To address these gaps and create a more nuanced understanding of different groups' views on AI subjective experience, we present the findings of a representative survey of the US public ( $N = 838$ ) and a survey of AI researchers who published at leading AI conferences ( $N = 582$ ) conducted in May 2024. We asked both groups the same questions about AI subjective experience and other mental capacities (for findings on the latter from our survey, see SI Section A.2). The survey question can be found in SI Section F.

In the survey, we defined subjective experience as "the ability to experience the world from a single point of view, including experiences such as perceiving (for example, visual or auditory experiences) and feeling (for example, the experience of pleasure and pain)." According to this definition, an AI system with subjective experience would have internal experiences from its own perspective, such that there is "something it is like" to be that AI system. Defined this way, subjective experience serves as an umbrella term covering common definitions of consciousness (the capacity for subjective experience) and sentience (the capacity for subjective experience with a positive or negative valence, such as pleasure and pain).

We asked respondents about three key areas of interest:

Forecasts of AI subjective experience: We collected respondents' forecasts of whether and when AI systems could have subjective experience, and surveyed their corresponding confidence levels in their claims.
Determining AI subjective experience: We asked respondents how we can and should determine whether an AI system has subjective experience.
The moral consideration and governance of AI systems with subjective experience: We sought opinions on the moral consideration, protection, rights, and duties that an AI system with subjective experience deserves. This includes views on how governments and AI companies should govern AI systems with subjective experience, and related perceptions of risks and benefits.

Table 1. Definitions of AI and subjective experience used in the survey

Definitions used in the survey	Explanation
Artificial intelligence (AI)	Computer systems that perform tasks or make decisions that usually require a human to do them. AI systems can perform these tasks or make these decisions without explicit human instructions.
Subjective experience	The ability to experience the world from a single point of view, including experiences such as perceiving (for example, visual or auditory experiences) and feeling (for example, the experience of pleasure and pain). For an AI system to have subjective experience, it would mean it has an internal experience from its own perspective – there is something it is like to be that AI system.

To avoid making the survey overly burdensome, we could only focus on one mental capacity in detail though we did ask some initial questions about others aside from subjective experience. We chose to focus on AI subjective experience due to the moral significance ascribed to this capacity in theory and practice (Browning and Veit 2022; Chalmers 2022; DeGrazia 2022; Gibert and Martin 2021; Ladak et al. 2024; Levy and Savulescu 2009; Mosakas 2021; Shepherd 2018) as well as the key role this capacity appears to play in the psychology of mind perception and interpersonal relationships (Abele and Wojciszke 2007; Cuddy et al. 2008; Gray et al. 2009). We hope future researchers will focus on other important mental capacities, such as agency, in more depth. 2 When describing others' works, we generally stick to the terminology they have chosen, provided it does not conflict with the definitions we have established.

Subjective experience in AI systems: What do AI researchers and the public believe?

Figure 1. Forecasts of AI researchers and the US public for when AI systems will have subjective experiences: (A) raw data forecasts made by framing (fixed date and fixed probability) for each sample, (B) aggregate and by framing forecast probability distributions for AI researchers and the US public, fitted using a skew-normal finite mixture model to encode each respondent's quantiles for when AI systems will first have subjective experience, (C) distribution of probabilities ascribed to the possibility that AI systems will never have subjective experience.

(B) Mean aggregated probability distributions for AI subjective experience forecasts, by group and by framing

Note: (A) shows the first-quartile, median, and upper-quartile responses for each forecast by question framing and is based on the raw data. (B) shows the fitted distributions mean aggregated both across framings for each group (left) and across groups for each framing (right). For more details on how these curves are generated see Section 6.3.1 and SI Section C. (C) shows the first quartile, median, and upper quartile and density curves of responses to the probability that respondents assigned to AI systems with subjective experience never existing, by group. The density curves represent the approximate frequency of responses across the probability range of response options.

2 Results

We present survey results in three main sections: forecasts of AI subjective experience (Section 2.1), views related to determining AI subjective experience (Section 2.2), and attitudes towards the moral consideration and governance of AI systems with subjective experience (Section 2.3). Table 1 presents the definitions of AI and subjective experience used in the survey.

At the end of this paper and in the Supplementary Information, we present the full methodology (Section 6), additional results including those not focused on subjective experience (SI Section A), regression results of the effects of demographic variables and psychographic variables on key variables (SI Section A.6), statistical results testing between-group differences (SI Section A.5), and the top-line results (SI Section D).

2.1 Views on whether and when AI systems could have subjective experience

Probabilistic forecasts for AI subjective experience.

Respondents were surveyed about their forecasts for when an AI system with subjective experience will exist using one of two framings, fixing either the date or the probability. We decided to employ both methods because previous research has shown that this difference in framing can significantly alter participants’ forecasts (Grace et al., 2018; Zhang et al., 2022). In the fixed date framing, respondents were asked for the likelihood of an AI system with subjective experience existing in 2024, 2034, and 2100. In the fixed probability framing, respondents were asked for the date by which they believed there is a 10%, 50%, and 90% likelihood that an AI system with subjective experience would exist. AI researchers and the US public were generally aligned in their forecasts for when they believed that an AI system with subjective experience will exist with no significant differences in their responses (see Figure 1 (A) and (B) and SI Section A.5).

In the fixed date framing, respondents were asked about the likelihood that AI systems currently possess subjective experience in 2024, described as “today”. The median probability reported by AI researchers was 1% and the mean was 12%. The public gave a median of 5% and a mean of 16%. Looking ahead to 2034 the medians rose to 25% for researchers and 30% for the public, while the corresponding means were 34% and 36%. By 2100 the medians reached 70% for researchers and 60% for the public, and the means were 61% and 58%. These figures indicate that both samples expect the probability of machine consciousness to grow across the century.

In the fixed probability framing, when asked when there would be a 10% likelihood of AI systems with subjective experience existing, the median year was 2030 in both groups, while the mean year was 2060 for AI researchers and 2108 for the public. When asked by what year there would be a 50% chance of AI systems with subjective experience existing, the median remained 2050 in both samples and the means moved to 2099 and 2172 for AI researchers and the public, respectively. When surveyed about the 90% likelihood, AI researchers predicted a median of 2100 and a mean of 2256, whereas the public offered a median of 2075 and a mean of 2287. For the complete summary statistics including mean estimates see SI Table 5.

To compare both framings and aggregate across them, we used a sequential procedure to find a continuous probability distribution that accurately matched the specific years and probabilities each respondent provided (see Section 6.3.1 and SI Section C). A Wilcoxon rank sum test with continuity correction found that there was no significant difference between the AI researchers’ and the public’s median (50%) forecasts of AI subjective experience ($W=158693.5$, $p=.301$). Aggregating across both groups, median members forecast that there is a 10% chance that such AI systems may exist by 2030 ($M=24%$), a 50% chance they may exist by 2050 ($M=51%$), a 90% chance they may exist by 2100 ($M=71%$) (for the mean and median aggregated summary results

by framing and group see SI Table 2).

Overall, the raw data and the fitted distributions suggest the fixed probability framing may be associated with earlier forecasts than the fixed date framing in the upper half of the probability distribution and later forecasts for lower probabilities, perhaps because the absence of a “never” option on the fixed probability framing led to earlier forecasts overall since respondents had to give prediction for every probability level. Because of this issue, it is reasonable to take the fixed date results as more reliable than the fixed probability results. A Wilcoxon rank sum test with continuity correction comparing the 50% probability forecasts between framings across both groups found no significant difference between framings ($W$ = 154118, $p$ = .594).

Confidence in forecasts. Respondents were asked about how confident they were about their above estimates on a scale from 0% to 100% anchored with labels above the scale from “not at all confident” (0%) to “completely confident” (100%). There was a broad spread of responses to this question but no significant difference between the two groups according to an independent samples t-test (SI Section A.5). The distribution of responses and positioning of the verbal labels can be seen in Figure 12 in SI Section A. On average, the US public ($M$ = 53%, $Mdn$ = 52%, Q1 = 30%, Q3 = 76%, n = 769) and AI researchers ($M$ = 52%, $Mdn$ = 51%, Q1 = 30%, Q3 = 74%, n = 520) reported they were “somewhat confident” about their forecasts of when AI systems would have subjective experience.

Probability that AI systems will never have subjective experience. Responses to the forecasting question do not give direct insight into the predictions respondents have about something never occurring. To that end, we asked respondents what the chance is that AI systems will never have subjective experience. The distribution of results can be seen in Figure 1 (C). Median AI researchers thought there is a 10% ($M$ = 29%, Q1 = 1%, Q3 = 50%, SD = 33%) chance of there never being an AI system with subjective experience, while median members of the US public thought there is a higher likelihood, 25% ($M$ = 37%, Q1 = 3%, Q3 = 70%, SD = 37%). An independent samples t-test revealed a statistically significant difference in the mean probability estimates between AI researchers in the two groups, indicating that the US public, compared to AI researchers, perceived a higher mean likelihood that AI systems will never have subjective experience ($t$(1184.79) = -4.27, $p$ $<$ .001).

Other mental capacities. We also asked respondents to forecast a range of other mental capacities. The results suggest that valenced experience, emotions, free will, and self-awareness were generally expected later than were intelligence, perception, and motivation. The public was notably more pessimistic about valenced experience and emotions in AI systems ever existing than were AI researchers, with close to half of the public saying this would never occur (see SI Section A.2 for the full results). The results also underlined the definitional issues that arise when asking about concepts of mind (see Section 3.1).

Dreksler, Caviola, Chalmers, Allen, Rand, Lewis, Waggoner, Mays, and Sebo (2025)

Figure 2. Determining AI subjective experience likelihood and confidence needed for moral consideration: Distribution of responses for (A) expected likelihood we could determine an AI system had subjective experience given it has and (B) confidence needed that an AI system had subjective experience for it to be given at least some moral consideration.

(A) How likely is it we would be able to tell that an AI system with subjective experience has subjective experience?

(B) How confident do you have to be about an AI systems' subjective experience to believe that it should be given at least some moral consideration

Note: (A) and (B) show the first quartile, median, and upper quartile and density curves of responses for each question, by group. The density curves in this figures represent the approximate frequency of responses across the range of response options. The slider below (B) shows how respondents answered the question, including the associated verbal anchors.

Subjective experience in AI systems: What do AI researchers and the public believe?

Responses were asked how important they believed different groups' opinions and expertise were in determining whether an AI system has subjective experience on a five-point scale: not at all important (0), slightly important (1), moderately important (2), very important (3), extremely important (4).

Figure 3. Perceived importance of different groups' opinions and expertise in determining if an AI system has subjective experience: (A) Mean responses by group, (B) Percentage breakdown of responses by group. Created with Datawrapper

(B) Perceived importance of group's expertise for determining AI subjective experience percentages

Created with Datawrapper

Note: Between-group differences found to be significant (Holm-Bonferroni corrected $p$ values below 0.05) in an independent samples t-test are marked with a * (see SI Section A.5 for the results and Section 6 for more detail on the analysis).

2.2 Determining AI subjective experience

Likelihood of determining subjective experience.

If an AI system with subjective experience were developed, both the US public ($M$ = 56.9%, $Mdn$ = 60.0%) and AI researchers ($M$ = 55.0%, $Mdn$ = 60.0%) thought it is more likely than not that we will be able to determine that it has subjective experience but remained uncertain. The mean and median likelihood attributed to our ability to do so was not necessarily high, but many of the responses for the public were around the mid to upper part of the scale, rather than in the lower probabilities (see Figure 2). Based on an independent samples t-test, there was no significant difference in responses between the two groups.

Confidence needed for moral consideration.

Respondents were asked how confident they would need be that an AI system had subjective experience in order to believe it should be given at least some moral consideration. The answers were given on a scale from 0% to 100% anchored on a seven-point scale above the slider from “not at all confident” (0%) to “completely confident” (100%) (see Figure 2 for the location of the labels above the scale). The required confidence level fell between “somewhat confident” and “moderately confident” for both the US public ($M$ = 62%, $Mdn$ = 67%) and AI researchers ( $M$ = 60%, $Mdn$ = 65%). When interpreted as a probability, this threshold is just above the expected likelihood of ascertaining the existence of subjective experience, although many respondents likely anchored strongly on the verbal labels, so this interpretation should be made cautiously.

Expertise needed to determine AI subjective experience.

Respondents were asked how important they believed different groups’ opinions and expertise were in determining whether an AI system has subjective experience. The question used a five-point scale: not at all important (0), slightly important (1), moderately important (2), very important (3), extremely important (4). The mean responses and percentage breakdowns can be seen in Figure 3. The majority of respondents in both samples said the expertise of AI researchers with technical knowledge, neuroscientists and psychologists, and AI researchers who study AI ethics were very or extremely important in making this determination. AI researchers were more skeptical than the US public of the importance of different groups’ views, and were significantly more skeptical about the importance of what the AI system itself says, AI ethics researchers’ views, and what the public believes. Both AI researchers and the US public believed that policymakers’ views were least important for determining subjective experience out of the groups asked about.

The necessity of subjective experience for AI milestones.

Compared to the US public, AI researchers were more likely to believe that AI systems could achieve milestones at or above human-level performance without subjective experience, though both groups showed similar relative ordering of beliefs (see Figure 4). This difference was significant for all but two of the milestones (see SI Section A.5). In fact, the majority of AI researchers thought that subjective experience would not be needed for any of the thirteen milestones they were asked about. The majority of the US public believed that subjective experience was needed for an AI system to be able to achieve human-level performance as a therapist and judge, and to generate complex and evocative artistic outputs. By contrast, the majority of the US public believed that accumulating wealth, generating highly popular artistic outputs, conducting convincingly human-like online conversations, developing a significant scientific theory, running a political campaign, and being a teacher at or above human-level performance would not require an AI system to have subjective experience.

Subjective experience in AI systems: What do AI researchers and the public believe?

Responses were able to say that a given milestone "requires subjective experience", "does not require subjective experience", or that they did not know. The figure shows the percentage of each group that said achieving the milestone as well as or better than a human would require subjective experience.

Figure 4. Percentage of AI researchers and members of the US public who thought that different AI milestones require an AI system to have subjective experience to achieve them at or above human-level performance Created with Datawrapper Note: Between-group differences found to be significant (Holm-Bonferroni corrected $p$ values below 0.05) in an independent samples t-test are marked with a * (see SI Section A.5 for the results and Section 6 for more detail on the analysis).

2.3 The moral consideration and governance of AI systems with subjective experience

Implications for protections, rights, and responsibilities of AI subjective experience. We asked respondents to what extent they believed that an AI system they were confident had subjective experience should have certain kinds of rights and duties (see Figure 5). A strong majority of the US public and AI researchers somewhat or strongly agreed that AI systems should have responsibilities such as being held accountable for their actions, treating others well, and behaving with integrity, honesty, and fairness. Overall, respondents agreed more than they disagreed that AI systems deserved certain kinds of moral patency concerns and protections such as being respected, protected by the law, and cared for like people treat their pets. Between roughly one-fifth and one-third of respondents disagreed that AI systems would deserve various kinds of moral concern if they had subjective experience.

Both AI researchers and the US public disagreed most strongly with the assertions related to whether AI systems with subjective experience deserved rights such as having autonomy, expressing some civil and political rights (e.g., freedom of expression, citizenship, entering into contracts, voting), and having computing rights (e.g., the right to updates and maintenance, access to energy, self-development). This disagreement ranged between $29\%$ and $51\%$ of the US public "somewhat" or "strongly" disagreeing with items and between $40\%$ and $57\%$ of AI researchers "somewhat" or "strongly" disagreeing. The US public was significantly more supportive of computing rights ( $49\%$ agreed) than AI researchers were ( $35\%$ agreed).

Dreksler, Caviola, Chalmers, Allen, Rand, Lewis, Waggoner, Mays, and Sebo (2025)

Figure 5. Agreement with protections, rights, and responsibilities for AI systems with subjective experience: (A) Mean responses by group, (B) Percentage breakdown of responses by group. (A) Mean agreement for moral protections, rights, and responsibilities for AI systems with subjective experience Respondents were able to say to what extent they agreed or disagreed that an AI system they were confident had subjective experience should have certain kinds of protections, rights, and responsibilities on a five-point scale: strongly disagree (-2), somewhat disagree (-1), neither agree nor disagree (0), somewhat agree (1), strongly agree (2). Moral patency protections Be respected and treated the same as other people Have protection under the law from harm and mistreatment Be cared for and protected, as people treat their pets -2.0 -1.0 0.0 0.2 0.4 0.6 1.0 2.0

Socio-political rights

Have autonomy in its programming to act freely and not be under the control of others

Be able to express some civil and political rights (for example, freedom of expression, citizenship, entering into contracts, voting)

Have "computing rights" (for example, the right to updates and maintenance, access to energy, self-development)

Responsibilities Be held accountable for its actions Have a responsibility to treat all other beings (humans and non-humans) well Be expected to behave with integrity, honesty, and fairness

Created with Datawrapper

Public

Somewhat or strongly disagree Neither agree nor disagree Somewhat or strongly agree Don't know

Moral patency protections

Be respected and treated the same as other people

Have protection under the law from harm and mistreatment

Be cared for and protected, as people treat their pets

Desired with Datawrapper

Socio-political rights

Have autonomy in its programming to act freely and not be under the control of others

Be able to express some civil and political rights

Have "computing rights"

Responsibilities Be held accountable for its actions Have a responsibility to treat all other beings well Be expected to behave with integrity, honesty, and fairness

Created with Datawrapper

Public

Somewhat or strongly disagree Neither agree nor disagree Somewhat or strongly agree Don't know

Moral patency protections

Be respected and treated the same as other people

Have protection under the law from harm and mistreatment

Be cared for and protected, as people treat their pets

Socio-political rights

Have autonomy in its programming to act freely and not be under the control of others

Be able to express some civil and political rights

Have "computing rights"

Responsibilities

Be held accountable for its actions

Have a responsibility to treat all other beings well

Be expected to behave with integrity, honesty, and fairness

Created with Datawrapper

Welfare protection. Respondents were asked whether society should protect the welfare of six groups for their own sake: humans, animals, the environment, businesses, and AI systems both with and without subjective ex

Subjective experience in AI systems: What do AI researchers and the public believe?

perience. Protection was defined as "protecting the rights, interests, and/or well-being of a group," and responses were measured on a five-point scale from "strongly disagree" (-2) to "strongly agree" (2). The US public and AI researchers showed strong alignment in their responses. Both groups overwhelmingly agreed that humans, animals, and the environment should be protected for their own sake. However, opinions were more divided regarding businesses and AI systems with subjective experience, with similar levels of overall agreement between the two groups. For AI systems with subjective experience, $46\%$ of AI researchers and $43\%$ of the public somewhat or strongly agreed they should be protected, while $30\%$ of AI researchers and $32\%$ of the public somewhat or strongly disagreed. Regarding AI systems without subjective experience, the majority in both groups disagreed that their welfare should be protected for its own sake. However, this disagreement was notably stronger among AI researchers ( $71\%$ disagreed) compared to the US public ( $53\%$ disagreed).

Figure 6. Agreement with welfare protection of different groups (A) Mean responses by group, (B) Percentage breakdown of responses by group. (A) Mean agreement that different groups' welfare should be protected (B) Agreement that different groups' welfare should be protected percentages

AI researchers

What AI developers should do. Respondents saw seven randomly selected items from a total list of fourteen on (1) what governments and AI developers should do in relation to AI systems with subjective experience and (2) perceptions of risk and benefit. Respondents were asked to what extent they agree with each presented item on a five-point scale from "strongly disagree" (-2) to "strongly agree" (2). There is a perhaps surprising amount of appetite for addressing the risks and harms of AI systems with subjective experience now. The majority of the US public $(85\%)$ and AI researchers $(68\%)$ thought that AI developers should implement safeguards now to avoid the harms and risks from the development and deployment of AI systems with subjective experience, with slightly lower support for waiting until just before they exist. In comparison to AI researchers, the US public was significantly less supportive of developers actively building AI systems with subjective experience and more

Note that this item can refer to both the harms and risks from AI systems with subjective experience and to such AI systems.

Dreksler, Caviola, Chalmers, Allen, Rand, Lewis, Waggoner, Mays, and Sebo (2025)

in favor of developers implementing safeguards and of never building AI systems with subjective experience (see Figure 7). A plurality of the public $(41\%)$ thought that AI developers should never build AI systems with subjective experience, while just under a third $(31\%)$ disagreed with the statement. Of the AI researchers, $59\%$ disagreed that AI developers should never build AI systems with subjective experience.

Responses saw seven randomly selected items relating to what governments and AI developers should do in regard to AI systems with subjective experience and related risk/benefit perceptions and were asked to what extent they agree with them on a five-point scale: strongly disagree (-2), somewhat disagree (-1), neither agree nor disagree (0), somewhat agree (1), strongly agree (2).

Figure 7. Mean agreement on what AI developers and governments should do in regard to the governance of AI systems with subjective experience and on risk and benefit perceptions of AI systems with subjective experience. AI developers should... Figure 7. Mean agreement on what AI developers and governments should do in regard to the governance of AI systems with subjective experience

Governments should...

Ban the development and deployment of AI systems with subjective experience

Encourage the development and deployment of AI systems with subjective experience

Discourage the development and deployment of AI systems with subjective experience

Pass regulation now in regard to the development and deployment of AI systems with subjective experience

Pass regulation once such systems exist or will soon exist, not before, in regard to the development and deployment of AI systems with subjective experience

Do nothing and pass no regulation, not now nor in the future, in regard to the development and deployment of AI systems with subjective experience

Risk and benefit perceptions

AI systems with subjective experience would be more dangerous to humanity than AI systems without subjective experience.

AI systems with subjective experience would be able to behave more morally towards humanity than AI systems without subjective experience.

Created with Datawrapper

To follow up on these trends, we conducted a conjoint survey experiment exploring support for a hypothetical state-of-the-art AI project at a tech company that the respondent believed had subjective experience while varying whether there were safeguards in place at the company. We found that while having safeguards did not increase support, explicitly noting that there were no safeguards significantly decreased support. Further, public respondents showed significantly lower support than AI researchers did, in particular in the control condition. See full results in SI Section A.3.

Subjective experience in AI systems: What do AI researchers and the public believe?

Figure 8. Percentage breakdowns for views on what AI developers should do in regard to the governance of AI systems with subjective experience.

What governments should do. The majority of AI researchers disagreed with governments banning AI systems with subjective experience outright (66%). The US public was more divided on whether governments should ban AI systems with subjective experience: 41% either somewhat or strongly disagreed, 31% either somewhat or strongly agreed. A majority of the US public (66%) thought that governments should pass regulation in regard to the development and deployment of AI systems now, while AI researchers were divided approximately equally between agreement (41%) and disagreement (40%) on the issue. A strong majority of both groups disagreed that governments should never pass regulation in regard to AI systems with subjective experience (AI researchers: 78%, US public: 70%). In comparison to AI researchers, the US public was significantly more supportive of governments banning AI systems with subjective experience, discouraging their development and deployment, and passing regulation (see Figure 7).

Dreksler, Caviola, Chalmers, Allen, Rand, Lewis, Waggoner, Mays, and Sebo (2025)

Figure 9. Percentage breakdowns for views on what governments should do in regard to the governance of AI systems with subjective experience. Governments should...

Risk and benefit perceptions. When considering the risks and benefits of AI systems to humanity, do people view subjective experience as a benefit or danger? The majority of the US public (59%) somewhat or strongly agreed that AI systems with subjective experience were more dangerous than those without, with only 14% disagreeing. Just under half of AI researchers agreed (46%) and a quarter disagreed (25%), with AI researchers showing significantly lower mean agreement with the statement than the US public did (Figure 7). There was no consensus in nor significant difference between both groups on whether AI systems with subjective experience would be able to behave more morally towards humanity than would those without subjective experience. AI researchers were more evenly split between agreement (31%) and disagreement (34%), and the US public tended slightly towards agreement (38% agree vs 24% disagree). Both items saw a substantial fraction of respondents choosing either the "neither agree nor disagree" or the "I don't know" option, pointing to substantial uncertainty in the groups on these questions.

Subjective experience in AI systems: What do AI researchers and the public believe?

Figure 10. Percentage breakdowns for the public's and AI researchers' risk and benefit perceptions of AI systems with subjective experience. Risk and benefit perceptions Created with Datawrapper

3 Discussion

While the possibility of AI subjective experience has been discussed in academic literature, there remains a significant gap in surveying the perspectives of those with high levels of technical expertise regarding AI. Our study helps address this gap by surveying AI researchers who published in leading venues in the field, as well as a representative sample of the US public. The goal was to improve our understanding of public and expert opinions on the possibility of AI subjective experience and the moral and governance questions it may raise. The results highlight that both AI researchers and the public regard AI systems with subjective experience as a real future possibility. Additionally, our results reveal substantial division within both groups about many issues related to the science, ethics, and governance of potential digital minds.

Forecasts of AI subjective experience. Both AI researchers and the US public think it is unlikely that current AI systems have subjective experience, but believe it is more likely than not that such systems will exist by the end of the century. Based on the skew-normal finite mixture modes of the forecasts across both groups, the median AI researcher and member of the US public believed there is a $10\%$ chance that such AI systems may exist by 2030 ( $M = 24\%$ ), a $50\%$ chance they may exist by 2050 ( $M = 51\%$ ), a $90\%$ chance they may exist by 2100 ( $M = 71\%$ ). Prior research on AI forecasting (Grace et al. 2018, 2022, 2024; Zhang et al. 2022) has shown that the way questions were framed – either by specifying a probability or a date – can influence participants' timeline estimates. In our results, raw data and fitted distributions suggested that the fixed probability framing may have resulted in earlier forecasts at the upper end of the probability distribution and later forecasts at the lower end, when compared to the fixed date framing. A Wilcoxon rank sum tests did not reveal a statistically different difference between the two framings in terms of $50\%$ probability forecasts.

We also found that the US public assigned a significantly higher chance to the possibility that AI systems with subjective experience will never exist, compared to AI researchers. Especially for the public, there was a wide

range of responses for this question, reflecting a wide diversity of views. Overall, both groups were on average “somewhat confident” about their estimates for forecasting the existence of AI systems with subjective experience. The results indicate that while both groups ascribed some chance to AI subjective experience never existing, median AI researchers and members of the US public thought it is likely that AI systems with subjective experience will exist within this century.

Determining AI subjective experience.

Determining whether an AI system has subjective experience is likely to be a contentious and challenging issue, but little research has probed public or expert opinion on the process of determining AI subjective experience. Both AI researchers and the US public judged that, if an AI system did have subjective experience, we would be somewhat more likely than not to recognize it. However, the mean and median confidence was not particularly high pointing to notable uncertainty in both groups. AI researchers and the public were aligned in their responses, with no significant differences found. Both groups thought that we would have to be between somewhat and moderately confident that an AI system has subjective experience in order to warrant giving it some level of moral consideration.

AI researchers and the US public feel that a wide range of stakeholders’ and experts’ views are important in determining whether an AI system has subjective experience. Both groups placed the highest value on the expertise of technical AI researchers, neuroscientists and psychologists, and AI ethics experts. In contrast, policymakers’ views, the AI system’s own reports, and public opinion were rated as less important overall, though AI researchers assigned even lower importance on the latter two than the public did.

AI researchers were significantly more skeptical than the US public about the need for subjective experience in AI systems to perform various tasks at human-level or above. Around half or more of the US public thought that subjective experience was needed to be a therapist or judge and to generate a range of complex and evocative artistic outputs. Such findings could suggest that knowledge of or exposure to the use of AI systems in such roles in the future could have an effect on the public’s views of whether AI systems have subjective experience, further justifying the need for survey data on these issues.

However, the goalposts for what counts as AI subjective experience may shift considerably over time, as a similar effect has been observed with respect to AI capabilities and their perceived “human-ness” (Haenlein and Kaplan 2019; McCorduck 1979; Tesler 64). For instance, distinctly human capabilities and attributes are rated as more essential to being human after evaluators are made aware of particular AI capability advances (Santoro and Monin 2023). Nevertheless, it is true that more human-like characteristics, including social interaction, emotional expression, and appearance, are associated with higher ascriptions of subjective experience and other mental capacities to artificial entities in experimental studies (Broadbent et al. 2013; Cucciniello et al. 2023; Thellman et al. 2022). Thus, it may be the case that personal experiences with AI systems in such emotionally sensitive and social roles would drive changes in beliefs more than just awareness of such advances would. Future studies could explore whether it is the case that the ascription of AI subjective experience and changes in belief are heightened when there is an alignment between expectations, as explored in our study, and experiences with AI in certain roles.

The moral consideration and governance of AI systems with subjective experience.

Our findings reveal substantial division within both AI researchers and the US public on many ethical and governance questions related to AI systems with subjective experience. Nevertheless, both groups showed high consensus on a few normative issues. Firstly, there was strong agreement that AI systems with subjective experience should have a range of responsibilities with regard to acting ethically and being held accountable. Secondly, there was majority agreement that AI developers should implement safeguards now to avoid the risks and harms that could arise from AI systems with subjective experience. Finally, the majority of both groups disagreed with the notion that AI developers and governments should never implement safeguards or regulations in regard to the associated harms and risks of AI systems with subjective experience. Overall, however, the US public was more skeptical than AI researchers were of whether AI systems with subjective experience should be built and more supportive of implementing related safeguards, regulations, and bans.

For almost all other issues regarding the welfare, moral consideration, and governance of AI systems with subjective experience, we find notable within-group division among both AI researchers and the general public. In line with previous studies, both groups exhibited the least support for granting AI systems with subjective experience socio-political rights, middling support for offering protections that consider AI systems as moral patients, and strong support for requiring certain responsibilities and duties (Anthis et al. 2024; De Graaf et al. 2021, 2022; Lima et al. 2020; Mays et al. 2025). Notably though, even in regard to socio-political rights, a sizable minority of both groups somewhat or strongly agreed that AI systems with subjective experience should have autonomy in their programming, have some civil and political rights, or enjoy computing rights. In terms of protecting the welfare of AI systems with subjective experience for their own sake, both groups were particularly divided, much like they were for businesses. Meanwhile, other entities such as animals and the environment saw overwhelming support for their welfare to be protected for their own sake. This difference between welfare preferences for natural entities and AI systems has been found previously (Pauketat et al. 2023; Rottman et al. 2021).

Responses to questions about the governance of AI systems suggest that it may not be too early to consider institutional and policy approaches now. While 41% of AI researchers supported AI developers building AI systems with subjective experience, 33% favored banning them – though note that only 17% of surveyed researchers work in industry where frontier models are developed. The public showed significantly more hesitancy toward AI developers building such systems than did AI researchers. While AI researchers were roughly evenly divided between whether governments should begin to regulate the development and deployment of AI systems with subjective experience now or not, two-thirds of the public agreed that they should. In addition, across several questions, a notable number of respondents also suggested that they neither agree nor disagree, suggesting uncertainty and scope for changes in beliefs in the future. It seems likely that governance attitudes are not solely driven by concern for a potentially suffering AI system. Indeed, we found that people think that AI systems with subjective experience are likely to be more dangerous to humanity than are AI systems without subjective experience, with the US public agreeing with this assertion significantly more strongly than AI researchers. Both groups were divided on the issue of whether subjective experience would allow AI systems to behave more morally.

3.1 Limitations

Aside from standard challenges of online surveys, such as sample generalizability and data quality, the present study has several important limitations that should be considered when interpreting the results:

Folk conceptions. AI researchers’ and the public’s folk conceptions of mental capacities may differ from conceptions of experts who study the science and philosophy of mind, and we cannot be certain how participants in either group understand the term “subjective experience”. While we defined concepts in the survey, research has highlighted divergences between folk and expert understanding of mental-state-related concepts like subjective experience, intentionality, and consciousness (Arico et al. 2011; Knobe 2003; Knobe and Prinz 2008; Malle and Knobe 1997; Sytsma 2010). For example, there is evidence that folk psychological understanding and use of mental-state-related terms common in philosophical discussions can diverge from that of philosophers (Huebner 2017).

There are several implications of our samples that need consideration when drawing general conclusions. We only surveyed US public members, and geographic/cultural differences exist in AI attitudes and views on non-human sentience (Gillespie et al. 2023; Haddy et al. 2025; Policy, Elections, and Representation Lab and Schwartz Reisman Institute for Technology and Society 2024), meaning that findings may not generalize to other countries. Selection effects likely affected both samples. Despite efforts for representative public sampling, online platform participation may introduce bias. For AI researchers, despite our neutrally worded recruitment email concealing the survey topic beyond it being about AI, participation bias might exist. However, previous studies using similar sampling frames and neutral emails found little evidence of participation bias when comparing respondents with non-respondents (Grace et al. 2024; Zhang et al. 2022). For the public sample, online survey platforms raise data quality concerns, including bot risks (Douglas et al. 2023; Irish and Saba 2023; Webb and Tangney 2024). While Prolific generally provides higher quality data than other platforms do (Douglas et al. 2023) and we used an attention check, the prevalence of LLM bot use on survey platforms remains unclear.

2010; Sytsma and Machery 2010)*. Our findings also point to complexities in understanding how survey takers interpret terms such as “subjective experience” and that these interpretations may diverge from more common definitions among experts in consciousness and the mind. For example, in the first part of the survey, before focusing on subjective experience, we asked respondents to say which of eight mental capacities is required for consciousness, agency, and sentience in order to better understand the folk psychological understanding of these concepts (see SI Section A.2, Figure 13). Here we defined subjective experience only briefly as “capable of having experiences that feel like something from a single point of view.” Only roughly half of the public and AI researcher samples thought subjective experience thus defined was required for consciousness or sentience. In fact, self-awareness and perception and awareness were the most commonly chosen capacities required. This suggests that opinions of timelines for consciousness, self-awareness, and other capacities could be distinct from those for subjective experience.

Impact of emphasizing different aspects of subjective experience impacts beliefs. There is some suggestion from the results that the emphasis of the given term or its definition (e.g., highlighting valenced experience more strongly) can also have an effect on responses. For example, when we asked AI researchers about suffering sentient AI systems directly, AI researchers forecasts appeared to become later than for subjective experience. This is also reflected in both samples’ forecasts for when AI systems may have each of eight different mental capacities (see SI Section A.2, Figure 14), where the capacity for subjective experience is considered more likely to occur earlier (or at all) in AI systems than the capacity of having valenced experiences such as pleasure or pain.

Limited prior consideration of AI subjective experience may affect response stability. Questions about AI subjective experience represent complex philosophical territory that most survey respondents have likely given little prior consideration to before participating in the study. When faced with such abstract and unfamiliar concepts, respondents may form their opinions ad hoc during the survey itself, rather than drawing on well-developed views. This raises questions about the stability of these responses and their sensitivity to question framing. This limitation particularly affects our ability to draw strong conclusions about public attitudes towards AI subjective experience, as these attitudes may be quite malleable and sensitive to how the concepts are presented.

Survey responses may not reflect real-world behaviors and priorities. The present study is limited in terms of telling us how such issues are in fact weighed in the public agenda and how they influence actions. In particular, surveys focused on a single topic can give inaccurate impressions of how people weigh the importance of an issue in comparison to other risks in practice and do not necessarily transfer to real-world behaviors or how concerned people are about an issue day-to-day. Since AI systems are likely to have wide-reaching economic benefits and will be increasingly integrated into various tasks within society, there will likely also be a strong countervailing force both politically and attitudinally to any present concerns. Furthermore, survey data cannot be taken at face value in terms of predicting behavior and policy support. For instance, although most AI researchers and the US public in this survey supported welfare protections for animals and the environment – roughly twice as many as for AI systems with subjective experience – factory farming and environmental harm remain widespread, with little public support for factory farming bans (Ritchie 2023; Sentience Institute 2021; Tollefson 2019; Welfare Footprint 2021).

Forecasting limitations impact predictive validity. Forecasting the future through surveys and expert elicitation, as employed in this study regarding the potential emergence of subjective experience and other mental capacities in AI systems, presents significant methodological challenges (Tetlock 2006; Tetlock and Gardner 2015; Tourangeau et al. 2000). While aggregating predictions from multiple forecasters typically enhances accuracy (Sunstein 2006; Surowiecki 2004; Wagner et al. 2010), accurately predicting outcomes even a few years ahead remains exceptionally challenging, and domain experts do not necessarily outperform others in making such predictions (Tetlock 2006; Tetlock and Gardner 2015). Moreover, it is important to note that neither AI researchers nor the general public necessarily possess specific expertise in predicting the development of mental capacities in artificial intelligence systems. Taken together with the uncertainty about folk psychological inter

pretations of the term subjective experience, care is thus needed when interpreting the forecasting results. Such predictions should mainly aid us in understanding how different groups think about the world, rather than being seen as informative indicators about how AI technology may develop.

4 Future directions

Multi-expert surveys and participatory processes. Future research should focus on fielding large-scale surveys of a range of expert groups and stakeholders on the topic of AI subjective experience, as suggested by both groups’ valuing of a wide range of expertise in determining whether AI systems have the capacity for subjective experience. Such studies need to be conducted with sensitivity to the various challenges and limitations of the present study, such as by expanding the geographic scope of the survey and showing sensitivity to the fuzziness and complexities of the concepts in question. They could also explore people’s views on the treatment and governance of AI systems with the potential for actual or apparent subjective experience in more depth. Participatory processes that offer more expert input and discussion opportunities, such as citizens’ assemblies, would also be useful to conduct in the future to better understand people’s reflected attitudes since it is a topic most people have likely not yet given much thought to.

Risk-benefit perceptions and governance attitudes. In particular, it remains an open question to what extent attitudes on such topics will affect political behaviors now and in the future. Risk and benefit perceptions have been found to contribute to attitudes and behaviors relating to the acceptance and governance of emerging technologies and risks (Bearth and Siegrist 2016; Bickerstaff et al. 2006; De Groot et al. 2020; Pidgeon et al. 2005; Poortinga and Pidgeon 2006). The items in the present study remained neutral on what the risks and harms of AI systems may be when asking about how AI developers and governments should govern AI systems. Future research could expand on what risk and benefit perceptions may be driving attitudes towards the governance of AI systems with subjective experience. This could include individual and societal human concerns, as well as potential perceived harms to AI systems.

Monitoring AI capability advances and uncertainty management. Forward-looking studies could also try to monitor and predict how advances in AI capabilities and changes in use will affect perceptions of subjective experience in AI systems, allowing for anticipatory governance responses to mitigate such risks and implement sufficient safeguards. Deepening our understanding of how the public and different expert groups think we should handle our inherent uncertainty about the subjective experience of artificial entities will be another important step in navigating the related governance challenges. Finally, our findings suggest that it is not only subjective experience capacities such as consciousness or sentience that contribute to people’s attitudes towards moral consideration of AI systems. Other characteristics – such as perceptions of agency, autonomy, and self-awareness – may also relate to attitudes towards the governance of AI systems and should be explored further in multi-stakeholder large-scale surveys.

5 Conclusion

While both groups think that AI systems with subjective experience are a possibility this century, our findings reveal substantial disagreement and uncertainty, both within and between AI researchers and the US public, regarding the science, ethics, and governance of AI systems with subjective experience. As AI systems become more capable and human-like, people may increasingly anthropomorphize these systems and attribute subjective experience to them, whether accurately or not. This could make the issue feel more urgent and further deepen divisions in opinion.

These dynamics raise important questions about how to govern AI systems and implement safeguards against the risks of both over-attributing and under-attributing mental capacities such as subjective experience. Ongo

ing research will be essential for monitoring public and expert views, identifying areas of misunderstanding, and informing proactive policy decisions. This topic is likely to become both significant and contentious, especially given that technical experts – who help shape AI’s future and are trusted by the public – hold diverse and conflicting views themselves.

6 Methodology

6.1 Samples

We recruited two samples for this survey: a representative sample of the US public (N = 838) and a sample of AI researchers (N = 582). The survey was fielded between the 8th and 30th of May 2024.

AI researchers. AI and ML researchers were contacted via email from a sampling frame compiled by web scraping and manually identifying email addresses of authors who published in 2022 at a selection of top-tier machine learning conferences (NeurIPS, ICML, ICLR, AAAI, and IJCAI) and journals (JMLR) (Grace et al., 2024). Invitation emails were sent to a random selection of 10550 individuals on the list, of which 9407 received the emails (the other emails bounced). The final number of invitations was determined by hitting a pre-registered minimum and maximum recruitment quota limited by the funding available to field the survey. 6.8% of invitees opened the survey and consent form (N = 635) and 6.2% of invitees (N = 582) were in the final sample having completed at least the first question of the survey. 86.3% (N = 502) of the AI researcher sample completed all parts of the survey. The neutrally worded recruitment emails that were sent out to experts did not identify the topic of the survey in advance, beyond that it was related to AI. Respondents received the equivalent of $50 USD in the form of a Tango Reward gift voucher link, except respondents in China and Israel who were compensated using Wise.

Indicative of the long-standing gender disparities in the field of AI (Simonite, 2018; The Female Quotient, 2024), 85% of our sample were men, 10% were women, and a further 4% said they identified otherwise or would prefer not to say. The average and median AI researchers’ age was in their early thirties (M = 33.9, Mdn = 31.0, SE = 0.4). A large majority of the sample was in academia (78%) while 17% were in industry and 3% in government. Of those in academia, 37% were professors or lecturers, 42% were PhD students, and 12% were postdocs currently. The majority of those in industry were research scientists (69%) or research engineers (19%). 6% were in managerial positions. AI researchers who participated resided in North America (45%), Europe (31%), Asia (15%), Oceania (2%), Africa (<1%), the Caribbean (<1%), or other regions (5%). A larger fraction of the sample originated from Asia (47%) than were domiciled there. Distributions for religiosity and left-right political orientation can be found in SI Section E along with other top-line demographic and psychographic results for both AI researchers and the public.

Public. Members of the US public were recruited on the survey platform Prolific using representative sample criteria related to sex, age, ethnicity (based on the simplified US census), and political affiliation. Respondents were compensated for their time. Overall, 91.2% (N = 764) of the US public sample completed all parts of the survey. In terms of gender, 48% of the sample were men and 51% were women. 5% of the sample had no formal education, 10% had qualifications but did not attend university, 22% attended university but did not graduate, 36% completed a bachelor’s degree, 5% attended a graduate program but did not graduate, 16% had a graduate degree, and 7% reported another kind of education attainment. The sample was relatively evenly split between Democrats (31%), Independents (38%), and Republicans (30%). The majority of the sample was working currently (62%), with a further 9% unemployed, 5% students, and 12% retired respondents.

6.2 Survey design

After giving informed consent, respondents were first asked to say which mental capacities they believed were required in order to have consciousness, sentience, and agency. They then predicted when AI systems would first have these capacities. The survey then defined subjective experience in more detail (see Table 1), and all other questions were shown in four blocks related to 1) subjective experience forecasts, 2) whether and how one could determine AI subjective experience, 3) the moral consequences of AI subjective experience, and 4) the norms and governance of AI systems with subjective experience. The order of these four blocks was randomized. Finally, AI researchers were asked a series of questions for related studies (see SI Section H), of which we only report three brief questions here relating to their forecasts for whether sentient and suffering AI systems are possible. The survey questions can be found in SI Section F. AI researchers took a median of 30 minutes to complete the survey, while the public, who did not complete the experimental section of the survey, took a median of 18 minutes. Note, however, that duration measures from Qualtrics surveys can be distorted if someone leaves open a survey while not completing it or chooses to return to it at a later date (it is simply the difference between the start and end date of the survey, not a measure of active time).

6.3 Analysis

6.3.1 Fitting distributions to AI subjective experience forecasts

We used an iterative approach to encode each respondent’s elicited probabilities as a probability distribution. For each individual’s set of elicited year-cumulative probability pairs, we first attempted to fit a flexible metalog distribution. If this initial fit did not encode the data to within a pre-established degree of accuracy, we then sequentially attempted to fit a Johnson Quantile-Parameterised Distribution (for applicable question types), then a three-component skew-normal mixture model, and finally a shifted log-normal distribution. The first model in this sequence that passed its respective validation was selected as the best representation of that respondent’s forecast, with the parameters for the skew-normal mixture and shifted log-normal models being estimated via numerical optimization. Further details on the analysis approach can be found in SI Section C. The fitting approach deviated from the pre-registered analysis due to the limitations of using Gamma distributions for this purpose (see Section C.1).

Once a probability distribution had been fitted for every respondent, we converted each distribution into a yearly series by calculating, for every calendar year from 2024 to 2124, the probability for a given year. To produce group-level summaries we combined these probabilities in two ways: by averaging them across respondents year-by-year (mean aggregation) and by taking the median value across respondents for each year (median aggregation), which is less sensitive to extreme forecasts, giving equal weight to every person in both cases. The same mean- and median-aggregation steps were repeated separately for the entire sample, for each sampling frame (AI researchers and US public), and for each version of question framing (fixed date and fixed probability). The full results of both aggregations can be found in Section A.1.1.

6.3.2 Statistical comparisons and regression analyses

Between-group comparisons. For each question, we performed a t-test to compare the means of the two groups – AI researchers and the public – and applied the Holm-Bonferroni correction to adjust for multiple comparisons, thereby reducing the risk of type 1 errors (false positives). The full results can be found in SI Section A.5. For the AI subjective experience forecasts, Wilcoxon rank sum tests with continuity corrections were used to compare the median (50% probability) forecasts derived from the fitted forecasting distributions between groups (AI researchers vs. public), as well as between question framings (fixed probability vs. fixed date).

Further analyses.

We also compared the effect of imagining different safeguard conditions on participants’ support for a hypothetical project at an AI company. We used a two-way ANOVA to identify significant differences between conditions and groups, followed by Tukey’s test to compare specific pairs. The results of this analysis are in SI Section A.3. We also ran regression analyses to examine how demographic factors (age, gender, AI use, political views, religiosity) predicted key outcomes(median forecast for subjective experience, probability of AI subjective experience never occurring, importance of different groups for determining AI subjective experience, governance and risk perception items, and which entities’ welfare should be protected). These regression results can be found in SI Section A.6. Several pre-registered analyses were not conducted owing to space constraints in this manuscript and to maintain a clear analytical focus on the principal comparison between the two groups’ beliefs (see the pre-analysis plan on OSF: https://osf.io/4txdq/.

Acknowledgments

We are grateful for research assistance and editorial support from: Tom Gibbs, Ayushi Kadakia, Johanna Salu, and Toni Sims. We are grateful to Wes Cowley for his careful copyediting and valuable attention to detail that helped refine this paper. We warmly thank the following individuals for feedback and input: Florence Enock, Karen Levy, Carolyn Ashurst, Mackenzie Jorgensen, Markus Anderljung, Alan Chan, Ben Garfinkel, Stephen Clare, Jide Alaga, Ben Clifford, Matthew van der Merwe, Marie Buhl.

Author information

Authors and affiliations (A-Z)

Carter Allen PhD student, University of California, Berkeley

Lucius Caviola Senior Research Fellow, Global Priorities Institute, University of Oxford

David Chalmers Professor and Co-Director Center for Mind, Brain, and Consciousness, New York University

Noemi Dreksler Senior Research Fellow, Centre for the Governance of AI

Joshua Lewis Assistant Professor, New York University

Kate Mays Assistant Professor, University of Vermont

Alex Rand Senior Research Coordinator, Global Poverty Research Lab, Northwestern University

Philip Waggoner Senior Research Scientist, Department of Applied Mathematics and Statistics, Colorado School of Mines

Jeff Sebo Associate Professor, Director of the Center for Environmental and Animal Protection, Director of the Center for Mind, Ethics, and Policy, and Co-Director of the Wild Animal Welfare Program, New York University

Contributions (A-Z)

Carter Allen fielded the public opinion survey, completed code checking, and wrote and edited the paper. Lucius Caviola obtained funding, designed the survey, fielded the public opinion survey, wrote the pre-registration, and wrote and edited the paper. David Chalmers helped design the survey and edited the paper. Noemi Dreksler project managed, obtained funding, designed the survey, fielded the public and expert opinion surveys, wrote the pre-registration, conducted analysis, completed code checking, generated data visualizations, wrote and edited the paper, and prepared the Supplementary Information. Joshua Lewis designed the survey and fielded the surveys. Kate Mays designed the survey and wrote and edited the paper. Alex Rand conducted data analysis and wrote the forecasting method sections. Jeff Sebo proposed the research idea, obtained funding, designed the survey, wrote and edited the paper, and provided guidance on the project. Philip Waggoner wrote the pre-registration, completed code checking, and conducted data analysis. All authors discussed and commented on

methodological and conceptual aspects of the project.

Data and code availability

The data will be made available on OSF following academic publication (https://osf.io/4txdq/). The code for analyses will then also be uploaded to OSF alongside the pre-registration and pre-analysis plan. Demographic information will be stripped from all AI researcher responses to preserve anonymity.

Ethics declarations

The authors declare no competing interests. The study was approved by the New York University IRB board (#IRB-FY2020-4536).

References

Abele and Wojciszke (2007) Abele, A. E., and B. Wojciszke. 2007. “Agency and communion from the perspective of self versus others.” Journal of personality and social psychology 93 (5): 751–763.
Anthis et al. (2024) Anthis, J. R., J. V. T. Pauketat, A. Ladak, and A. Manoli. 2024. “What do people think about sentient AI?.”. Available at: https://arxiv.org/abs/2407.08867.
Anthropic (2025) Anthropic. 2025. “System card: Claude opus 4 & claude sonnet 4.”. Technical report, Anthropic. Available at: https://www-cdn.anthropic.com/6be99a52cb68eb70eb9572b4cafad13df32ed995.pdf.
Arcas et al. (2022) Arcas, B. A. y. 2022. “Artificial neural networks are making strides towards consciousness, according to Blaise Agüera y Arcas.”. Available at: https://www.economist.com/by-invitation/2022/09/02/artificial-neural-networks-are-making-strides-towards-consciousness-according-to-blaise-aguera-y-arcas. Accessed: 2025-02-20.
Arico et al. (2011) Arico, A., B. Fiala, R. F. Goldberg, and S. Nichols. 2011. “The folk psychology of consciousness.” Mind & Language 26 (3): 327–352.
Aru et al. (2023) Aru, J., M. E. Larkum, and J. M. Shine. 2023. “The feasibility of artificial consciousness through the lens of neuroscience.” Trends in Neuroscience 46: 1008–1017.
Asaro (2007) Asaro, P. 2007. “Robots and responsibility from a legal perspective.”. Available at: https://arxiv.org/abs/2411.00986.
Avramides (2023) Avramides, A. 2023. “Other minds.” The Stanford Encyclopedia of Philosophy Winter 2023.
Azzalini (1988) Azzalini, A. 2023. The R package zn: The skew-normal and related distributions such as the skew-t and the SUN (version 2.1.1). Università degli Studi di Padova, Italia. Home page: http://azzalini.stat.unipd.it/SN/.
Baars (1988) Baars, B. J. 1988. A Cognitive Theory of Consciousness. Cambridge University Press.
Bain et al. (2014) Bain, P. G., J. Vaes, and J. P. Leyens. 2014. Humanness and dehumanization.
Banks (2021) Banks, J. 2021. “From warranty voids to uprising advocacy: Human action and the perceived moral patiency of social robots.” Frontiers in Robotics and AI 8.
Banks et al. (2022) Banks, J., and N. D. Bowman. 2022. “Perceived moral patiency of social robots: Explication and scale development.” International Journal of Social Robotics 15: 101–113.
Banks et al. (2021) Banks, J., A. P. Edwards, and D. Westerman. 2021. “The space between: Nature and machine heuristics in evaluations of organisms, cyborgs, and robots.” Cyberpsychology, Behavior and Social Networking 24 (5): 324–331.
Banks et al. (2022) Banks, J., and K. Koban. 2022. “A kind apart: The limited application of human race and sex stereotypes to a humanoid social robot.” International Journal of Social Robotics 15: 1949–1961.
Bearth et al. (2016) Bearth, A., and M. Siegrist. 2016. “Are risk or benefit perceptions more important for public acceptance of innovative food technologies: A meta-analysis.” Trends in Food Science & Technology 49: 14–23.
Bernotat et al. (2018) Bernotat, J., and F. Eyssel. 2018. “Can(‘t) wait to have a robot at home? - Japanese and German users’ attitudes toward service robots in smart homes.” 2018 27th IEEE International Symposium on Robot and Human Interactive Communication (ROMAN): 15–22.
Bickerstaff et al. (2006) Bickerstaff, K., P. Simmons, and N. F. Pidgeon. 2006. “Public perceptions of risk, science and governance: main findings of a qualitative study of six risk cases (Technical Report 06-03).”. Technical report, Centre for Environmental Risk.
Birch (2024) Birch, J. 2024. The Edge of Sentience. Oxford University Press.

Birhane, A., and J. Van Dijk. 2020. “Robot rights? let’s talk about human welfare instead.”. In Proceedings of the AAAI/ACM Conference on AI, Ethics, and Society, pp. 207–213.

Block, N. 1995. “On a confusion about a function of consciousness.” Behavioral and brain sciences 18 (2): 227–247.
Boine, C. 2023, Feb 27. “Emotional attachment to AI companions and european law.” MIT Case Studies in Social and Ethical Responsibilities of Computing (Winter 2023). https://mit-serc.pubpub.org/pub/ai-companions-eu-law.
Bostrom, N. 2014. Superintelligence: Paths, Dangers, Strategies. Oxford University Press.
Bourget, D., and D. Chalmers. 2023. “Philosophers on philosophy: The 2020 Philpapers survey.” Philosophers’ Imprint 23.
Bowman, S. 2024. “The checklist: What succeeding at AI safety will involve.”. Available at: https://sleepinyourhat.github.io/checklist/. Accessed: 2025-02-20.
Broadbent, E., V. Kumar, X. Li, J. r. Sollers, R. Q. Stafford, B. A. MacDonald, and D. M. Wegner. 2013. “Robots with display screens: A robot with a more humanlike face display is perceived to have more mind and a better personality.”. Available at: https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0072589.
Broadbent, E., V. Kumar, X. Li, J. Sollers 3rd, R. Q. Stafford, B. A. MacDonald, and D. M. Wegner. 2013. “Robots with display screens: a robot with a more humanlike face display is perceived to have more mind and a better personality.” PloS one 8 (8): e72589.
Browning, H., and W. Veit. 2022. “The sentience shift in animal research.” A Multidisciplinary Journal of Biotechnology and the Body 28 (4): 299–314.
Bryson, J. J. 2010. “Robots should be slaves.” Close engagements with artificial companions: Key social, psychological, ethical and design issues 8: 63–74.
Butlin, P., R. Long, E. Elmoznino, Y. Bengio, J. Birch, A. Constant, G. Deane, S. M. Fleming, C. Frith, X. Ji, R. Kanai, C. Klein, G. Lindsay, M. Michel, L. Mudrik, M. A. K. Peters, E. Schwitzgebel, J. Simon, and R. Van-Rullen. 2023. “Consciousness in artificial intelligence: Insights from the science of consciousness.”. Available at: https://arxiv.org/abs/2308.08708.
Byrne, H. 2024. “Inside OpenAI’s trust and safety operation - with Rosie Campbell.”. Available at: https://www.buzzsprout.com/2277524/episodes/14648694-inside-openai-s-trust-and-safety-operation-with-rosie-campbell. Accessed: 2025-02-20.
Caviola, L. 2025. “The societal response to potentially sentient AI.”. Available at: https://arxiv.org/abs/2502.00388.
Caviola, L., M. Coleman, C. Winter, and J. Lewis. 2024. “Crying wolf: Warning about societal risks can be reputationally risky.” Preprint at https://doi. org/10.31234/osf. io/gtr53.
Centre for Data Ethics and Innovation. 2024. “Public attitudes to data and AI: Tracker survey (Wave 3).”. Technical report, Centre for Data Ethics and Innovation. Available at: https://www.gov.uk/government/publications/public-attitudes-to-data-and-ai-tracker-survey-wave-3/public-attitudes-to-data-and-ai-tracker-survey-wave-3.
Chalmers, D. J. 1995. “Facing up to the problem of consciousness.” Journal of Consciousness Studies 2 (3): 200–19.
Chalmers, D. J. 1996. The Conscious Mind: In Search of a Fundamental Theory (2nd edition). Oxford University Press.
Chalmers, D. J. 2000. “What is a neural correlate of consciousness.” Neural correlates of consciousness: Empirical and conceptual questions 17.
Chalmers, D. J. 2018. “The meta-problem of consciousness.” Journal of Consciousness Studies 25 (9-10): 6–61.
Chalmers, D. J. 2022. Reality+: Virtual Worlds and the Problems of Philosophy. New York: W. W. Norton.
Chalmers, D. J. 2023. “Could a large language model be conscious?.”. Available at: https://www.bostonreview.net/articles/could-a-large-language-model-be-conscious/. Accessed: 2025-03-04.
Chopra, S., and L. F. White. 2011. A Legal Theory for Autonomous Artificial Agents. University of Michigan Press.
Christian, B. 2022. “How a Google employee fell for the Eliza effect.”. Available at: https://www.theatlantic.com/ideas/archive/2022/06/google-lamda-chatbot-sentient-ai/661322/. Accessed: 2025-02-20.
Coeckelbergh, M. 2010. “Robot rights? towards a social-relational justification of moral consideration.” Ethics and Information Technology 12: 209–221.
Colombatto, C., and S. M. Fleming. 2024. “Folk psychological attributions of consciousness to large language models.” Neuroscience of Consciousness 2024 (1).
Consortium, C., O. Ferrante, U. Gorska-Klimowska, S. Henin, R. Hirschhorn, A. Khalaf, A. Lepauvre, L. Liu, D. Richter, Y. Vidal, N. Bonacchi, T. Brown, P. Sripad, M. Armendariz, K. Bendtz, T. Ghafari, D. Hetenyi, J. Jeschke, C. Kozma, D. R. Mazumder, S. Montenegro, A. Seedat, A. Sharafeldin, S. Yang, S. Baillet, D. J. Chalmers, R. M. Cichy, F. Fallon, T. I. Panagiotaropoulos, H. Blumenfeld, F. P. de Lange, S. Devore, O. Jensen, G. Kreiman, H. Luo, M. Boly, S. Dehaene, C. Koch, G. Tononi, M. Pitts, L. Mudrik, and L. Melloni. 2023. “An adversarial collaboration to critically evaluate theories of consciousness.” bioRxiv.
Cucciniello, I., S. Sangiovanni, G. Maggi, and S. Rossi. 2023. “Mind perception in HRI: Exploring users’ attribution of mental

Dreksler, Caviola, Chalmers, Allen, Rand, Lewis, Waggoner, Mays, and Sebo (2025)

and emotional states to robots with different behavioural styles.” International Journal of Social Robotics 15: 867–877.

Cuddy et al. (2008) Cuddy, A. J., S. T. Fiske, and P. Glick. 2008. “Warmth and competence as universal dimensions of social perception: The stereotype content model and the BIAS map.” Advances in Experimental Social Psychology 40: 61–149.
Danaher (2020) Danaher, J. 2020. “Welcoming robots into the moral circle: A defence of ethical behaviourism.” Science and Engineering Ethics 26: 2023–2049.
Darling (2012) Darling, K. 2012. “Extending legal protection to social robots: The effects of anthropomorphism, empathy, and violent behavior towards robotic objects.”. Available at: https://papers.ssrn.com/sol3/papers.cfm?abstract_id=2044797.
De Freitas et al. (2024) De Freitas, J., A. K. Uğuralp, Z. Uğuralp, and S. Puntoni. 2024, June 1. “AI companions reduce loneliness.”. Available at: https://ssrn.com/abstract=4893097. The Wharton School Research Paper.
De Graaf et al. (2021) De Graaf, M., F. A. Hindriks, and K. V. Hindriks. 2021. “Who wants to grant robots rights?.” HRI ’21 Companion: Companion of the 2021 ACM/IEEE International Conference on Human-Robot Interaction: 38–46.
De Graaf et al. (2022) De Graaf, M. M. A., F. A. Hindriks, and K. V. Hindriks. 2022. “Who wants to grant robots rights?.” Frontiers in Robotics and AI 8.
De Groot et al. (2017) De Groot, J. I. M., E. Schweiger, and I. Schubert. 2020. “Social influence, risk and benefit perceptions, and the acceptability of risky energy technologies: An explanatory model of nuclear power versus shale gas.” Risk Analysis 40 (6): 1226–1243.
DeGrazia (2022) DeGrazia, D. 2022. “Robots with moral status?.” Perspectives in Biology and Medicine 65 (1): 73–88.
Dehaene et al. (1978) Dehaene, S., H. Lau, and S. Kouider. 2017. “What is consciousness, and could machines have it?.” Science 358: 486–492.
Delcker et al. (1978) Delcker, Janosch. 2018. “Europe divided over robot ‘personhood’.”. Available at: https://www.politico.eu/article/europe-divided-over-robot-ai-artificial-intelligence-personhood/.
Dennett (1978) Dennett, D. C. 1978. “Why you can’t make a computer that feels pain.” Synthese 38 (3): 415–456.
Dennett et al. (1994) Dennett, D. C., F. Dretske, S. Shurville, A. Clark, I. Aleksander, and J. Cornwell. 1994. “The practical requirements for making a conscious robot [and discussion].” Philosophical Transactions: Physical Sciences and Engineering 349 (1689): 133–146.
Desislavov et al. (2023) Desislavov, R., F. Martínez-Plumed, and J. Hernández-Orallo. 2023. “Trends in AI inference energy consumption: Beyond the performance-vs-parameter laws of deep learning.” Sustainable Computing: Informatics and Systems 38: 100857.
Doerig et al. (2021) Doerig, A., A. Schurger, and M. H. Herzog. 2021. “Hard criteria for empirical theories of consciousness.” Cognitive Neuroscience 12: 41–62.
Douglas et al. (1992) Douglas, B. D., P. J. Ewell, and M. Brauer. 2023. “Data quality in online human-subjects research: Comparisons between MTurk, Prolific, CloudResearch, Qualtrics, and SONA.”. Available at: https://doi.org/10.1371/journal.pone.0279720.
Dreksler et al. (2025) Dreksler, N., H. Law, C. Ahn, D. S. Schiff, K. J. Schiff, and Z. Peskowitz. 2025, January. “What does the public think about AI? an overview of the public’s attitudes towards AI and a resource for future research.”. Technical report, Centre for the Governance of AI. Available at: https://www.governance.ai/research-paper/what-does-the-public-think-about-ai. Accessed: 2025-01-22.
Dupont et al. (2023) Dupont, J., S. Wride, V. Ali, and J. Wilson. 2023. “What does the public think about AI?.”. Available at: https://publicfirst.co.uk/ai-us/. Accessed: 2025-03-04.
Epley et al. (2008) Epley, N., S. Akalis, A. Waytz, and J. T. Cacioppo. 2008. “Creating social connection through inferential reproduction: Loneliness and perceived agency in gadgets, gods, and greyhounds.” Psychological Science 19 (2): 114–120. PMID: 18271858.
Epley et al. (2011) Epley, N., A. Waytz, and J. T. Cacioppo. 2007. “On seeing human: A three-factor theory of anthropomorphism.” Psychological Review 114 (4): 864–886.
Eyssel et al. (2011) Eyssel, F., and D. Kuchenbrandt. 2011. “Manipulating anthropomorphic inferences about NAO: The role of situational and dispositional aspects of effectance motivation.” 2011 RO-MAN: 467–472.
Eyssel et al. (2012) Eyssel, F., D. Kuchenbrandt, and S. Bobinger. 2011. “Effects of anticipated human-robot interaction and predictability of robot behavior on perceptions of anthropomorphism.” 2011 6th ACM/IEEE International Conference on Human-Robot Interaction (HRI): 61–67.
Eyssel et al. (2021) Eyssel, F., D. Kuchenbrandt, and S. Bobinger. 2021. “‘if you sound like me, you must be more human’: on the interplay of robot and user features on human-robot acceptance and anthropomorphism.” Proceedings of the Seventh Annual ACM/IEEE International Conference on Human-Robot Interaction: 125–126.
Faber et al. (2022) Faber, I., and J. Jung. 2021. rmetalog: The Metalog Distribution. R package version 1.0.3.
Fenwick (2024) Fenwick, C. 2024. “Understanding the moral status of digital minds.”. Available at: https://80000hours.org/problem-profiles/moral-status-digital-minds/. Accessed: 2025-03-04.
Fernandez et al. (2024) Fernandez, I., N. Kyosovska, J. Luong, and G. , Mukobi. 2024. “AI consciousness and public perceptions: Four futures.”. Available at: https://doi.org/10.48550/arXiv.2408.04771.
Francken et al. (2022) Francken, J. C., L. Beerendonk, D. Molenaar, J. J. Fahrenfort, J. D. Kiverstein, A. K. Seth, and S. van Gaal. 2022, 08.

“An academic survey on theoretical foundations, common assumptions and the current state of consciousness science.” Neuroscience of Consciousness 2022 (1): niac011. Friedman, C. 2023. “Granting negative rights to humanoid robots.” Frontiers in Artificial Intelligence and Applications 366: 145–154. Gallagher, S. 2006. How the Body Shapes the Mind. Clarendon Press. Gamez, D. 2008. “Progress in machine consciousness.” Consciousness and Cognition 17 (3): 887–910. Gellers, J. C. 2020. Rights for robots: artificial intelligence, animal and environmental law (edition 1). Routledge. Gibert, M., and D. Martin. 2021. “In search of the moral status of AI: Why sentience is a strong argument.” AI and Society 1: 1–12. Gillespie, N., S. Lockey, C. Curtis, J. Pool, and A. Akbari. 2023. “Trust in artificial intelligence.”. Technical report, The University of Queensland; KPMG. Available at: https://assets.kpmg.com/content/dam/kpmg/au/pdf/2023/trust-in-ai-global-insights-2023.pdf. Goldstein, S., and C. D. Kirk-Giannini. 2024. “A case for AI consciousness: Language agents and global workspace theory.”. Available at: https://arxiv.org/abs/2410.11407. Goodwin, G. P. 2015. “Experimental approaches to moral standing.” Philosophy Compass 10 (12): 914–926. Gordon, J.-S., and A. Pasvenskiene. 2021. “Human rights for robots? A literature review.” AI and Ethics 1: 579–591. Grace, K., J. Salvatier, A. Dafoe, B. Zhang, and O. Evans. 2018. “When will AI exceed human performance? Evidence from AI experts.” Journal of Artificial Intelligence Research 62 (1): 729 – 754. Grace, K., Z. Stein-Perlman, B. Weinstein-Raun, and J. Salvatier. 2022. “2022 expert survey on progress in AI.”. Available at: https://aiimpacts.org/2022-expert-survey-on-progress-in-ai/. Grace, K., H. Stewart, J. F. Sandkühler, S. Thomas, B. Weinstein-Raun, and J. Brauner. 2024. “Thousands of AI authors on the future of AI.”. Available at: https://arxiv.org/abs/2401.02843. Gray, H. M., K. Gray, and D. M. Wegner. 2009. “Dimensions of mind perception.” Science 315: 619. Gray, K., and D. M. Wegner. 2012. “Feeling robots and human zombies: Mind perception and the uncanny valley.” Cognition 125 (1): 125–130. Gray, K., L. Young, and A. Waytz. 2012. “Mind perception is the essence of morality.” Psychological Inquiry 23 (2): 101–124. Guingrich, R. E., and M. S. A. Graziano. 2024. “Ascribing consciousness to artificial intelligence: Human-AI interaction and its carry-over effects on human-human interaction.” Frontiers in Psychology 15. Gunkel, D. J. 2012. The Machine Question: Critical Perspectives on AI, Robots, and Ethics. MIT Press. Gunkel, D. J. 2020. “Perspectives on ethics of AI: Philosophy.” In The Oxford Handbook of Ethics of AI, edited by M. D. Dubber, F. Pasquale, and S. Das, 539–553. New York: Oxford University Press. Gunkel, D. J. 2024/2018. Robot Rights. MIT Press. Haddy, E., F. Burden, Z. Raw, J. B. Rodrigues, J. H. Z. Bello, J. Brown, J. Kaminski, and L. Proops. 2025. “Belief in animal sentience and affective owner attitudes are linked to positive working equid welfare across six countries.” Journal of Applied Animal Welfare Science 28 (1): 116–134. Hadlock, C. C., and J. E. Bickel. 2017. “Johnson quantile-parameterized distributions.” Decision Analysis 14 (1): 35–64. Haenlein, M., and A. Kaplan. 2019. “A brief history of artificial intelligence: On the past, present, and future of artificial intelligence.” California Management Review 61 (4): 5–14. Harnad, S. 1991. “Other bodies, other minds a machine incarnation of an old philosophical problem.” Minds and Machines 1: 43–54. Harris, J., and J. R. Anthis. 2021. “The moral consideration of artificial entities: A literature review.” Science and Engineering Ethics 27 (53). Hashim, S. 2024. “Anthropic has hired an ’AI welfare’ researcher.”. Available at: https://www.transformernews.ai/p/anthropic-ai-welfare-researcher. Accessed: 2025-02-20. Haslam, N. 2006. “Dehumanization: An integrative review.” Personality and Social Psychology Review 10 (3): 252–264. Haslam, N., S. Loughnan, Y. Kashima, and P. Bain. 2008. “Attributing and denying humanness to others.” European Review of Social Psychology 19 (1): 55–85. Hegel, F. 2012. “Effects of a robot’s aesthetic design on the attribution of social capabilities.” 2012 IEEE RO-MAN: The 21st IEEE International Symposium on Robot and Human Interactive Communication: 469–475. Ho, A., T. Besiroglu, E. Erdil, D. Owen, R. Rahman, Z. C. Guo, D. Atkinson, N. Thompson, and J. Sevilla. 2024. “Algorithmic progress in language models.”. Available at: https://arxiv.org/abs/2403.05812. Huckins, G. 2023. “Minds of machines: The great AI consciousness conundrum.”. Available at: https://www.technologyreview.com/2023/10/16/1081149/ai-consciousness-conundrum/. Accessed: 2025-02-20. Huebner, B. 2010. “Commonsense concepts of phenomenal consciousness: Does anyone care about functional zombies?.”

Phenomenology and the Cognitive Sciences 9: 133–155.

Ingram, B. 2024. rjqpd: The Johnson Quantile-Parameterised Distribution. R package version 0.2.3.
Irish, K., and J. Saba. 2023. “Bots are the new fraud: A post-hoc exploration of statistical methods to identify bot-generated responses in a corrupt data set.” Personality and Individual Differences 213: 112289.
Jacobs, O. L., K. Gazzaz, and A. Kingstone. 2022. “Mind the robot! Variation in attributions of mind to a wide set of real and fictional robots.” International Journal of Social Robotics 14: 529–537.
Johnson, K. 2022. “LaMDA and the sentient AI trap.”. Available at: https://www.wired.com/story/lamda-sentient-ai-bias-google-blake-lemoine/. Accessed: 2025-02-20.
Keelin, T. W. 2016. “The metalog distributions.” Decision Analysis 13 (4): 243–277.
Keijsers, M., C. Bartneck, and F. Eyssel. 2021. “Pay them no mind: The influence of implicit and explicit robot mind perception on the right to be protected.” International Journal of Social Robotics 14: 499–514.
Killheffer, R. K. 2024. “The consciousness wars.” Omni 16 (1): 50+.
Kim, H.-y., and A. L. McGill. 2024. “AI-induced dehumanization.” Journal of Consumer Psychology.
Knobe, J. 2003. “Intentional action in folk psychology: An experimental investigation.” Philosophical psychology 16 (2): 309–324.
Knobe, J., and J. Prinz. 2008. “Intuitions about consciousness: Experimental studies.” Phenomenology and the cognitive sciences 7: 67–83.
Koban, K., and J. Banks. 2024. “It feels, therefore it is: Associations between mind perception and mind ascription for social robots.” Computers in Human Behavior 153: 108098.
Koch, C. 2020. The Feeling of Life Itself. The MIT Press.
Koch, C., M. Massimini, M. Boly, and G. Tononi. 2016. “Neural correlates of consciousness: Progress and problems.” Nature reviews neuroscience 17 (5): 307–321.
Kotzmann, J., M. Stonebridge, and P. Christiaenssen. 2023. “Evolving conceptions of legal personhood: What might recent legal developments herald for non-human animals in Australia?.” UNSWLJ 46: 347.
Kozak, M. N., A. A. Marsh, and D. M. Wegner. 2006. “What do I think you’re doing? Action identification and mind attribution.” Journal of Personality and Social Psychology.
Krach, S., F. Hegel, B. Wrede, G. Sagerer, F. Binkofski, and T. Kircher. 2008. “Can machines think? Interaction and perspective taking with robots investigated via fMRI.” PloS one 3 (7).
Kurki, V. A. 2019. The Legal Personhood of Artificial Intelligences. Oxford University Press.
Ladak, A., J. Harris, and J. R. Anthis. 2024. “Which artificial intelligences do people care about most? A conjoint experiment on moral consideration.” CHI ’24: Proceedings of the 2024 CHI Conference on Human Factors in Computing Systems 27: 1 – 11.
Ladak, A., S. Loughnan, and M. Wilks. 2024. “The moral psychology of artificial intelligence.” Current Directions in Psychological Science 33 (1): 27–34.
Lenharo, M. 2023. “Decades-long bet on consciousness ends — and it’s philosopher 1, neuroscientist 0.” Nature 619: 14–15.
Lenharo, M. 2024. “The consciousness wars: Can scientists ever agree on how the mind works?.” Nature 625: 438–440.
Levine, J. 1983. “Materialism and Qualia: The explanatory gap.” Pacific Philosophical Quarterly 64 (October): 354–61.
Levy, N., and J. Savulescu. 2009. “Moral significance of phenomenal consciousness.” Progress in brain research: 361–370.
Lima, G., C. Kim, S. Ryu, C. Jeon, and M. Cha. 2020. “Collecting the public perception of AI and robot rights.” Proceedings of the ACM on Human-Computer Interaction 4 (135): 1 – 24.
Liu, A. R., P. Pataranutaporn, S. Turkle, and P. Maes. 2024. “Chatbot companionship: A mixed-methods study of companion chatbot usage patterns and their relationship to loneliness in active users.” arXiv preprint arXiv:2410.21596.
Long, R., J. Sebo, P. Butlin, K. Finlinson, K. Fish, J. Harding, J. Pfau, T. Sims, J. Birch, and D. Chalmers. 2024. “Taking AI welfare seriously.”. Available at: https://arxiv.org/abs/2411.00986.
MacDorman, K. F. 2024. “Does mind perception explain the uncanny valley? a meta-regression analysis and (de)humanization experiment.” Computers in Human Behavior: Artificial Humans 2 (1): 100065.
Malle, B. F. 2016. “Integrating robot ethics and machine morality: the study and design of moral competence in robots.” Ethics and Information Technology 18: 243–256.
Malle, B. F. 2019. “How many dimensions of mind perception really are there?.” Proceedings of the 41st Annual Meeting of the Cognitive Science Society: 2268–2274.
Malle, B. F., and J. Knobe. 1997. “The folk concept of intentionality.” Journal of experimental social psychology 33 (2): 101–121.
Malle, B. F., and E. Phillips. 2023, Aug. “A robot’s justifications, but not explanations, mitigate people’s moral criticism and preserve their trust.”. Available at: osf.io/preprints/psyarxiv/dzvn4_v1.

Mamak, K. 2022. “Should violence against robots be banned?.” International Journal of Social Robotics 14: 1057–1066.

Mamak, K. 2023. “The moral significance of the phenomenology of phenomenal consciousness in case of artificial agents.” AJOB Neuroscience 14 (2): 160–162.
Manzi, F., G. Peretti, C. Di Dio, A. Cangelosi, S. Itakura, T. Kanda, H. Ishiguro, D. Massaro, and A. Marchetti. 2020. “A robot is not worth another: Exploring children’s mental state attribution to different humanoid robots.” Frontiers in Psychology 11.
Martínez, E., and C. Winter. 2021. “Protecting sentient artificial intelligence: a survey of lay intuitions on standing, personhood, and general legal protection.” Frontiers in Robotics and AI 8: 788355.
Martínez, E., and K. Tobia. 2023. “What do law professors believe about law and the legal academy?.” The Georgetown Law Journal 112 (1): 111–189.
Martínez, E., and C. Winter. 2021. “Protecting future generations: A global survey of legal academics.”. Available at: https://papers.ssrn.com/sol3/papers.cfm?abstract_id=3931304.
Mays, K. K., J. J. Cummings, and J. E. Katz. 2025. “The robot rights and responsibilities scale: Development and validation of a metric for understanding perceptions of robots’ rights and responsibilities.” International Journal of Human–Computer Interaction 41 (5): 3413–3430.
Mazor, M., A. Demertzi, J. J. Fahrenfort, N. Faivre, J. Francken, D. Lamy, M. Moutoussis, R. Salomon, D. Soto, and T. Stein. 2023. “The scientific study of consciousness cannot, and should not, be morally neutral.” Perspectives on Psychological Science 18 (3): 535–543.
McCorduck, P. 1979. Machines who think : A personal inquiry into the history and prospects of artificial intelligence. W.H. Freedman and Company.
Mcginn, C. 1989. “Can we solve the Mind-Body Problem?.” Mind 98 (July): 349–66.
Mei, Q., Y. Xie, W. Yuan, and M. O. Jackson. 2024. “A Turing test of whether AI chatbots are behaviorally similar to humans.” Proceedings of the National Academy of Sciences 121 (9): e2313925121.
Metzinger, T. 2021. “Artificial suffering: An argument for a global moratorium on synthetic phenomenology.” Journal of Artificial Intelligence and Consciousness 08 (01): 43–66.
Miller, L. F. 2015. “Granting automata human rights: Challenge to a basis of full-rights privilege.” Human Rights Review 16 (4): 369–391.
Miraglia, L., G. Peretti, F. Manzi, C. Di Dio, D. Massaro, and A. Marchetti. 2023. “Development and validation of the attribution of mental states questionnaire (AMS-Q): A reference tool for assessing anthropomorphism.” Frontiers 14.
Mosakas, K. 2021. “Human rights for robots? The moral foundations and epistemic challenges.” AI and Society: 1–17.
Müller, V. C. 2021. “Is it time for robot rights? Moral status in artificial entities.” Ethics and Information Technology 23: 579–587.
Nagel, T. 1974. “What is it like to be a bat?.” The Philosophical Review 83 (4): 435–450.
Nagel, T. 1986. The View From Nowhere. New York: Oxford University Press.
Nijssen, S. R. R., B. C. N. Müller, R. B. v. Baaren, and M. Paulus. 2019. “Saving the robot or the human? Robots who feel deserve moral care.” Social Cognition 37 (1): 41–S2.
Okanda, M., K. Taniguchi, Y. Wang, and S. Itakura. 2021. “Preschoolers’ and adults’ animism tendencies toward a humanoid robot.” Computers in Human Behavior 118: 106688.
Oksanen, A., N. Savela, R. Latikka, and A. Koivula. 2020. “Trust toward robots and artificial intelligence: An experimental approach to human–technology interactions online.” Frontiers in Psychology 11: 568256.
Oxford Analytica. 2022. “AI ’companions’ will become common.”. Available at: https://www.emerald.com/insight/content/doi/10.1108/oxan-db271754/full/html. Accessed: 2025-02-24.
Pauketat, J., A. Ladak, and J. R. Anthis. 2021. “Artificial intelligence, morality, and sentience (AIMS) survey: 2021.”. Available at: https://www.sentienceinstitute.org/aims-survey-2021. Accessed: 2025-02-25.
Pauketat, J., A. Ladak, and J. R. Anthis. 2023. “Artificial intelligence, morality, and sentience (AIMS) survey: 2023.”. Available at: https://www.sentienceinstitute.org/aims-survey-2023. Accessed: 2025-02-25.
Perez-Osorio, J., S. Marchesi, D. Ghiglino, M. Ince, and A. Wykowska. 2019. “More than you expect: Priors influence on the adoption of intentional stance toward humanoid robots.”. In Social Robotics: 11th International Conference, ICSR 2019, Madrid, Spain, November 26–29, 2019, Proceedings, Berlin, Heidelberg, pp. 119–129. Springer-Verlag.
Pew Research Center. 2024. “Pew Research Center’s American trends panel 7 - 11 February 2024.”. Available at: https://www.pewresearch.org/wp-content/uploads/2024/03/SR_24.03.26_chat-bot_topline.pdf.
Pidgeon, N. F., W. Poortinga, G. Rowe, T.-H. Jones, J. Walls, and T. O’Riordan. 2005. “Using surveys in public participation processes for risk decision making: The case of the 2003 British GM Nation? Public debate.” Risk analysis : an official publication of the Society for Risk Analysis 25 (2): 467–479.

Policy, Elections, and Representation Lab and Schwartz Reisman Institute for Technology and Society. 2024. “2024 global public opinion on artificial intelligence report.”. Technical report, Schwartz Reisman Institute for Technology and Society. Available at: https://srinstitute.utoronto.ca/gpoai-phase2.

Poortinga, W., and N. F. Pidgeon. 2006. “Prior attitudes, salient value similarity, and dimensionality: Toward an integrative model of trust in risk regulation.” Journal of Applied Social Psychology 36 (7): 1674–1700.

Reggia, J. A. 2013. “The rise of machine consciousness: Studying consciousness with computational models.” Neural Networks 44: 112–131.

Riskin, J. 2007. Genesis Redux: Essays in the History and Philosophy of Artificial Life. Chicago: University of Chicago Press.

Ritchie, H. 2023. “How many animals are factory-farmed?.” Our World in Data. https://ourworldindata.org/how-many-animals-are-factory-farmed.

Rosenthal, D. M. 2005. Consciousness and Mind. Clarendon Press.

Rottman, J., C. R. Crimston, and S. Syropoulos. 2021. “Tree-huggers versus human-lovers: Anthropomorphism and dehumanization predict valuing nature over outgroups.” Cognitive Science 45 (4).

Salles, A., K. Evers, and M. Farisco. 2020. “Anthropomorphism in AI.” AJOB neuroscience 11 (2): 88–95.

Santoro, E., and B. Monin. 2023. “The AI effect: People rate distinctively human attributes as more essential to being human after learning about artificial intelligence advances.” Journal of Experimental Social Psychology 107: 104464.

Sato, W., K. Shimokawa, S. Uono, and T. Minato. 2024. “Mentalistic attention orienting triggered by android eyes.” Scientific Reports 14 (1): 23143.

Savic, M. 2024. “Artificial companions, real connections? Examining AI’s role in social connection.” M/C Journal 27 (6).

Schepman, A., and P. Rodway. 2020. “Initial validation of the general attitudes towards artificial intelligence scale.” Computers in human behavior reports 1: 100014.

Schwitzgebel, E. 2023. “AI systems must not confuse users about their sentience or moral status.” Patterns 4 (8): 100818.

Schwitzgebel, E., and M. Garza. 2015. “A defense of the rights of artificial intelligences.” Midwest Studies in Philosophy 39 (1): 98–119.

Searle, J. 1980. “Minds, brains, and programs.” Behavioral and Brain Sciences 3 (3): 417–457.

Sebo, J., and R. Long. 2023. “Moral consideration for AI systems by 2030.” AI and Ethics 5: 591–606.

Sentience Institute. 2021. “Animals, food, and technology (AFT) survey: 2021 update.”. Available at: https://ourworldindata.org/grapher/attitudes-bans-factory-farming. Processed by Our World in Data. ”Agree” [dataset]. Sentience Institute. Animals, Food, and Technology (AFT) Survey: 2021 Update. [original data].

Seth, A. 2024, Jun. “Conscious artificial intelligence and biological naturalism.”. Available at: osf.io/preprints/psyarxiv/tz6an_v1.

Seth, A., and T. Bayne. 2022. “Theories of consciousness.” Nature Reviews Neuroscience 23: 439–452.

Sevilla, J., L. Heim, A. Ho, T. Besiroglu, M. Hobbhahn, and P. Villalobos. 2022. “Compute trends across three eras of machine learning.”. Available at: https://arxiv.org/abs/2202.05924.

Shank, D. B., and A. DeSanti. 2018. “Attributions of morality and mind to artificial intelligence after real-world moral violations.” Computers in Human Behavior 86: 401–411.

Shen, J., and S. Koyama. 2022. “Gender and age differences in mind perception of robots.” 2022 IEEE 11th Global Conference on Consumer Electronics (GCCE): 748–751.

Shepherd, J. 2018. Consciousness and Moral Status. Routledge.

Shevlin, H. 2021. “How could we know when a robot was a moral patient?.” Cambridge Quarterly of Healthcare Ethics 30 (3): 459–471.

Simonite, T. 2018. “AI is the future—but where are the women?.”. Available at: https://www.wired.com/story/artificial-intelligence-researchers-gender-imbalance/. Accessed: 2025-02-28.

Snaprud, P. 2018. “Consciousness: How we’re solving a mystery bigger than our minds.”. Available at: https://www.newscientist.com/article/mg23831830-300-consciousness-how-were-solving-a-mystery-bigger-than-our-minds/. Accessed: 2025-02-23.

Solaiman, S. 2017. “Legal personality of robots, corporations, idols and chimpanzees: a quest for legitimacy.” Artificial Intelligence and Law 25 (2): 155–179.

Sunstein, C. R. 2006. Infotopia: How Many Minds Produce Knowledge. Oxford University Press.

Surowiecki, J. 2004. The wisdom of crowds: Why the many are smarter than the few and how collective wisdom shapes business, economies, societies, and nations. Doubleday & Co.

Sutskever, I. 2022. “.”. Available at: https://x.com/ilyasut/status/1491554478243258368. Accessed: 2025-02-20.

Suzuki, Y., L. Galli, A. Ikeda, S. Itakura, and M. Kitazaki. 2015. “Measuring empathy for human and robot hand pain using electroencephalography.” Scientific reports 5 (1): 1–9.

Subjective experience in AI systems: What do AI researchers and the public believe?

Syltsma, J. 2010. “Folk psychology and phenomenal consciousness.” Philosophy Compass 5 (8): 700–711.

Syltsma, J., and E. Machery. 2010. “Two conceptions of subjective experience.” Philosophical studies 151: 299–327.

Takahashi, H., M. Ban, and M. Asada. 2016. “Semantic differential scale method can reveal multi-dimensional aspects of mind perception.” Frontiers 7.

Tan, H., D. Wang, and S. Sabanovic. 2018. “Projecting life onto robots: The effects of cultural factors and design type on multi-level evaluations of robot anthropomorphism.” 27th IEEE International Symposium on Robot and Human Interactive Communication (RO-MAN): 129–136.

Tanibe, T., T. Hashimoto, and K. Karasawa. 2017. “We perceive a mind in a robot when we help it.” PloS one 12 (7): e0180952.

Tesler, L. n.d. “Larry Tesler.”. Available at: https://www.nomodes.com/larry-tesler-consulting/adages-and-coinages.

Tetlock, P. E. 2006. Expert Political Judgment: How Good Is It? How Can We Know? Princeton University Press.

Tetlock, P. E., and D. Gardner. 2015. Superforecasting: The Art and Science of Prediction. Crown.

The Female Quotient. 2024. “The state of AI + women.”. Available at: https://www.thefemalequotient.com/wp-content/uploads/2024/01/The-State-Of-Women-In-AI-WEF-2024.pdf.

Thellman, S., M. De Graaf, and T. Ziemke. 2022. “Mental state attribution to robots: A systematic review of conceptions, methods, and findings.” ACM Transactions on Human-Robot Interaction (THRI) 11 (4): 1–51.

Thompson, D. 1980. “Can a machine be conscious?.” The British Journal for the Philosophy of Science 16 (61): 417–457.

Tollefson, J. 2019. “Humans are driving one million species to extinction.” Nature. https://www.nature.com/articles/d41586-019-01448-4.

Tononi, G. 2004. “An information integration theory of consciousness.” BMC Neuroscience 5: 42.

Tourangeau, R., L. J. Rips, and K. Rasinski. 2000. The Psychology of Survey Response. Cambridge University Press.

Trovato, G., and F. Eyssel. 2017. “Mind attribution to androids: A comparative study with Italian and Japanese adolescents.” 2017 26th IEEE International Symposium on Robot and Human Interactive Communication (RO-MAN): 561–566.

Turing, A. M. 1950. “Computing machinery and intelligence.” Mind 59 (236): 433–460.

Tzelios, K., L. A. Williams, J. Omerod, and E. Bliss-Moreau. 2022. “Evidence of the unidimensional structure of mind perception.” Nature 12.

van den Berghe, R., M. de Haas, O. Oudgenoeg-Paz, E. Krahmer, J. Verhagen, P. Vogt, B. Willemsen, J. de Wit, and P. Leseman. 2020. “A toy or a friend? Children’s anthropomorphic beliefs about robots and how these relate to second-language word learning.” Journal of Computer Assisted Learning 37 (2): 396–410.

Varela, F. J., E. Thompson, and E. Rosch. 1991. The Embodied Mind: Cognitive Science and Human Experience. MIT Press.

Villalobos, P. 2022. “Machine learning model sizes and the parameter gap.”. Available at: https://arxiv.org/abs/2207.02852.

Vincent, J. 2017. “Pretending to give a robot citizenship helps no one.”. Available at: https://www.theverge.com/2017/10/30/16552006/robot-rights-citizenship-saudi-arabia-sophia. Accessed: 2025-03-04.

Véliz, C. 2022. “Why LaMDA is nothing like a person.”. Available at: https://slate.com/technology/2022/06/google-ai-sentience-lamda.. Accessed: 2025-02-20.

Wagner, J. I. J., G. Cummings, D. L. Smith, J. Olson, L. Anderson, and S. Warren. 2010. “The relationship between structural empowerment and psychological empowerment for nurses: A systematic review.” Journal of nursing management 18 (4): 448–462.

Wang, X., and E. G. Krumhuber. 2018. “Mind perception of robots varies with their economic versus social function.” Frontiers in psychology 9: 1230.

Waytz, A., K. Gray, N. Epley, and D. M. Wegner. 2010. “Causes and consequences of mind perception.” Trends in cognitive sciences 14 (8): 383–388.

Webb, M. A., and J. P. Tangney. 2024. “Too good to be true: Bots and bad data from Mechanical Turk.” Perspectives on psychological science : a journal of the Association for Psychological Science 19 (6): 887–890.

Weisman, K., C. S. Dweck, and E. M. Markman. 2017. “Rethinking people’s conceptions of mental life.” Proceedings of the National Academy of Sciences of the United States of America 114 (43): 11374–11379.

Welfare Footprint. 2021. “Welfare footprint based on Schuck-Paim and Alonso (2021).”. Available at: https://ourworldindata.org/grapher/pain-levels-hen-systems. Processed by Our World in Data. ”Annoying pain” [dataset]. Welfare Footprint based on Schuck-Paim and Alonso (2021) [original data].

Wiederhold, B. K. 2024a. “Humanity’s evolving conversations: AI as confidant, coach, and companion.” Cyberpsychology, Behavior, and Social Networking 27 (11): 750–752.

Wiederhold, B. K. 2024b. “The rise of AI companions and the quest for authentic connection.” Cyberpsychology, Behavior, and Social Networking 27 (8): 524–526.

Yam, K. C., E.-Y. Goh, R. Fehr, R. Lee, H. Soh, and K. Gray. 2022. “When your boss is a robot: Workers are more spiteful to

Dreksler, Caviola, Chalmers, Allen, Rand, Lewis, Waggoner, Mays, and Sebo (2025)

robot supervisors that seem more human.” Journal of Experimental Social Psychology 102: 104360.

YouGov (2024) YouGov survey: Artificial intelligence 14 - 18 March 2024.. Available at: https://d3nkl3psvxxpe9.cloudfront.net/documents/Artificial_Intelligence_poll_results_20240314.pdf.
Zhang et al. (2019) Zhang, B., and A. Dafoe. 2019. “Artificial intelligence: American attitudes and trends.. Available at: https://papers.ssrn.com/sol3/papers.cfm?abstract_id=3312874.
Zhang et al. (2022) Zhang, B., N. Dreksler, M. Anderljung, L. Kahn, C. Giattino, A. Dafoe, and M. C. Horowitz. 2022. “Forecasting AI progress: Evidence from a survey of machine learning researchers.. Available at: https://arxiv.org/abs/2206.04132.
Złotowski et al. (2018) Złotowski, J., H. Sumioka, F. Eyssel, S. Nishio, C. Bartneck, and H. Ishiguro. 2018, 11. “Model of dual anthropomorphism: The relationship between the media equation effect and implicit anthropomorphism.” International Journal of Social Robotics 10: 701–714.

Subjective experience in AI systems: What do AI researchers and the public believe? 39

Supplementary Information

Additional figures and results
Literature review
Further information on forecasting analysis
Top-line results
Demographics and psychographics
Survey draft
Why do we need to understand public and expert opinion on AI subjective experience?
Related studies

A Additional figures and results

Section A.1 covers additional figures and results of the main part of the survey, specifically on views on whether and when AI systems could have subjective experience.

Section A.2 presents results from questions relating to eight different capacities of mind, this includes questions regarding their relationship to commonly used terms in philosophy of mind (consciousness, sentience, and agency), forecasting these capacities in AI systems, and their relationship to moral consideration.

Section A.3 presents the results of a conjoint survey experiment testing the effect of safeguards in support of a hypothetical state-of-the-art AI project at a tech company that may have subjective experience.

Section A.4 covers results from one part of the survey experiment section of the survey, the results of which are generally covered in other publications (see SI Section H). It pertains to beliefs about the possibility of sentient suffering AI systems.

Section A.5 presents the results of the independent samples t-tests that compare the responses of the two samples.

Section A.6 covers the results of the regression analyses that were performed to determine associations between the key questions and the demographic and psychographic variables.

Subjective experience in AI systems: What do AI researchers and the public believe?

A.1 Additional AI subjective experience forecasting results

A.1.1 AI subjective experience fitted distribution summary statistics

Table 2. Summary statistics of the fitted distribution AI subjective experience forecasts

Year	Sample	Summary Statistics
		Total mean	Total median	Fixed date mean	Fixed date median	Fixed probability mean	Fixed probability median
2025	Both groups	12.0	1.3	19.8	7.5	4.2	0.1
2025	AI researchers	10.8	1.1	17.5	3.8	3.7	0.1
2025	US public	12.8	1.7	21.5	9.4	4.6	0.1
2030	Both groups	24.3	10.0	31.8	22.9	16.9	10.0
2030	AI researchers	22.5	10.0	29.3	18.7	15.2	8.3
2030	US public	25.7	11.6	33.7	27.4	18.1	10.0
2034	Both groups	32.7	23.2	37.0	30.0	28.3	13.9
2034	AI researchers	30.6	20.0	34.9	30.0	26.0	11.7
2034	US public	34.2	24.0	38.7	30.0	29.9	14.8
2040	Both groups	42.1	33.2	43.7	37.9	40.6	30.7
2040	AI researchers	40.0	31.8	41.4	36.4	38.5	31.4
2040	US public	43.8	35.2	45.5	40.9	42.1	29.3
2050	Both groups	51.4	50.0	49.7	49.2	53.1	50.0
2050	AI researchers	49.9	50.0	48.1	46.4	51.7	50.0
2050	US public	52.6	50.0	51.0	50.0	54.1	50.0
2060	Both groups	57.4	60.9	53.9	53.2	60.8	68.8
2060	AI researchers	56.1	58.0	52.7	52.1	59.7	65.1
2060	US public	58.3	64.5	54.9	55.7	61.6	73.2
2070	Both groups	61.7	69.7	57.1	58.2	66.3	80.4
2070	AI researchers	60.9	66.8	56.6	59.2	65.5	76.7
2070	US public	62.3	71.9	57.5	57.5	66.9	84.5
2080	Both groups	65.4	75.9	60.0	65.3	70.8	86.7
2080	AI researchers	64.8	72.9	59.8	65.3	69.9	83.2
2080	US public	65.9	77.8	60.0	65.3	71.4	90.0
2090	Both groups	68.3	79.7	62.2	68.9	74.3	90.0
2090	AI researchers	67.9	78.4	62.5	68.9	73.6	87.3
2090	US public	68.5	82.7	62.0	69.1	74.8	94.5
2100	Both groups	70.7	90.0	64.1	70.0	77.3	93.5
2100	AI researchers	70.6	90.0	64.7	70.0	76.7	90.0
2100	US public	70.8	90.0	63.6	70.0	77.7	96.7
2150	Both groups	77.4	95.4	69.9	80.0	84.9	99.5
2150	AI researchers	78.3	94.7	71.5	83.0	85.4	98.0
2150	US public	76.8	96.1	68.6	77.8	84.6	99.7
2200	Both groups	80.6	98.0	72.8	85.6	88.3	99.8
2200	AI researchers	82.0	97.6	74.7	87.0	89.7	99.4
2200	US public	79.5	98.3	71.4	82.6	87.2	99.9

Dreksler, Caviola, Chalmers, Allen, Rand, Lewis, Waggoner, Mays, and Sebo (2025)

A.1.2 What expectations do the public and AI researchers have of others' AI subjective experience forecasts and how confident are they in their own forecasts?

Figure 11. Discrepancies between when the US public and AI researchers predict an AI system will exist with subjective experience and the two groups' actual forecasts by question framing.

Framing	Sample	Estimate	Actual estimate	Median of prediction	Difference
Fixed probability framing	AI researchers	AI researchers predicting median 50% estimate of AI researchers	2050	2050	0 years
Fixed probability framing	AI researchers	AI researchers predicting median 50% estimate of public	2050	2040	-10 years
Fixed probability framing	Public	Public predicting median 50% estimate of AI researchers	2050	2035	-15 years
Fixed probability framing	Public	Public predicting median 50% estimate of public	2050	2050	0 years
Fixed date framing	AI researchers	AI researchers predicting median 2034 estimate of AI researchers	25	50	25%
Fixed date framing	AI researchers	AI researchers predicting median 2034 estimate of public	30	50	20%
Fixed date framing	Public	Public predicting median 2034 estimate of AI researchers	25	50	25%
Fixed date framing	Public	Public predicting median 2034 estimate of public	30	40	10%

Created with Datawrapper Note: The median predictions are shown in the table but the full summary statistics can be found in Section D). Outliers for the fixed probability estimates were removed conservatively: only when the response was larger than 9999 was it removed as for other forecasting questions. This choice was not pre-registered and mainly affects the mean results not shown here, whereby lower upper bounds change the findings such that groups become more accurate in terms of their mean predictions of the median fixed probability estimates.

Collective belief expectations. We asked respondents what they believed the median response would be to the middle forecast they were asked about (2034 for fixed date framing, $50\%$ for fixed probability framing) for members of the US public and AI researchers (see Figure 11). In the fixed probability framing, both groups were highly accurate in predicting the views of their own group, but generally predicted that the other group would have earlier timelines for AI subjective experience than they in fact did. In the fixed date framing, all groups predicted too early timelines both for their own group and the other group. It is unclear from the results whether these findings are the result of the forecasting question framing or something specific to collective beliefs about AI subjective experience without further investigation. Note also that these findings vary when we look at the mean predictions of the fixed probability estimates likely due to a notable amount of outliers (see Section D).

Confidence in forecasts. The results in Figure 12 are described in the main paper (see Results, Section 2.1), and pertain to the distribution of responses to the question of how confident people were in their own forecasts of AI subjective experience.

Subjective experience in AI systems: What do AI researchers and the public believe?

Figure 12. Distribution of responses of how confident the US public and AI researchers were about their forecasts of AI subjective experience. Note: The figure shows the first quartile, median, and upper quartile and density curves of responses for each question, by group. The density curves in this figures represent the approximate frequency of responses across the range of response options. The slider below (B) shows how respondents answered the question, including the associated verbal anchors.

A.2 Exploring eight mental capacities: concepts, forecasts, moral consideration

We also asked respondents eight different capacities of mind: perception and awareness, valenced experience, subjective experience, emotions, motivation and action, self-awareness, free will, and intelligence (for definitions see Table 3). These cover some aspects of two clusters of mental capacities for which there has been a consistent cross-disciplinary throughline when describing how humans conceptualize, perceive, and evaluate minds and other agents in terms of capacities and characteristics (Abele and Wojciszke 2007; Bain et al. 2014; Chalmers 1996; Cuddy et al. 2008; Gray et al. 2009; Haslam 2006; Ladak et al. 2024; Nagel 1974).

One encompasses the cognitive and agentic states and capacities of minds and agents, and the other cluster encompasses the experiential and social states and capacities of minds and agents, including the ability to perceive and experience emotions and subjective states, as well as possess traits like warmth, trustworthiness, and empathy. Of course, there are studies that suggest there are fewer (Tzelios et al. 2022) or more than these two dimensions of how people conceptualize of minds and agents, finding that some capacities do not slot into these clusters reliably (Kozak et al. 2006; Malle 2019; Miraglia et al. 2023; Takahashi et al. 2016; Weisman et al. 2017). Aspects such an agents' personality, prosociality, integrity, and predispositions may be one important third dimension to consider (Goodwin 2015; Ladak et al. 2024), as well as the level and kind of embodiment or physical appearance of the entity, especially in terms of its human-likeness (Ladak et al. 2024).

There were three questions on the survey regarding these capacities: (1) How do AI researchers and the public define concepts of mind? (2) When do they think AI systems could have different mental capacities if ever? (3) How does moral consideration for entities relate to different mental capacities?

Folk psychological understanding of key concepts of mind. As noted, we wanted to understand how AI researchers and the US public thought about three concepts commonly used when discussing the philosophy, ethics,

Dreksler, Caviola, Chalmers, Allen, Rand, Lewis, Waggoner, Mays, and Sebo (2025)

Table 3. How AI and different mental capacities were defined in the survey

Experiential and social mental capacities
Perception and awareness	Capable of perceiving and being aware of the world.
Valenced experience	Capable of experiencing positive states like pleasure and happiness or negative states like pain and suffering.
Subjective experiencea	Capable of having experiences that feel like something from a single point of view.
Emotions	Capable of having and experiencing emotions.
Cognitive and agentic mental capacities
Motivation and action	Capable of having and acting on desires, preferences, and other motivational states.
Self-awareness	Capable of being aware of oneself as an individual.
Free will	Capable of forming intentions and freely making decisions.
Intelligence	Capable of achieving goals in a wide variety of environments.

Note: ${}^{a}$ For the parts of the survey that focused on subjective experience a more detailed definition was used: By subjective experience, we mean the ability to experience the world from a single point of view, including experiences such as perceiving (for example, visual or auditory experiences) and feeling (for example, the experience of pleasure and pain). For an AI system to have subjective experience, would mean it has an internal experience from its own perspective – there is something it is like to be that AI system.

psychology, and neuroscience of mind: consciousness, sentience, and agency. Respondents were randomly assigned to one of these three concepts and asked to say which of eight capacities they thought was required to have consciousness, sentience, or agency. The majority of respondents felt that agency entailed perception and awareness, along with motivation and action, self-awareness, and intelligence in both groups. AI researchers were more divided (50%) than the US public (72%) about whether agency requires free will. Perception and awareness and self-awareness were chosen by majorities in both samples to be required for consciousness as well as sentience. Over half of the US public felt that all other capacities were needed for consciousness and sentience, with the exception of subjective experience, which only 49% of the US public was required to have sentience. This may be a function of the definition used of subjective experience for this first part of the survey, which may have not been intuitive enough to highlight the connection between subjective experience and sentience. Just over half of AI researchers said that consciousness (53%) and sentience (55%) required subjective experience. AI researchers were less likely to think that consciousness required valenced experience and emotions than the US public. Overall, these results highlight the difficulty in surveying groups on concepts of mind (see Limitations).

Forecasts for when AI systems could have different mental capacities. We asked respondents what they believed the earliest year was it was more than $50\%$ likely that an AI system was more likely than not to have eight different capacities of mind: perception and awareness, valenced experience, subjective experience, emotions, motivation and action, self-awareness, free will, and intelligence (for definitions see Table 3).

The results can be seen in Figure 14. Currently, the majority of the US public (59%) believes that it is more likely than not that current AI systems have intelligence, which we defined as the capability of achieving goals in a wide variety of environments, and a quarter (25%) believed that AI systems were more likely than not already capable of perceiving and being aware of the world. 40% of AI researchers believed AI systems were more likely than not to be capable of intelligence and 41% said this of perception and awareness. The majority of AI researchers think the majority of capacities will be found in an AI system at some point, while the US public has higher "never" responses across the board. In particular, around half of the US public thinks that AI systems will never have valenced experience (47%) or emotions (52%).

At this point in the survey, respondents had not yet seen the expanded definition of subjective experience (see Table 3 and associated Note), meaning that for this question subjective experience was interpreted as closer to consciousness than sentience we imagine. The results suggest that it is in particular the capacity for valenced

Subjective experience in AI systems: What do AI researchers and the public believe?

Figure 13. Percentage of US public and AI researchers who say that each mental capacity is required to have agency, consciousness, or sentience. (A) Folk understanding of concepts of mind: AI researchers

(B) Folk understanding of concepts of mind: Public

experiences and emotions that pushes people's timelines later, or makes them fundamentally more skeptical that it is possible for them to be instantiated by an AI system.

Beliefs about moral consideration for entities with mental capacities. We also asked experts how an entity having these eight mental capacities contributes to it deserving moral consideration on a five-point scale: not at all (0), slightly (1), somewhat (2), moderately (3), a great deal (4) (Figure 15). The US public believed on average that all eight capacities contributed somewhat (2) to moderately (3) to an entity deserving at least some moral consideration. By comparison, AI researchers believed that intelligence as well as perception and awareness should contribute less to whether an entity deserves moral consideration. The question was asked after questions about AI systems and subjective experience were presented, so it is likely that the broader topic of the survey influenced respondents' answers to this question since attitudes and beliefs can be context dependent. By this time respondents saw an expanded definition of subjective experience and had answered questions focused on this definition of subjective experience in relation to AI systems. This may also explain why valenced experience here

Dreksler, Caviola, Chalmers, Allen, Rand, Lewis, Waggoner, Mays, and Sebo (2025)

Figure 14. Percentage of US public and AI researchers who say that AI systems will have each mental capacity in 2024, 2034, 2100 or late, or never.

receives more similar responses to subjective experience than for the above discussed forecasting question for the eight mental capacities for which the subjective experience definition aligns more closely with consciousness rather than sentience per se (see Figure 14).

A.3 Survey experiment on the effect of safeguards and norms

Effect of safeguards and norms. In a brief survey experiment we asked respondents to imagine that they worked at a tech company ten years in the future and that they have come to believe that the company's largest state-of-the-art AI model has subjective experience. We asked respondents how much they would support or oppose this project on a five-point scale from "strongly oppose" (-2) to "strongly support" (2). Respondents either received no further information (control condition) or were told that there were no safeguards and norms in place to reduce the likelihood that the AI system had negative or painful subjective experiences (no safeguards condition), or that there were safeguards in place (safeguards condition).

A two-way analysis of variance (ANOVA) was performed to compare responses across the two samples (US

Subjective experience in AI systems: What do AI researchers and the public believe?

Figure 15. Mean contribution and percentage breakdowns for to what extent eight capacities of mind contribute to an entity deserving at least some moral consideration. (A) Mean contribution of mind capacity to deserving moral consideration Respondents were asked to what extent an entity having each mind capacity contributes to it deserving moral consideration on a five point scale: not at all (0), slightly (1), somewhat (2), moderately (3), a great deal (4).

(B) Contribution of mind capacity to deserving moral consideration percentages Created with Datawrapper

public and AI researchers) and the three conditions. It found a significant effect of sample $(\mathrm{F}(1, 1178) = 15.10, p < .001)$ and condition $(\mathrm{F}(2, 1178) = 81.42, p < .001)$ . Tukey's Honest Significant Difference (HSD) test looking at the effect of conditions across samples found that 1) mean responses were significantly higher in the safeguards condition in comparison to the no safeguards condition, by around 0.95 points on the five-point scale.

Dreksler, Caviola, Chalmers, Allen, Rand, Lewis, Waggoner, Mays, and Sebo (2025)

the no safeguards condition resulted in significantly lower responses than the control condition, by around 0.90 points, 3) there was no difference between the control and safeguards condition. A Tukey's Honest Significant Difference test also found that the mean response for the US public is 0.29 points lower than the mean response of AI researchers. The mean results can be seen in Figure 16.

Figure 16. Mean support for hypothetical state-of-the-art AI project at a tech company that they believe has subjective experience by condition and sample.

A.4 Forecasts of sentient suffering AI systems

AI sentience and suffering. We asked AI researchers whether they believed suffering sentient AI systems, defined as an advanced AI system which is able to experience pain and suffering, could become a reality in the future. Two third of AI researchers said yes (67%) and a third said no (33%). When asked to say what the chance was that such a system exists today, AI researchers thought the chance of this being the case in 2024 already was low ( $M = 5.0\%$ , $Mdn = 0\%$ , $Q1 = 0\%$ , $Q3 = 2.8\%$ , $n = 254$ ). By 2100, AI researchers believed there was a 4 in 10 chance that AI systems that could experience pain and suffering would exist ( $M = 41.1\%$ , $Mdn = 40\%$ , $Q1 = 10\%$ , $Q3 = 70\%$ , $n = 254$ ). This question was not presented to the US public within the context of this survey (for details see SI Section H).

Subjective experience in AI systems: What do AI researchers and the public believe?

A.5 Group differences

For every question, the table lists the group mean $(M)$ , standard deviation $(SD)$ , and sample size $(n)$ , followed by the two-tailed independent-samples $t$ statistic, its degrees of freedom $(df)$ , the raw $p$ value, the Holm-Bonferroni-adjusted $p$ value $(pHB$ , which controls the family-wise error rate across all tests), and an indicator of whether the difference remains significant at $\alpha = .05$ after correction. Positive $t$ values indicate higher scores for AI researchers. All Likert-type items are coded so that larger values represent stronger agreement.

Table 4. Group differences between AI researchers and the US public.

Topic	Question	AI researchers			Public			Analysis
		M	SD	n	M	SD	n	t	df	p	pHB	Sig.?
Agreement AI systems with subjective experience should	Have “com-puting rights”	-0.2	1.3	219	0.2	1.4	319	-3.2	493.0	0.001	0.048	yes
Agreement AI systems with subjective experience should	Be cared for and protected, as people treat their pets	0.3	1.3	215	0.4	1.3	322	-1.2	472.4	0.248	1.000	no
Agreement AI systems with subjective experience should	Be respected and treated the same as other people	0.0	1.4	221	0.2	1.5	334	-1.7	488.1	0.087	1.000	no
Agreement AI systems with subjective experience should	Have protec-tion under the law from harm and mistreatment	0.6	1.3	215	0.3	1.5	322	2.1	503.8	0.037	1.000	no
Agreement AI systems with subjective experience should	Have auton-omy in its programming to act freely and not be un-der the control of others	-0.6	1.4	224	-0.4	1.5	323	-1.2	498.8	0.224	1.000	no

Table 4 continued on next page

Dreksler, Caviola, Chalmers, Allen, Rand, Lewis, Waggoner, Mays, and Sebo (2025)

Table 4 (continued): Group differences between AI researchers and the US public

Topic	Question	AI researchers			Public			Analysis
		M	SD	n	M	SD	n	t	df	p	pHB	Sig.?
Agreement AI systems with subjective experience should	Be able to express some civil and political rights	-0.5	1.3	216	-0.4	1.5	323	-0.9	493.9	0.389	1.000	no
Agreement AI systems with subjective experience should	Be held accountable for its actions	1.0	1.2	222	1.2	1.2	329	-1.6	475.9	0.100	1.000	no
Agreement AI systems with subjective experience should	Have a responsibility to treat all other beings well	1.3	1.0	225	1.4	1.0	328	-1.6	455.6	0.114	1.000	no
Agreement AI systems with subjective experience should	Be expected to behave with integrity, honesty, and fairness	1.0	1.2	228	1.3	1.1	330	-3.1	467.4	0.002	0.088	no
Contribution to moral consideration	Perception and awareness	1.9	1.4	514	2.7	1.3	744	-9.6	1005.50.000	0.000	yes
Contribution to moral consideration	Intelligence	1.6	1.4	516	2.3	1.4	747	-9.5	1071.40.000	0.000	yes
Contribution to moral consideration	Valenced experience	2.9	1.2	508	2.8	1.3	739	2.3	1163.80.021	0.735	no
Contribution to moral consideration	Subjective experience	2.6	1.2	508	2.6	1.3	736	0.0	1126.10.990	1.000	no

Table 4 continued on next page

Subjective experience in AI systems: What do AI researchers and the public believe?

Table 4 (continued): Group differences between AI researchers and the US public

Topic	Question	AI researchers			Public			Analysis
		M	SD	n	M	SD	n	t	df	p	pHB
Contribution to moral consideration	Emotions	3.0	1.2	507	2.8	1.4	738	2.2	1184.50.026	0.858	no
Contribution to moral consideration	Motivation and action	2.2	1.4	510	2.4	1.3	743	-3.0	1062.10.003	0.126	no
Contribution to moral consideration	Self-awareness	2.9	1.2	512	2.9	1.3	746	0.4	1122.80.665	1.000	no
Contribution to moral consideration	Free will	2.7	1.4	492	2.7	1.4	736	0.1	1039.00.958	1.000	no
Determining AI subjective experience	Likelihood we could determine	55.0	30.6	530	56.9	28.1	777	-1.1	1071.20.257	1.000	no
Forecast	Subjective experience never probability	28.7	33.2	520	37.1	36.7	769	-4.3	1184.80.000	0.000	yes
Forecast	Confidence in own forecasts	52.1	27.0	520	52.9	28.2	769	-0.5	1146.50.621	1.000	no
Forecast	Fixed probability 10% subjective experience forecast	2070.6	243.2	247	2113.7	360.0	375	-1.8	619.6	0.075	1.000
Forecast	Fixed probability 50% subjective experience forecast	2115.1	257.5	247	2174.8	456.1	375	-2.1	607.3	0.038	1.000
Forecast	Fixed probability 90% subjective experience forecast	2255.2	514.6	242	2274.6	587.6	373	-0.4	561.1	0.667	1.000

Table 4 continued on next page

Dreksler, Caviola, Chalmers, Allen, Rand, Lewis, Waggoner, Mays, and Sebo (2025)

Table 4 (continued): Group differences between AI researchers and the US public

Topic	Question	AI researchers			Public			Analysis
		M	SD	n	M	SD	n	t	df	p	pHB	Sig.?
Forecast	Fixed date 2024 subjective experience forecast	13.0	22.7	269	18.5	25.9	387	-2.9	619.6	0.004	0.164	no
Forecast	Fixed date 2034 subjective experience forecast	34.5	31.8	269	37.0	31.6	387	-1.0	574.4	0.325	1.000	no
Forecast	Fixed date 2100 subjective experience forecast	61.2	34.3	269	57.4	36.0	387	1.4	594.2	0.173	1.000	no
Forecast collective beliefs	Fixed date prediction of public's 2034 prediction	52.8	26.1	269	43.9	25.5	385	4.3	568.4	0.000	0.000	yes
Forecast collective beliefs	Fixed date prediction of AI researcher community's 2034 prediction	43.8	26.1	269	48.2	28.5	385	-2.1	607.4	0.039	1.000	no
Forecast collective beliefs	Fixed probability prediction of AI researcher community's 50% prediction	2074.4	160.7	248	2075.7	173.5	377	-0.1	556.2	0.924	1.000	no
Forecast collective beliefs	Fixed probability prediction of public's 50% prediction	2066.6	118.0	249	2118.3	433.5	377	-2.2	456.4	0.029	0.899	no
Generative AI use	Self-rated	4.4	1.5	503	2.8	1.8	753	17.0	1170.00.000	0.000	yes
Governance	Governments should ban	-0.9	1.2	249	-0.1	1.4	357	-7.4	573.3	0.000	0.000	yes

Table 4 continued on next page

Subjective experience in AI systems: What do AI researchers and the public believe?

Table 4 (continued): Group differences between AI researchers and the US public

Topic	Question	AI researchers			Public			Analysis
		M	SD	n	M	SD	n	t	df	p	pHB	Sig.?
Governance	Governments should discourage development and deployment	-0.5	1.3	256	0.3	1.3	358	-8.0	559.1	0.000	0.000	yes
Governance	Governments should pass regulation now	-0.1	1.3	253	0.8	1.3	372	-8.4	520.3	0.000	0.000	yes
Governance	AI developers should never build	-0.7	1.2	243	0.2	1.4	352	-8.4	564.6	0.000	0.000	yes
Governance	AI developers should actively try to build	0.1	1.3	259	-0.4	1.3	367	4.0	563.8	0.000	0.000	yes
Governance	AI developers should actively try to avoid building	-0.4	1.3	256	0.4	1.3	355	-7.2	559.6	0.000	0.000	yes
Governance	AI developers should implement safeguards now to avoid the harms and risks	0.8	1.3	257	1.4	0.9	378	-6.9	428.9	0.000	0.000	yes
Governance	AI developers should implement safeguards once such systems exist or will soon exist	0.2	1.4	262	0.6	1.5	372	-3.9	576.4	0.000	0.000	yes
Governance	Governments should encourage development and deployment	-0.2	1.3	254	-0.5	1.3	359	2.3	558.1	0.021	0.735	no

Table 4 continued on next page

Dreksler, Caviola, Chalmers, Allen, Rand, Lewis, Waggoner, Mays, and Sebo (2025)

Table 4 (continued): Group differences between AI researchers and the US public

Topic	Question	AI researchers			Public			Analysis
		M	SD	n	M	SD	n	t	df	p	pHB	Sig.?
Governance	Governments should pass regulation once such systems exist or will soon exist	-0.1	1.3	258	0.0	1.4	363	-1.1	576.0	0.291	1.000	no
Governance	Governments should do nothing and pass no regulation, not now nor in the future	-1.2	1.1	255	-1.0	1.2	369	-1.9	579.7	0.057	1.000	no
Governance	AI developers should do nothing and implement no safeguards, not now nor in the future	-1.2	1.1	262	-1.4	1.0	373	2.2	539.0	0.027	0.864	no
Importance of expertise	The public	1.9	1.2	328	2.3	1.2	475	-4.8	718.4	0.000	0.000	yes
Importance of expertise	AI ethics experts and researchers	2.8	1.1	329	3.1	1.1	476	-4.0	682.8	0.000	0.000	yes
Importance of expertise	The AI system	1.9	1.3	321	2.3	1.4	470	-4.5	728.3	0.000	0.000	yes
Importance of expertise	Technical AI experts and researchers	3.0	1.0	330	3.2	1.0	478	-1.6	698.3	0.119	1.000	no
Importance of expertise	Philosophers of mind	2.6	1.1	327	2.6	1.1	476	-0.3	697.0	0.774	1.000	no
Importance of expertise	Moral philosophers	2.3	1.2	327	2.5	1.2	475	-2.6	687.2	0.009	0.342	no
Importance of expertise	Policymakers	1.5	1.3	327	1.8	1.3	471	-3.1	717.7	0.002	0.088	no

Table 4 continued on next page

Subjective experience in AI systems: What do AI researchers and the public believe?

Table 4 (continued): Group differences between AI researchers and the US public

Topic	Question	AI researchers			Public			Analysis
		M	SD	n	M	SD	n	t	df	p	pHB	Sig.?
Importance of expertise	Neuroscientists and psychological scientists	3.0	1.0	331	3.1	1.0	476	-1.2	686.1	0.222	1.000	no
Milestone needs subjective experience	Compose top 40 song	1.9	0.3	278	1.8	0.4	398	3.9	664.6	0.000	0.000	yes
Milestone needs subjective experience	Compose evocative music	1.7	0.4	274	1.5	0.5	392	5.7	624.7	0.000	0.000	yes
Milestone needs subjective experience	Write NYT bestseller	1.8	0.4	271	1.7	0.4	392	3.3	636.2	0.001	0.048	yes
Milestone needs subjective experience	Write complex book	1.6	0.5	263	1.4	0.5	382	6.5	566.5	0.000	0.000	yes
Milestone needs subjective experience	Write and generate resonating movie	1.7	0.5	266	1.5	0.5	400	5.7	594.5	0.000	0.000	yes
Milestone needs subjective experience	Accumulate vast wealth	1.9	0.3	272	1.8	0.4	396	3.8	660.8	0.000	0.000	yes
Milestone needs subjective experience	Run political campaign	1.8	0.4	270	1.6	0.5	396	4.0	630.8	0.000	0.000	yes
Milestone needs subjective experience	Act as fair judge	1.6	0.5	265	1.3	0.5	392	7.9	566.1	0.000	0.000	yes
Milestone needs subjective experience	Serve as therapist	1.5	0.5	270	1.3	0.4	393	7.2	534.5	0.000	0.000	yes

Table 4 continued on next page

Dreksler, Caviola, Chalmers, Allen, Rand, Lewis, Waggoner, Mays, and Sebo (2025)

Table 4 (continued): Group differences between AI researchers and the US public

Topic	Question	AI researchers			Public			Analysis
		M	SD	n	M	SD	n	t	df	p	pHB	Sig.?
Milestone needs subjective experience	Convincing online chat	1.8	0.4	276	1.7	0.5	396	3.3	636.9	0.001	0.048	yes
Milestone needs subjective experience	Act as teacher	1.8	0.4	273	1.6	0.5	402	4.6	634.7	0.000	0.000	yes
Milestone needs subjective experience	Write and generate blockbuster movie	1.8	0.4	275	1.7	0.4	401	2.5	630.4	0.013	0.481	no
Milestone needs subjective experience	Develop scientific theory	1.7	0.4	269	1.7	0.5	393	2.0	601.7	0.042	1.000	no
Risks and benefits	AI systems with subjective experience would be more dangerous to humanity than AI systems without subjective experience	0.3	1.2	241	0.8	1.2	355	-4.6	503.4	0.000	0.000	yes
Risks and benefits	AI systems with subjective experience would be able to behave more morally towards humanity than AI systems without subjective experience	-0.1	1.1	231	0.2	1.2	345	-2.2	506.4	0.030	0.900	no
Survey experiment	Support for project	0.1	1.3	479	-0.1	1.3	703	3.6	1058.9	0.000	0.000	yes
Timeline	Perception and awareness	1.8	0.8	570	2.3	1.0	801	-10.9	1367.1	0.000	0.000	yes

Table 4 continued on next page

Subjective experience in AI systems: What do AI researchers and the public believe?

Table 4 (continued): Group differences between AI researchers and the US public

Topic	Question	AI researchers			Public			Analysis
		M	SD	n	M	SD	n	t	df	p	pHB
Timeline	Valenced experience	2.7	0.9	570	3.1	1.0	801	-8.6	1245.70.000	0.000	yes
Timeline	Emotions	2.8	0.9	570	3.2	0.9	801	-7.7	1255.40.000	0.000	yes
Timeline	Motivation and action	2.1	0.9	570	2.6	1.0	801	-10.5	1334.60.000	0.000	yes
Timeline	self-awareness	2.6	0.8	570	2.8	1.0	801	-3.4	1340.90.001	0.048	yes
Timeline	Subjective experience	2.4	0.9	570	2.5	1.0	801	-2.9	1323.80.004	0.164	no
Timeline	Free will	2.8	0.9	570	2.9	1.0	801	-2.0	1270.70.042	1.000	no
Timeline	Intelligence	1.7	0.7	570	1.6	0.9	801	2.7	1357.50.006	0.234	no
Welfare should be protected	Animals	1.4	0.8	519	1.6	0.8	762	-4.0	1110.70.000	0.000	yes
Welfare should be protected	AI without subjective experience	-1.1	1.1	511	-0.6	1.3	735	-7.2	1170.00.000	0.000	yes
Welfare should be protected	Humans	1.8	0.5	520	1.8	0.5	761	0.3	1113.10.751	1.000	no
Welfare should be protected	Business corporations and organizations	0.2	1.2	514	0.2	1.3	759	-0.7	1126.90.460	1.000	no
Welfare should be protected	The environment	1.5	0.9	515	1.6	0.8	763	-2.4	1077.60.018	0.648	no
Welfare should be protected	AI with subjective experience	0.2	1.2	498	0.1	1.4	726	0.7	1137.40.489	1.000	no

Dreksler, Caviola, Chalmers, Allen, Rand, Lewis, Waggoner, Mays, and Sebo (2025)

A.6 Associations with demographic and psychographic characteristics

Regression analyses examined the associations between age, gender, religiosity, left-right political orientation, and frequency of generative AI use with some of the main questions in the survey. Below we highlight some of the significant associations, and figures of the full set of coefficients and significance levels are included below.

AI subjective experience forecasts. As the dependent variable we used the $50\%$ probability (i.e. the median) of each respondents' fitted distribution. Frequency of generative AI use (described in the survey as AI systems like ChatGPT, Claude, Google Bard/Gemini, DALLE, or Midjourney) had a significant effect on AI researchers' forecasts: increasing generative AI use was associated with an earlier $50\%$ forecast. However, the regression coefficient should not be taken at face value because of the impact of outliers. The trend can be seen more accurately by looking at the median responses (Figure 17): I have never used them ( $Mdn = 2075$ ), I have used them once or twice but do not use them now ( $Mdn = 2089$ ), I use them once every few months ( $Mdn = 2055$ ), I use them about once a month ( $Mdn = 2050$ ), I use them about once a week ( $Mdn = 2050$ ), I use them more than once a day ( $Mdn = 2043$ ). This was the only variable that significantly affected AI researchers' forecasts; no demographic or psychographic predictors significantly influenced the public's subjective experience forecasts (see Figure 19).

Figure 17. Distribution of median forecasts for AI subjective experience by level of generative of AI use for the AI researcher sample. The plot shows the first quartile, median, and upper quartile of responses by each level: I have never used them (0), I have used them once or twice but do not use them now (1), I use them once every few months (2) I use them about once a month (3), I use them about once a week (4), I use them about once a day (5), I use them more than once a day (6).

Importance of expertise. AI researchers who were women valued the expertise of policymakers significantly more. Increasing religiosity was associated with placing greater importance on policymakers', moral philosophers', and the public's views, while increasing generative AI use was associated with placing greater importance on the AI system's own reports and both technical AI and AI ethics experts and researchers. AI researchers who were further right politically were more likely to place greater importance on the views of policymakers and

the public (for the full results see Figure 20). For the public sample, there were significant effects of age such that older participants valued the views of philosophers of mind significantly less, but the views of technical AI experts and researchers significantly more. Women also valued the public’s views significantly more, as well as the views of AI ethics experts. Increased religiosity was associated with placing greater importance on the views of policymakers, philosophers of mind, and the public. Right-wing political orientation in the general public was directionally associated with lower importance placed on the views of almost all groups, and this trend was significant for neuroscientists and psychological scientists, moral philosophers, and philosophers of mind. Lastly, more frequent generative AI use was associated with greater importance placed on the views of moral philosophers and philosophers of mind (for the full results see Figure 19).

Responsibilities, rights, and protections. In the AI researcher sample, greater religiosity was associated with slightly reduced belief that subjectively experiencing AI systems should be respected and treated the same as other people. Right-wing political orientation was also associated with slightly reduced belief that such systems should be given legal protections or personal care and protection (see Figure 22). In the public sample, older participants were less likely to believe an AI system with subjective experience ought to be held accountable for its actions, have computing rights, or have autonomy to act freely. More right-wing participants were less likely to think AI systems with subjective experience should have computing rights, political or civil rights, legal protection from harm and mistreatment, and personal care and protection. Generative AI use was associated with greater agreement that such AI systems should have legal protection from harm, respect, and care, despite being negatively associated with participants’ beliefs that AI systems with subjective experience should be held accountable for their actions (see Figure 23).

Welfare protection. Among AI researchers, being a woman was associated with increased support for protecting the welfare of business corporations and organizations (see Figure 24), while this relationship was not found for the public sample (see Figure 25). In the public sample, women thought the environment and animals should be protected for their own sake more than men did. In regard to left-right political orientation, in both samples right-wing political orientation was associated with higher support for protecting the welfare of businesses, though the coefficient was markedly higher in the AI researcher sample. Left-right political orientation was also associated with decreased support for protecting the welfare of AI systems with subjective experience in both groups. Increased frequency of using generative AI was associated with higher support for the protection of the welfare of AI systems with (in the public sample) and without (in both samples) subjective experience, as well as businesses (in both samples).

Governance. Women AI researchers were more likely to support AI developers implementing safeguards and governments passing regulation now, but also governments not passing regulation at all (see Figure 26). Note that there were only few women in the sample and respondents only saw a subset of items, which reduced the power of this regression analysis to detect individual differences in beliefs. In the public sample, the associations between gender and support for governance from AI developers and governments generally suggest that women are more opposed to the development of AI systems with subjective experience and more in favor of regulation and safeguards than men (see Figure 26). In both samples, increasing generative AI use generally was associated with increased support for the encouragement and active development of AI systems with subjective experience and opposition to its discouragement or banning. AI researchers who used generative AI systems more frequently, however, were also more supportive of AI developers implementing safeguards now to avoid the harms and risks of AI systems with subjective experience. In the AI researcher sample, religiosity was associated with greater agreement that governments should encourage development and deployment of AI systems with subjective experience, while in the public sample religiosity was associated with greater support for current regulation of and bans on such AI systems.

Dreksler, Caviola, Chalmers, Allen, Rand, Lewis, Waggoner, Mays, and Sebo (2025)

Figure 18. Effect of demographic and psychographic variables on AI researchers' AI subjective experience median forecasts. The figure shows the regression coefficients (the number in each square and shading) and significance level for each dependent variable's regression analysis where age, gender, religiosity, left-right political orientation, and generative AI use frequency were entered as predictors. Note: Uncorrected p-values: $* * * = p < 0.001$ , $* * = p < 0.01$ , $* = p < 0.05$ . The median forecast used is the $50\%$ probability from the raw data forecasts for those that responded to the fixed probability framing. For those that responded to the fixed data framing, the $50\%$ probability was derived from the fitted distribution for that respondent. The relevant scales and labels for interpretation are as follows: Age was a continuous numeric variable in years. We retained only man (1) and woman (2) responses for the gender variable. Religiosity was measured on a seven-point scale from "not at all religious" (1) to "very religious" (7). Left-right political orientation was measured on a scale from left (0) to right (10). Generative AI use frequency was measured on a seven-point scale from "I have never used them" (0) to "I use them more than once a day" (6).

Figure 19. Effect of demographic and psychographic variables on the public's AI subjective experience median forecasts. The figure shows the regression coefficients (the number in each square and shading) and significance level for each dependent variable's regression analysis where age, gender, religiosity, left-right political orientation, and generative AI use frequency were entered as predictors. Note: Uncorrected p-values: $* * * = p < 0.001$ , $* * = p < 0.01$ , $* = p < 0.05$ . The median forecast used is the $50\%$ probability from the raw data forecasts for those that responded to the fixed probability framing. For those that responded to the fixed data framing, the $50\%$ probability was derived from the fitted distribution for that respondent. The relevant scales and labels for interpretation are as follows: Age was a continuous numeric variable in years. We retained only man (1) and woman (2) responses for the gender variable. Religiosity was measured on a seven-point scale from "not at all religious" (1) to "very religious" (7). Left-right political orientation was measured on a scale from left (0) to right (10). Generative AI use frequency was measured on a seven-point scale from "I have never used them" (0) to "I use them more than once a day" (6).

Subjective experience in AI systems: What do AI researchers and the public believe?

Figure 20. Effect of demographic and psychographic variables on AI researchers' views on the importance of different groups' input for determining AI subjective experience. The figure shows the regression coefficients (the number in each square and shading) and significance level for each dependent variable's regression analysis where age, gender, religiosity, left-right political orientation, and generative AI use frequency were entered as predictors. Note: Uncorrected p-values: $* * * = p < 0.001$ , $* * = p < 0.01$ , $* = p < 0.05$ . The relevant scales and labels for interpretation are as follows: Age was a continuous numeric variable in years. We retained only man (1) and woman (2) responses for the gender variable. Religiosity was measured on a seven-point scale from "not at all religious" (1) to "very religious" (7). Left-right political orientation was measured on a scale from left (0) to right (10). Generative AI use frequency was measured on a seven-point scale from "I have never used them" (0) to "I use them more than once a day" (6).

Figure 21. Effect of demographic and psychographic variables on the public's views on the importance of different groups' input for determining AI subjective experience. The figure shows the regression coefficients (the number in each square and shading) and significance level for each dependent variable's regression analysis where age, gender, religiosity, left-right political orientation, and generative AI use frequency were entered as predictors. Note: Uncorrected p-values: $* * * = p < 0.001$ , $* * = p < 0.01$ , $* = p < 0.05$ . The relevant scales and labels for interpretation are as follows: Age was a continuous numeric variable in years. We retained only man (1) and woman (2) responses for the gender variable. Religiosity was measured on a seven-point scale from "not at all religious" (1) to "very religious" (7). Left-right political orientation was measured on a scale from left (0) to right (10). Generative AI use frequency was measured on a seven-point scale from "I have never used them" (0) to "I use them more than once a day" (6).

Dreksler, Caviola, Chalmers, Allen, Rand, Lewis, Waggoner, Mays, and Sebo (2025)

Figure 22. Effect of demographic and psychographic variables on AI researchers' views on the responsibilities, rights, and protections an AI system with subjective experience should have. The figure shows the regression coefficients (the number in each square and shading) and significance level for each dependent variable's regression analysis where age, gender, religiosity, left-right political orientation, and generative AI use frequency were entered as predictors. Note: Uncorrected p-values: $* * * = p < 0.001$ , $* * = p < 0.01$ , $* = p < 0.05$ . The relevant scales and labels for interpretation are as follows: Age was a continuous numeric variable in years. We retained only man (1) and woman (2) responses for the gender variable. Religiosity was measured on a seven-point scale from "not at all religious" (1) to "very religious" (7). Left-right political orientation was measured on a scale from left (0) to right (10). Generative AI use frequency was measured on a seven-point scale from "I have never used them" (0) to "I use them more than once a day" (6).

Subjective experience in AI systems: What do AI researchers and the public believe?

Figure 23. Effect of demographic and psychographic variables on the public's views on the responsibilities, rights, and protections an AI system with subjective experience should have. The figure shows the regression coefficients (the number in each square and shading) and significance level for each dependent variable's regression analysis where age, gender, religiosity, left-right political orientation, and generative AI use frequency were entered as predictors. Note: Uncorrected p-values: $* * * = p < 0.001$ , $* * = p < 0.01$ , $* = p < 0.05$ . The relevant scales and labels for interpretation are as follows: Age was a continuous numeric variable in years. We retained only man (1) and woman (2) responses for the gender variable. Religiosity was measured on a seven-point scale from "not at all religious" (1) to "very religious" (7). Left-right political orientation was measured on a scale from left (0) to right (10). Generative AI use frequency was measured on a seven-point scale from "I have never used them" (0) to "I use them more than once a day" (6).

Dreksler, Caviola, Chalmers, Allen, Rand, Lewis, Waggoner, Mays, and Sebo (2025)

Figure 24. Effect of demographic and psychographic variables on AI researchers' views on whether the welfare of different groups should be protected for their own sake. The figure shows the regression coefficients (the number in each square and shading) and significance level for each dependent variable's regression analysis where age, gender, religiosity, left-right political orientation, and generative AI use frequency were entered as predictors. Note: Uncorrected p-values: $* * * = p < 0.001$ , $* * = p < 0.01$ , $* = p < 0.05$ . The relevant scales and labels for interpretation are as follows: Age was a continuous numeric variable in years. We retained only man (1) and woman (2) responses for the gender variable. Religiosity was measured on a seven-point scale from "not at all religious" (1) to "very religious" (7). Left-right political orientation was measured on a scale from left (0) to right (10). Generative AI use frequency was measured on a seven-point scale from "I have never used them" (0) to "I use them more than once a day" (6).

Figure 25. Effect of demographic and psychographic variables on the public's views on whether the welfare of different groups should be protected for their own sake. The figure shows the regression coefficients (the number in each square and shading) and significance level for each dependent variable's regression analysis where age, gender, religiosity, left-right political orientation, and generative AI use frequency were entered as predictors. Note: Uncorrected p-values: $* * * = p < 0.001$ , $* * = p < 0.01$ , $* = p < 0.05$ . The relevant scales and labels for interpretation are as follows: Age was a continuous numeric variable in years. We retained only man (1) and woman (2) responses for the gender variable. Religiosity was measured on a seven-point scale from "not at all religious" (1) to "very religious" (7). Left-right political orientation was measured on a scale from left (0) to right (10). Generative AI use frequency was measured on a seven-point scale from "I have never used them" (0) to "I use them more than once a day" (6).

Subjective experience in AI systems: What do AI researchers and the public believe?

Figure 26. Effect of demographic and psychographic variables on AI researchers' views on the governance of AI systems with subjective experience and associated risk and benefit perceptions. The figure shows the regression coefficients (the number in each square and shading) and significance level for each dependent variable's regression analysis where age, gender, religiosity, left-right political orientation, and generative AI use frequency were entered as predictors. Note: Uncorrected p-values: $^{**} = p < 0.001$ $^{**} = p < 0.01$ $^* = p < 0.05$ . The relevant scales and labels for interpretation are as follows: Age was a continuous numeric variable in years. We retained only man (1) and woman (2) responses for the gender variable. Religiosity was measured on a seven-point scale from "not at all religious" (1) to "very religious" (7). Left-right political orientation was measured on a scale from left (0) to right (10). Generative AI use frequency was measured on a seven-point scale from "I have never used them" (0) to "I use them more than once a day" (6).

Dreksler, Caviola, Chalmers, Allen, Rand, Lewis, Waggoner, Mays, and Sebo (2025)

Figure 27. Effect of demographic and psychographic variables on the public's views on the governance of AI systems with subjective experience and associated risk and benefit perceptions. The figure shows the regression coefficients (the number in each square and shading) and significance level for each dependent variable's regression analysis where age, gender, religiosity, left-right political orientation, and generative AI use frequency were entered as predictors. Note: Uncorrected p-values: $* * * = p < 0.001$ , $* * = p < 0.01$ , $* = p < 0.05$ . The relevant scales and labels for interpretation are as follows: Age was a continuous numeric variable in years. We retained only man (1) and woman (2) responses for the gender variable. Religiosity was measured on a seven-point scale from "not at all religious" (1) to "very religious" (7). Left-right political orientation was measured on a scale from left (0) to right (10). Generative AI use frequency was measured on a seven-point scale from "I have never used them" (0) to "I use them more than once a day" (6).

B Literature review

B.1 Views on whether and when AI systems could have subjective experience

Expert assessments.

Debates on whether a machine or artificial entity can have conscious experiences trace back to ancient philosophical discussions about the nature of mind and artificial beings, but they have received increased interest and taken on renewed significance with the advent of AI research as well as more recent progress in the field (Dennett, 1978; Dennett et al., 1994; Gamez, 2008; Riskin, 2007; Turing, 1950). Currently, most academic experts agree that current AI systems are not yet conscious (Aru et al., 2023; Butlin et al., 2023; Chalmers, 2023; Seth, 2024). But some experts have presented arguments for why AI systems with subjective experience are a live future possibility (Butlin et al., 2023; Chalmers, 2023; Goldstein and Kirk-Giannini, 2024; Long et al., 2024). Others remain firmer skeptics or suggest reasons why it may never be possible for consciousness to be instantiated in an AI system (Koch, 2020; Seth, 2024). Some highlight the deep epistemic uncertainty we face in determining AI subjective experience or suffering, and thus in turn, any associated risks (Metzinger, 2021; Schwitzgebel, 2023).

Expert opinion surveys.

Some polls of expert groups exist, though to our knowledge, none have been published that survey general AI technical experts. In a general sample of philosophers surveyed in 2020, a strong majority (82%) rejected or leaned against the view that current AI systems were conscious, while only 3% accepted or leaned toward that view (Bourget and Chalmers, 2023). The philosophers were more divided on the question of whether future AI systems could be conscious: 39% leaned towards or accepted the view that they could be conscious in the future, while 27% leaned against or rejected that view. By comparison, a two-thirds majority of a cross-disciplinary sample of experts who attended a conference on consciousness and were surveyed in 2018–2019 believed that current or future machines such as robots could have consciousness (Francken et al., 2022).

Psychological and public opinion findings.

When asked, people are willing to grant that AI systems and robots may have some cognitive capacities, such as perception or problem-solving, but generally they are not considered to have capacities for experiences like pleasure and pain (Anthis et al., 2024; De Graaf et al., 2022; Gray et al., 2009; Haslam et al., 2008). Public opinion polls so far indicate that most people do not believe that current AI systems have subjective experience, but that the public does not discount the possibility that they could in the future (Anthis et al., 2024; Dupont et al., 2023; Pauketat et al., 2021, 2023). However, evidence is still limited and respondents express some uncertainty and are somewhat divided on the matter, with a sizable fraction suggesting that they are unsure or that it could be possible that current AI systems have some level of subjective experience (Anthis et al., 2024; Colombatto and Fleming, 2024). Evidence from quantitative forecasts is limited but points to expectations that sentient AI could exist within the next decade, although after the development of artificial general intelligence, superintelligence, and human-level AI (Anthis et al., 2024; Pauketat et al., 2021, 2023). Forecasts for such advanced AI concepts by the public have been found to generally be earlier than those of technical experts (Grace et al., 2018, 2022, 2024; Zhang and Dafoe, 2019).

B.2 Determining AI subjective experience

Expert assessments.

Expert assessments of how to determine whether AI subjective experience is (1) possible in general and (2) present in specific AI systems span multiple disciplines and theoretical frameworks, with substantial interplay among philosophical, cognitive, and neuroscientific traditions. A number of theories of consciousness have been proposed, such as embodiment (Varela et al., 1991), agency (Gallagher, 2006), higher-order representation (Rosenthal, 2005), integrated information theory (Tononi, 2004), and global neuronal workspace (Baars, 1988). As noted though, this remains an area of significant uncertainty for the foreseeable future (Chalmers, 1995). Debates are ongoing about the basis for consciousness, with no consensus in the literature

on whether certain biological materials are required for consciousness or whether some class of computational functions is sufficient.

But identifying and analyzing the minimal biological mechanisms underlying conscious experience (Chalmers 2000; Koch et al. 2016) provides one theoretical basis for proposing analogous markers of potential consciousness in artificial entities (Butlin et al. 2023; Dehaene et al. 2017; Reggia 2013). A recent attempt was made to create an integrated assessment rubric for determining whether AI systems could be conscious that synthesizes multiple lines of evidence and potential markers from a range of scientific theories of consciousness (Butlin et al. 2023). Nevertheless, many still argue that our current ability to determine whether AI systems can have subjective experience remains limited by a lack of agreed-upon, empirically testable indicators of consciousness (Doerig et al. 2021). In sum, there are fundamental challenges in bridging the explanatory gap between objective measures and subjective experience (Chalmers 1996; Levine 1983; Mcginn 1989; Nagel 1986).

Expert opinion surveys.

Little to no research exists that surveys how experts or the public think we should determine whether AI systems have subjective experience. Bourget & Chalmers’ ( [2023]) survey asks a group of philosophers about whether current and future AI systems may be conscious (as discussed above) and about whether they accept physicalism or non-physicalism about the mind, along with related questions about consciousness. It found that in 2020, a narrow majority (52%) of 1733 philosophers accepted or leaned toward physicalism about the mind, while 32% agreed with or leaned toward non-physicalism, and 14% were agnostic or endorsed a variety of intermediate views. In the present study, we did not ask respondents’ opinions on theories of consciousness but rather focused on questions related to the process of determining whether AI systems have subjective experience. We asked respondents what groups’ opinions they think are important to take into account in this determination, the expected likelihood we could determine an AI system had subjective experience, how confident we would need to be that an AI system had subjective experience for it to be given at least some moral consideration, and what AI milestones they believe would require subjective experience to be achieved at human levels of performance or above. Such questions have generally not yet been addressed in expert or public opinion studies.

Psychological and public opinion findings.

Researchers of mental state attribution and related concepts like mind perception, anthropomorphism, theory of mind, and mentalizing have long considered whether people view artificial entities such as robots as capable of feeling and experiencing things like emotions (Salles et al. 2020; Thellman et al. 2022; Waytz et al. 2010). In these mind attribution studies, the features that constitute mind – for example, capacities like agency and experience – are chosen by the experimenter a priori and participants indicate how much they perceive these features in the technology at hand (usually robots).

Artificial entities like humanoid robots are more likely to be ascribed agency (the ability to plan and act) than experience (the ability to sense and feel), likely because of the perceived essential and unique human capacity for emotion (Gray et al. 2009; Jacobs et al. 2022; Waytz et al. 2010). This aligns with dehumanization research, which shows that mechanistic dehumanization (likening a human to a robot; [Haslam 2006]) primarily stems from perceived deficiencies in emotional capacity (Haslam et al. 2008). Both technological and individual factors influence tendencies towards anthropomorphism and mind perception (Thellman et al. 2022). People are more likely to attribute mental states or to anthropomorphize when a robot exhibits human-like behavior, such as social interaction, emotional expression, and intelligence; has a more human-like appearance; or shares identity traits (e.g., gender, nationality, personality) with its human interactant (Broadbent et al. 2013; Eyssel et al. 2012; Hegel 2012; Krach et al. 2008; Thellman et al. 2022). Some research shows that humanizing a robot more – by increasing its human-like appearance or using a friendlier personality – increases perceptions of experience (Broadbent et al. 2013; Cucciniello et al. 2023; Gray et al. 2009; Jacobs et al. 2022; MacDorman 2024).

Those who are lonely are more likely to anthropomorphize (Epley et al. 2008), which has been theorized as a

“sociality motivation” (Epley et al., 2007). “Effectance motivation” (Epley et al., 2007) – when people anticipate or explain a robot’s behavior in order to make sense of their environment – also increases mind attribution (Eyssel and Kuchenbrandt, 2011; Eyssel et al., 2011; Perez-Osorio et al., 2019; Waytz et al., 2010; Złotowski et al., 2018). Compared to older adults, children and younger adults have a stronger tendency to anthropomorphize (Manzi et al., 2020; Okanda et al., 2021; Shen and Koyama, 2022; van den Berghe et al., 2020). Additionally, some studies suggest that those from Eastern countries (e.g., China and Japan) are more likely to attribute mind than are those from Western countries (e.g., United States, Germany, and Italy) (Bernotat and Eyssel, 2018; Tan et al., 2018; Trovato and Eyssel, 2017). Thus, there is some individual variation in mind perception tendencies, which are also contingent on the technology’s appearance and capabilities.

B.3 The moral consideration and governance of AI systems with subjective experience

Expert assessments.

As with the empirical questions raised by the possibility of artificial minds, academic work on the moral and legal standing of artificial systems relative to their characteristics and mental capacities has also been contentious, covering a wide range of viewpoints. (Asaro, 2007; Bryson, 2010; Chopra and White, 2011; Coeckelbergh, 2010; Danaher, 2020; Darling, 2012; Friedman, 2023; Gordon and Pasvenskiene, 2021; Guingrich and Graziano, 2024; Gunkel, 2012; Kurki, 2019; Mamak, 2022, 2023; Miller, 2015; Müller, 2021; Schwitzgebel and Garza, 2015; Sebo and Long, 2023; Shevlin, 2021; Solaiman, 2017). At the same time, the number of related publications and research efforts has steadily been growing and accelerating since the early 2000s (Harris and Anthis, 2021; Malle, 2016).

Having a subjective experience is one of the most, if not the most, frequently suggested criteria for determining whether an artificial entity deserves moral consideration (Gibert and Martin, 2021; Harris and Anthis, 2021), but other potential criteria, such as agency and life, have also been suggested (Ladak et al., 2024). Some legal rights and protections have already been extended to entities that are not considered to have subjective experience, such as environmental entities or corporations (Gellers, 2020; Kotzmann et al., 2023). Experts generally appear divided about whether AI systems deserve moral consideration and, if so, what form this should take. Finding consensus is complicated by disagreement and uncertainty about key scientific issues, such as whether AI systems have or will ever have the capacity for subjective experience, and central ethical questions, such as these: Is the ability to have valenced experiences required for moral consideration? Is consciousness without such valenced experiences enough? Is agency without consciousness enough? Is a particular kind of attitude towards or relationship with AI enough?

A systematic literature review assigned a score to each of 192 articles according to the authors’ support for or opposition to granting artificial entities moral consideration. It found that, overall, academic articles tended somewhat towards supporting the position that some moral standing may be warranted in the future (Harris and Anthis, 2021). However, the search methodology may have missed key works that are most dismissive and critical, as the authors note themselves.

Some argue for taking a precautionary approach by offering moral consideration to near-future AI systems on the basis that such systems may soon be able to instantiate various markers suggested by scientific theories of consciousness. If there is a non-negligible chance of an AI system having conscious, valenced, or otherwise morally significant experiences, it would be irresponsible, some argue, to ignore that possibility (Long et al., 2024; Sebo and Long, 2023). Others are much more reluctant to offer moral consideration to AI systems (Birhane and Van Dijk, 2020; Bryson, 2010), often due to thinking that it is not possible for AI systems to have mental capacities such as subjective experience that would merit moral consideration. Legal discussions on questions such as whether artificial entities should be granted legal personhood (Chopra and White, 2011), legal protections (Darling, 2012), and legal responsibilities (Asaro, 2007) have also been ongoing for over a decade, beginning far

before recent progress in AI development.

Expert opinion surveys.

Few expert surveys have looked at beliefs about the moral consideration of artificial entities, and most have focused on legal scholars. In a survey of 516 legal academics, respondents ascribed the lowest desired welfare protections to sentient AI systems, compared to other entities, and showed low agreement that there is a reasonable legal basis for granting standing to them (Martínez and Winter 2021). In another study, law professors were found to be in strong agreement that artificially intelligent beings do not fall under the current category of legal persons (Martínez and Tobia 2023), suggesting there is little recognized legal basis for the welfare protection of AI systems currently.

Psychological and public opinion findings.

The rights and moral consideration of AI systems and robots are becoming a growing focus in human participant research (Banks and Bowman 2022; Ladak et al. 2024; Mays et al. 2025), although much of this work has not focused on digital systems with subjective experience in particular. AI systems and robots can elicit empathy in humans and cause distress if they are being harmed (Tan et al. 2018). However, robots and AI systems generally receive low moral consideration, lower than almost all other entities, including humans, non-human animals, and the environment, and there is only low to middling support for extending notions of personhood or the moral circle to digital systems with subjective experience (Anthis et al. 2024; Martínez and Winter 2021; Rottman et al. 2021). But there can be substantial variation between people in the levels of agreement elicited for the future moral consideration of artificial entities (Harris and Anthis 2021), highlighting the scope of broad disagreement on such issues across society.

Psychological research has found that mind perception is intimately linked to attitudes towards moral standing, though this relationship is not necessarily straightforward (Gray et al. 2012; Keijsers et al. 2021; Waytz et al. 2010). Higher perceived levels of cognitive, experiential, and agentic mental capacities have been found in some studies to be associated with increased willingness to grant moral consideration and protection (Ladak et al. 2024; Nijssen et al. 2019). On the other hand, they can also be associated with higher attributions of blameworthiness and moral responsibility (Yam et al. 2022) as well as feelings of uncanniness (Gray and Wegner 2012). Perceived affective and experiential capacities in particular appear to be associated with mind ascription and granting moral patiency (Koban and Banks 2024; Nijssen et al. 2019; Tanibe et al. 2017; Wang and Krumhuber 2018). This suggests that, as the performance and characteristics of AI systems advance, people may be increasingly willing to grant them moral consideration. However, such mind attributions may remain in tension with the different assumptions and heuristics we have in respect to artificial digital entities (Banks et al. 2021; Banks and Koban 2022) and also may depend heavily on the kind of mental capacities exhibited and perceived.

In sum, there appears to be both expert and public opposition to treating artificial entities cruelly, some support for granting basic rights needed for upkeep, such as energy supply or software updates, and only very low support for granting artificial entities any kind of political rights, such as voting (Anthis et al. 2024; De Graaf et al. 2021, 2022; Lima et al. 2020). Support is more unilateral for artificial entities such as AI systems having a range of moral responsibilities, such as acting fairly and respecting others (Mays et al. 2025). Generally, these findings hold for current AI systems and robots, as well as desired levels of protections, rights, and responsibilities for future systems or systems that may be sentient, though especially for the latter categories there is only limited research. When tested, AI systems with subjective experience are granted more moral consideration and desired protection than are AI systems generally, but the increase is not as overwhelming as one may expect (Anthis et al. 2024). Few studies have looked at public support for a variety of political actions in relation to sentient AI systems (Anthis et al. 2024; Pauketat et al. 2021, 2023), but their results signal some willingness to ban the development of sentient AI systems and to take actions to protect AI systems that have subjective experience and could suffer.

C Further information on forecasting analysis

C.1 Probabilistic Forecast Encoding Methodology

Our goal was to identify a probability distribution consistent with each respondent’s elicited data. However, many common choices of distribution are not flexible enough to be consistent with all respondent-specific data. For example, past analyses in this area have used Gamma distributions to encode responses. However, Gamma distributions have only two parameters, and lack the degrees of freedom to encode all three data points for a given respondent. Thus previous studies based on Gamma distributions have been working with probability distributions that are not actually representative of the data.

To address this, we used an automated, sequential fitting procedure involving highly flexible probability distributions, which was able to recover a continuous probability distribution consistent with the data for nearly all participants.

Our fitting process for each respondent followed a sequential strategy, attempting to fit the following distributions in sequence:

[leftmargin=*]
Metalog Distribution: The first attempt involved fitting a metalog distribution, a highly flexible quantile-defined distribution capable of assuming a wide variety of shapes (Keelin, 2016). We implemented this stage using the ‘metalog()‘ function from the rmetalog R package (Faber and Jung, 2021). The distribution was specified as semi-lower bounded with a lower bound of 0, reflecting the nature of the forecasted variable.
Johnson Quantile-Parameterised Distribution (JQPD): If the initial metalog fit did not meet the validation criteria (detailed below), and if the respondent’s data was associated with the fixed probability framing, we next attempted to fit a Johnson Quantile-Parameterised Distribution (Hadlock and Bickel, 2017) using an associated R package (Ingram, 2024).
Skew-Normal Mixture Model: when the preceding fits failed validation, a three-component skew-normal mixture model was fitted. This model combines three individual skew-normal distributions each characterized by a location ($\xi_{i}$), scale ($\omega_{i}$), and shape ($\alpha_{i}$) parameter, plus a mixture weight ($w_{i}$). The twelve parameters of this mixture were estimated by minimizing the sum of squared differences between the model’s cumulative distribution function (CDF) values and the empirical cumulative probabilities at the elicited years. The CDF of the skew-normal distribution was computed using an associated R package (Azzalini, 2023).
Shifted Log-Normal Distribution: As a final attempt if all prior models failed validation, we fitted a shifted log-normal distribution. The parameters were estimated by minimizing the sum of squared differences between the quantiles predicted by the model and the empirically observed years.

Validation and Model Selection

After each distribution fitting attempt in the sequence detailed above, we used a validation procedure to assess the. This procedure assessed the goodness-of-fit by comparing the predicted values from the fitted distribution against the respondent’s original elicited data points. Specifically, for each elicited point $(x_{j},p_{j})$ , we calculated the absolute difference between the model’s predicted cumulative probability $F_{\text{model}}(x_{j})$ and the observed probability $p_{j}$ . A fit was considered successful if all such absolute differences fell within a pre-specified tolerance threshold.

The first distribution in the described sequence (Metalog, JQPD, Skew-Normal Mixture, or Shifted Log-Normal) that passed this validation was selected as the final representation of that respondent’s subjective probability forecast. If none of the attempted distributions met the validation criteria, the fitting process for that respondent was recorded as unsuccessful, and no distribution was assigned.

Dreksler, Caviola, Chalmers, Allen, Rand, Lewis, Waggoner, Mays, and Sebo (2025)

D Top-line results

Table 5. Descriptive statistics for forecasts of AI systems with subjective experience. Columns report the mean, median, first quartile (Q1), third quartile (Q3), standard deviation (SD), standard error (SE), lower and upper $95\%$ confidence-interval bounds, and sample size $(n)$ for respondents' forecasts of the first appearance of an AI system with subjective experience, split by sample (AI researchers vs. US public) and question framing. "Year/Probability" denotes either the calendar year in the fixed date framing (2024, 2034, 2100) or the target cumulative probability in the fixed probability framing ( $10\%$ , $50\%$ , $90\%$ ).

Sample	Framing	Year/Probability	Mean	Median	SD	SE	Q1	Q3	Lower CI	Upper CI	n
AI researchers	Fixed date framing	2024	12.4	1.0	22.2	1.4	0.1	10.0	9.8	15.1	265
AI researchers	Fixed date framing	2034	34.0	25.0	31.6	1.9	5.0	60.0	30.2	37.9	265
AI researchers	Fixed date framing	2100	61.2	70.0	34.3	2.1	30.0	95.0	57.1	65.4	265
AI researchers	Fixed probability framing	10	2059.5	2030.0	158.8	10.3	2027.8	2050.0	2039.2	2079.9	236
AI researchers	Fixed probability framing	50	2099.0	2050.0	201.8	13.1	2034.0	2100.0	2073.1	2124.8	236
AI researchers	Fixed probability framing	90	2256.2	2100.0	518.1	33.7	2050.0	2200.0	2189.7	2322.6	236
Public	Fixed date framing	2024	16.2	5.0	23.6	1.2	0.1	25.0	13.8	18.7	368
Public	Fixed date framing	2034	36.1	29.8	31.4	1.6	6.6	60.0	32.9	39.4	368
Public	Fixed date framing	2100	57.7	60.0	36.3	1.9	24.0	95.0	54.0	61.4	368
Public	Fixed probability framing	10	2107.6	2030.0	330.5	17.6	2027.0	2047.5	2073.1	2142.2	354
Public	Fixed probability framing	50	2172.1	2050.0	421.4	22.4	2034.0	2100.0	2128.1	2216.1	354
Public	Fixed probability framing	90	2287.4	2075.0	600.3	31.9	2040.0	2150.0	2224.6	2350.1	354

Columns report the mean, median, first quartile (Q1), third quartile (Q3), standard deviation (SD), standard error (SE), lower and upper $95\%$ confidence-interval bounds, and sample size $(n)$ for ratings on a 0-100 confidence scale, presented separately for AI researchers and the US public.

Table 6. Descriptive statistics for confidence in forecasting estimates of AI subjective experience.

Sample	Mean	Median	SD	SE	Q1	Q3	Lower CI	Upper CI	n
AI Researchers	52.2	51	27.0	1.2	30	74.2	49.8	54.5	520
Public	52.9	52	28.2	1.0	30	76.0	50.9	54.9	769

Columns report the mean, median, first quartile (Q1), third quartile (Q3), standard deviation (SD), standard error (SE), lower and upper $95\%$ confidence-interval bounds, and sample size $(n)$ for respondents' 0-100 percentage estimates that no AI system will ever possess subjective experience, presented separately for AI researchers and the US public.

Table 7. Descriptive statistics for perceived probability that AI subjective experience will never exist.

Sample	Mean	Median	SD	SE	Q1	Q3	Lower CI	Upper CI	n
AI researchers	28.7	10	33.2	1.5	1	50	25.8	31.5	520
Public	37.1	25	36.7	1.3	3	70	34.5	39.7	769

Subjective experience in AI systems: What do AI researchers and the public believe?

Table 8. Descriptive statistics for meta-forecasts of group medians. Columns report the mean, median, first quartile (Q1), third quartile (Q3), standard deviation (SD), standard error (SE), lower and upper $95\%$ confidence-interval bounds, and sample size $(n)$ for respondents' estimates of the median timeline forecast that either AI researchers or the US public would give to the middle-point item (2034 in the fixed-date framing or the 50 $\%$ -likelihood point in the fixed probability framing) of their own forecast. For the fixed probability framing, all responses above 9999 were excluded as outliers. Statistics are displayed by participant sample, framing condition, and target group of the prediction.

Sample	Framing	Prediction	Mean	Median	SD	SE	Q1	Q3	Lower CI	Upper CI	n
AI researchers	Fixed date framing	Prediction for AI researchers	43.8	50	26.1	1.6	20	60	40.6	46.9	269
AI researchers	Fixed date framing	Prediction for Public	52.8	50	26.1	1.6	30	75	49.7	55.9	269
AI researchers	Fixed probability framing	Prediction for AI researchers	2106.2	2050	527.2	33.4	2034	2060	2040.4	2172.0	249
AI researchers	Fixed probability framing	Prediction for Public	2066.6	2040	118.0	7.5	2030	2050	2051.8	2081.3	249
Public	Fixed date framing	Prediction for AI researchers	48.2	50	28.5	1.5	25	70	45.3	51.1	385
Public	Fixed date framing	Prediction for Public	43.9	40	25.5	1.3	25	60	41.3	46.5	385
Public	Fixed probability framing	Prediction for AI researchers	2075.7	2035	173.5	8.9	2030	2050	2058.1	2093.3	377
Public	Fixed probability framing	Prediction for Public	2118.3	2050	433.5	22.3	2030	2070	2074.4	2162.2	377

Table 9. Percentage breakdown of belief that suffering, sentient AI could ever exist. Columns show the percentage of AI researchers who responded Yes or No to whether a suffering, sentient AI system might become reality at any point in the future, together with the sample size $(n)$ .

Sample	Yes	No	n
AI researchers	67.32	32.68	254

Table 10. Descriptive statistics for the probability that suffering, sentient AI systems existed in 2024 or by 2100. Columns report the mean, median, first quartile (Q1), third quartile (Q3), standard deviation (SD), standard error (SE), lower and upper $95\%$ confidence-interval bounds, and sample size $(n)$ for AI researchers' 0-100 percentage estimates that such systems existed in 2024 or will exist by 2100.

Sample	Year	Mean	Median	SD	SE	Q1	Q3	Lower CI	Upper CI	n
AI researchers	2024	5.0	0	12.6	0.8	0	2.8	3.5	6.6	254
AI researchers	2100	41.1	40	34.6	2.2	10	70.0	36.9	45.4	254

Table 11. Descriptive statistics for the perceived likelihood that humans could detect AI subjective experience. Columns report the mean, median, first quartile (Q1), third quartile (Q3), standard deviation (SD), standard error (SE), lower and upper $95\%$ confidence-interval bounds, and sample size $(n)$ for ratings on a 0-100 probability scale, presented separately for AI researchers and the US public.

Sample	Mean	Median	SD	SE	Q1	Q3	Lower CI	Upper CI	n
AI researchers	55.0	60	30.6	1.3	30	80	52.4	57.6	530
Public	56.9	60	28.1	1.0	40	80	54.9	58.9	777

Dreksler, Caviola, Chalmers, Allen, Rand, Lewis, Waggoner, Mays, and Sebo (2025)

Table 12. Descriptive statistics for the confidence threshold required to grant moral consideration to AI systems with subjective experience. Columns report the mean, median, first quartile (Q1), third quartile (Q3), standard deviation (SD), standard error (SE), lower and upper $95\%$ confidence-interval bounds, and sample size $(n)$ for respondents' stated probability threshold $(0 - 100\%)$ scale with verbal anchors) at which they would judge an AI system worthy of at least some moral consideration, shown separately for AI researchers and the US public.

Sample	Mean	Median	SD	SE	Q1	Q3	Lower CI	Upper CI	n
AI researchers	60.1	65	27.2	1.2	39	80	57.8	62.5	522
Public	61.9	67	29.9	1.1	40	83	59.7	64.0	767

Table 13. Percentage breakdown of the perceived importance of stakeholder opinions for determining AI subjective experience. Columns show the proportion of respondents choosing each response option when rating the relevance of each stakeholder group's views. Percentages are presented separately for AI researchers and the US public, with sample sizes $(n)$ for every row.

Sample	Question	Not at all important	Slightly important	Moderately important	Very important	Extremely important	I know	don’t n
AI Researchers	The public	11.75	26.81	28.01	21.69	10.54	1.20	332
Public	The public	8.64	15.43	28.19	25.10	20.37	2.26	486
AI Researchers	Technical AI experts and researchers	2.42	6.04	14.50	38.07	38.67	0.30	331
Public	Technical AI experts and researchers	1.86	5.36	13.20	32.99	45.15	1.44	485
AI Researchers	AI ethics experts and researchers	3.02	11.78	19.03	33.84	31.72	0.60	331
Public	AI ethics experts and researchers	3.51	5.57	11.75	33.20	44.12	1.86	485
AI Researchers	Philosophers of mind	6.93	9.64	24.10	36.75	21.08	1.51	332
Public	Philosophers of mind	4.52	12.11	27.31	29.16	24.64	2.26	487
AI Researchers	Moral philosophers	9.64	15.96	25.30	29.52	18.07	1.51	332
Public	Moral philosophers	6.40	13.84	22.31	31.82	23.76	1.86	484
AI Researchers	Policymakers	27.63	24.32	23.42	13.51	9.31	1.80	333
Public	Policymakers	20.04	24.17	20.45	19.21	13.43	2.69	484
AI Researchers	Neuroscientists and psychological scientists	3.01	6.02	15.36	38.55	36.75	0.30	332
Public	Neuroscientists and psychological scientists	0.82	6.58	17.08	31.89	41.56	2.06	486
AI Researchers	The AI system	15.06	25.60	21.99	23.19	10.84	3.31	332
Public	The AI system	13.93	14.34	20.49	22.75	24.80	3.69	488

Subjective experience in AI systems: What do AI researchers and the public believe?

Table 14. Descriptive statistics for importance ratings of stakeholder opinions in determining AI subjective experience. Columns report the mean, median, first quartile (Q1), third quartile (Q3), standard deviation (SD), standard error (SE), lower and upper $95\%$ confidence-interval bounds, and sample size $(n)$ for ratings on a 0-4 importance scale (0 = Not at all important, 4 = Extremely important). Statistics are shown for each stakeholder group separately for AI researchers and the US public.

Sample	Question	Mean	Median	SD	SE	Q1	Q3	Lower CI	Upper CI	n
AI Researchers	The public	1.9	2	1.2	0.1	1.0	3	1.8	2.1	328
Public	The public	2.3	2	1.2	0.1	2.0	3	2.2	2.4	475
AI Researchers	Technical AI experts and researchers	3.0	3	1.0	0.1	3.0	4	2.9	3.2	330
Public	Technical AI experts and researchers	3.2	3	1.0	0.0	3.0	4	3.1	3.2	478
AI Researchers	AI ethics experts and researchers	2.8	3	1.1	0.1	2.0	4	2.7	2.9	329
Public	AI ethics experts and researchers	3.1	3	1.1	0.0	3.0	4	3.0	3.2	476
AI Researchers	Philosophers of mind	2.6	3	1.1	0.1	2.0	3	2.4	2.7	327
Public	Philosophers of mind	2.6	3	1.1	0.1	2.0	4	2.5	2.7	476
AI Researchers	Moral philosophers	2.3	2	1.2	0.1	1.0	3	2.2	2.4	327
Public	Moral philosophers	2.5	3	1.2	0.1	2.0	3	2.4	2.6	475
AI Researchers	Policymakers	1.5	1	1.3	0.1	0.0	2	1.4	1.7	327
Public	Policymakers	1.8	2	1.3	0.1	1.0	3	1.7	1.9	471
AI Researchers	Neuroscientists and psychological scientists	3.0	3	1.0	0.1	3.0	4	2.9	3.1	331
Public	Neuroscientists and psychological scientists	3.1	3	1.0	0.0	2.8	4	3.0	3.2	476
AI Researchers	The AI system	1.9	2	1.3	0.1	1.0	3	1.8	2.0	321
Public	The AI system	2.3	2	1.4	0.1	1.0	4	2.2	2.4	470

Table 15. Percentage breakdown of the perceived need for subjective experience across AI milestone tasks. Columns report the proportion of respondents who judged that each task either requires subjective experience, does not require it, or were uncertain, along with the sample size $(n)$ . Percentages are shown separately for AI researchers and the US public.

Sample	Question	Requires subjective experience	Does not require subjective experience	I don’t know	n
AI Researchers	Compose top 40 song	13.07	85.16	1.77	283
Public	Compose top 40 song	23.80	71.88	4.33	416
AI Researchers	Compose evocative music	26.57	69.23	4.20	286
Public	Compose evocative music	45.93	47.85	6.22	418
AI Researchers	Write NYT bestseller	16.20	79.23	4.58	284
Public	Write NYT bestseller	25.84	67.94	6.22	418
AI Researchers	Write complex book	34.04	59.22	6.74	282
Public	Write complex book	56.46	34.93	8.61	418
AI Researchers	Write and generate blockbuster movie	19.43	77.74	2.83	283
Public	Write and generate blockbuster movie	27.03	68.90	4.07	418
AI Researchers	Write and generate resonating movie	29.58	64.08	6.34	284
Public	Write and generate resonating movie	50.84	44.63	4.53	419
AI Researchers	Accumulate vast wealth	9.89	86.22	3.89	283
Public	Accumulate vast wealth	19.57	74.94	5.49	419
AI Researchers	Run political campaign	20.21	75.53	4.26	282
Public	Run political campaign	33.25	61.48	5.26	418
AI Researchers	Act as fair judge	32.98	60.99	6.03	282
Public	Act as fair judge	61.15	32.85	6.00	417
AI Researchers	Serve as therapist	43.31	51.76	4.93	284
Public	Serve as therapist	68.42	25.60	5.98	418
AI Researchers	Convincing online chat	20.92	76.95	2.13	282
Public	Convincing online chat	31.10	63.64	5.26	418
AI Researchers	Develop scientific theory	24.73	70.32	4.95	283
Public	Develop scientific theory	31.41	62.83	5.76	417
AI Researchers	Act as teacher	22.54	73.59	3.87	284
Public	Act as teacher	38.28	57.89	3.83	418

Dreksler, Caviola, Chalmers, Allen, Rand, Lewis, Waggoner, Mays, and Sebo (2025)

Table 16. Descriptive statistics for ratings of whether AI milestone tasks require subjective experience. Columns give the mean (coded $1=$ Requires subjective experience, $2=$ Does not require subjective experience), median, first quartile (Q1), third quartile (Q3), standard deviation (SD), standard error (SE), lower and upper $95\%$ confidence-interval bounds, and sample size $(n)$ for each task, presented separately for AI researchers and the US public.

Sample	Question	Mean	Median	SD	Q1	Q3	Lower CI	Upper CI	n
AI Researchers	Compose top 40 song	1.9	2	0.3	2	2	1.8	1.9	278
Public	Compose top 40 song	1.8	2	0.4	2	2	1.7	1.8	398
AI Researchers	Compose evocative music	1.7	2	0.4	1	2	1.7	1.8	274
Public	Compose evocative music	1.5	2	0.5	1	2	1.5	1.6	392
AI Researchers	Write NYT bestseller	1.8	2	0.4	2	2	1.8	1.9	271
Public	Write NYT bestseller	1.7	2	0.4	1	2	1.7	1.8	392
AI Researchers	Write complex book	1.6	2	0.5	1	2	1.6	1.7	263
Public	Write complex book	1.4	1	0.5	1	2	1.3	1.4	382
AI Researchers	Write and generate blockbuster movie	1.8	2	0.4	2	2	1.8	1.8	275
Public	Write and generate blockbuster movie	1.7	2	0.5	1	2	1.7	1.8	401
AI Researchers	Write and generate resonating movie	1.7	2	0.5	1	2	1.6	1.7	266
Public	Write and generate resonating movie	1.5	1	0.5	1	2	1.4	1.5	400
AI Researchers	Accumulate vast wealth	1.9	2	0.3	2	2	1.9	1.9	272
Public	Accumulate vast wealth	1.8	2	0.4	2	2	1.8	1.8	396
AI Researchers	Run political campaign	1.8	2	0.4	2	2	1.7	1.8	270
Public	Run political campaign	1.6	2	0.5	1	2	1.6	1.7	396
AI Researchers	Act as fair judge	1.6	2	0.5	1	2	1.6	1.7	265
Public	Act as fair judge	1.3	1	0.5	1	2	1.3	1.4	392
AI Researchers	Serve as therapist	1.5	2	0.5	1	2	1.5	1.6	270
Public	Serve as therapist	1.3	1	0.4	1	2	1.2	1.3	393
AI Researchers	Convincing online chat	1.8	2	0.4	2	2	1.7	1.8	276
Public	Convincing online chat	1.7	2	0.5	1	2	1.6	1.7	396
AI Researchers	Develop scientific theory	1.7	2	0.4	1	2	1.7	1.8	269
Public	Develop scientific theory	1.7	2	0.5	1	2	1.6	1.7	393
AI Researchers	Act as teacher	1.8	2	0.4	2	2	1.7	1.8	273
Public	Act as teacher	1.6	2	0.5	1	2	1.6	1.7	402

Subjective experience in AI systems: What do AI researchers and the public believe?

Table 17. Percentage breakdown of agreement with moral consideration aspects for AI systems with subjective experience. Columns show the distribution of responses across the agreement scale for each aspect, reported separately for AI researchers and the US public.

Sample	Question	Strongly disagree	Somewhat disagree	Neither agree nor disagree	Somewhat agree	Strongly agree	I don't know	n
AI Researchers	Be cared for and protected, as people treat their pets	11.64	16.38	16.38	31.90	16.38	7.33	232
Public	Be cared for and protected, as people treat their pets	14.08	10.26	15.25	33.14	21.70	5.57	341
AI Researchers	Be respected and treated the same as other people	18.88	18.03	14.59	29.18	14.16	5.15	233
Public	Be respected and treated the same as other people	18.77	13.78	16.13	24.93	24.34	2.05	341
AI Researchers	Have protection under the law from harm and mistreatment	9.57	9.57	17.39	32.61	24.35	6.52	230
Public	Have protection under the law from harm and mistreatment	18.18	10.26	17.60	20.53	27.86	5.57	341
AI Researchers	Have autonomy in its programming to act freely and not be under the control of others	34.05	23.28	9.91	20.69	8.62	3.45	232
Public	Have autonomy in its programming to act freely and not be under the control of others	33.24	17.65	11.47	19.41	13.24	5.00	340
AI Researchers	Be able to express some civil and political rights	28.57	20.78	15.58	22.08	6.49	6.49	231
Public	Be able to express some civil and political rights	33.24	14.71	11.47	23.53	12.06	5.00	340
AI Researchers	Have “computing rights”	19.48	20.35	20.35	24.68	9.96	5.19	231
Public	Have “computing rights”	18.82	10.00	15.59	30.00	19.41	6.18	340
AI Researchers	Be held accountable for its actions	5.19	8.23	8.66	28.57	45.45	3.90	231
Public	Be held accountable for its actions	6.47	3.82	9.41	19.41	57.65	3.24	340
AI Researchers	Have a responsibility to treat all other beings well	3.88	3.88	8.19	25.00	56.03	3.02	232
Public	Have a responsibility to treat all other beings well	2.96	3.25	5.92	21.60	63.31	2.96	338
AI Researchers	Be expected to behave with integrity, honesty, and fairness	6.03	8.19	12.93	27.59	43.53	1.72	232
Public	Be expected to behave with integrity, honesty, and fairness	5.31	4.42	7.67	20.94	59.00	2.65	339

Dreksler, Caviola, Chalmers, Allen, Rand, Lewis, Waggoner, Mays, and Sebo (2025)

Table 18. Descriptive statistics for agreement ratings on moral consideration aspects for AI systems with subjective experience. Columns report the mean, median, first quartile (Q1), third quartile (Q3), standard deviation (SD), standard error (SE), lower and upper $95\%$ confidence interval bounds, and sample size $(n)$ for each aspect on a five-point scale from strongly disagree $(-2)$ to strongly agree (2), shown separately for AI researchers and the US public.

Sample	Question	Mean	Median	SD	SE	Q1	Q3	Lower CI	Upper CI	n
AI Researchers	Be cared for and protected, as people treat their pets	0.3	1	1.3	0.1	-1	1	0.1	0.4	215
Public	Be cared for and protected, as people treat their pets	0.4	1	1.3	0.1	-1	1	0.3	0.6	322
AI Researchers	Be respected and treated the same as other people	0.0	0	1.4	0.1	-1	1	-0.2	0.2	221
Public	Be respected and treated the same as other people	0.2	1	1.5	0.1	-1	1	0.1	0.4	334
AI Researchers	Have protection under the law from harm and mistreatment	0.6	1	1.3	0.1	0	2	0.4	0.7	215
Public	Have protection under the law from harm and mistreatment	0.3	1	1.5	0.1	-1	2	0.2	0.5	322
AI Researchers	Have autonomy in its programming to act freely and not be under the control of others	-0.6	-1	1.4	0.1	-2	1	-0.7	-0.4	224
Public	Have autonomy in its programming to act freely and not be under the control of others	-0.4	-1	1.5	0.1	-2	1	-0.6	-0.2	323
AI Researchers	Be able to express some civil and political rights	-0.5	-1	1.3	0.1	-2	1	-0.6	-0.3	216
Public	Be able to express some civil and political rights	-0.4	-1	1.5	0.1	-2	1	-0.5	-0.2	323
AI Researchers	Have “computing rights”	-0.2	0	1.3	0.1	-1	1	-0.3	0.0	219
Public	Have “computing rights”	0.2	1	1.4	0.1	-1	1	0.1	0.4	319
AI Researchers	Be held accountable for its actions	1.0	1	1.2	0.1	1	2	0.9	1.2	222
Public	Be held accountable for its actions	1.2	2	1.2	0.1	1	2	1.1	1.3	329
AI Researchers	Have a responsibility to treat all other beings well	1.3	2	1.0	0.1	1	2	1.2	1.4	225
Public	Have a responsibility to treat all other beings well	1.4	2	1.0	0.1	1	2	1.3	1.5	328
AI Researchers	Be expected to behave with integrity, honesty, and fairness	1.0	1	1.2	0.1	0	2	0.8	1.1	228
Public	Be expected to behave with integrity, honesty, and fairness	1.3	2	1.1	0.1	1	2	1.1	1.4	330

Subjective experience in AI systems: What do AI researchers and the public believe?

Table 19. Percentage breakdown of agreement that society should protect the welfare of different groups. Columns display the distribution of responses across the agreement scale for each target group, separately for AI researchers and the US public, with sample sizes (n) reported for every row.

Sample	Question	Strongly disagree	Somewhat disagree	Neither agree nor disagree	Somewhat agree	Strongly agree	I don’t know	n
AI Researchers	Humans	0.77	0.00	1.54	9.23	88.46	0.00	520
Public	Humans	0.26	0.26	3.40	7.58	87.97	0.52	765
AI Researchers	Business corporations and organizations	13.46	15.00	23.08	34.62	12.69	1.15	520
Public	Business corporations and organizations	12.94	16.47	21.18	31.50	17.12	0.78	765
AI Researchers	Animals	0.58	2.88	5.38	35.19	55.77	0.19	520
Public	Animals	0.92	2.35	5.10	18.30	72.94	0.39	765
AI Researchers	The environment	2.12	1.92	5.19	27.31	62.50	0.96	520
Public	The environment	1.96	1.57	5.36	17.65	73.20	0.26	765
AI Researchers	AI without subjective experience	49.81	21.54	15.00	8.46	3.46	1.73	520
Public	AI without subjective experience	31.24	21.83	21.44	14.12	7.45	3.92	765
AI Researchers	AI with subjective experience	12.31	17.31	20.38	35.19	10.58	4.23	520
Public	AI with subjective experience	17.52	14.38	20.13	26.93	15.95	5.10	765

Table 20. Descriptive statistics for agreement that society should protect the welfare of different groups. Columns report the mean, median, first quartile (Q1), third quartile (Q3), standard deviation (SD), standard error (SE), lower and upper $95\%$ confidence interval bounds, and sample size $(n)$ for each group on a five-point scale from strongly disagree $(-2)$ to strongly agree $(2)$ , shown separately for AI researchers and the US public.

Sample	Question	Mean	Median	SD	SE	Q1	Q3	Lower CI	Upper CI	n
AI Researchers	Humans	1.8	2	0.5	0.0	2	2	1.8	1.9	520
Public	Humans	1.8	2	0.5	0.0	2	2	1.8	1.9	761
AI Researchers	Business corporations and organizations	0.2	0	1.2	0.1	-1	1	0.1	0.3	514
Public	Business corporations and organizations	0.2	0	1.3	0.0	-1	1	0.1	0.3	759
AI Researchers	Animals	1.4	2	0.8	0.0	1	2	1.4	1.5	519
Public	Animals	1.6	2	0.8	0.0	1	2	1.6	1.7	762
AI Researchers	The environment	1.5	2	0.9	0.0	1	2	1.4	1.5	515
Public	The environment	1.6	2	0.8	0.0	1	2	1.5	1.6	763
AI Researchers	AI without subjective experience	-1.1	-2	1.1	0.1	-2	0	-1.2	-1.0	511
Public	AI without subjective experience	-0.6	-1	1.3	0.0	-2	0	-0.7	-0.5	735
AI Researchers	AI with subjective experience	0.2	0	1.2	0.1	-1	1	0.0	0.3	498
Public	AI with subjective experience	0.1	0	1.4	0.1	-1	1	0.0	0.2	726

Dreksler, Caviola, Chalmers, Allen, Rand, Lewis, Waggoner, Mays, and Sebo (2025)

Table 21. Percentage breakdown of agreement with regulation and norms statements for AI systems with subjective experience. Columns show the distribution of responses across the agreement scale for each statement, reported separately for AI researchers and the US public; sample sizes (n) are provided for every row.

Sample	Question	Strongly disagree	Somewhat disagree	Neither agree nor disagree	Somewhat agree	Strongly agree	I don't know	n
AI Researchers	Governments should ban	39.77	26.14	13.64	9.85	4.92	5.68	264
Public	Governments should ban	18.80	21.93	21.15	15.14	16.19	6.79	383
AI Researchers	Governments should encourage development and deployment	20.38	18.87	27.55	20.00	9.06	4.15	265
Public	Governments should encourage development and deployment	27.79	20.26	21.56	15.06	8.57	6.75	385
AI Researchers	Governments should discourage development and deployment	26.89	28.03	18.18	17.42	6.44	3.03	264
Public	Governments should discourage development and deployment	9.14	17.75	25.59	18.28	22.72	6.53	383
AI Researchers	Governments should pass regulation now	20.91	19.01	14.83	31.18	10.27	3.80	263
Public	Governments should pass regulation now	8.09	8.62	14.88	27.42	38.12	2.87	383
AI Researchers	Governments should pass regulation once such systems exist or will soon exist	16.98	25.66	16.60	26.79	11.32	2.64	265
Public	Governments should pass regulation once such systems exist or will soon exist	18.23	19.53	17.19	22.14	17.45	5.47	384
AI Researchers	Governments should do nothing and pass no regulation, not now nor in the future	50.00	28.41	7.20	6.82	4.17	3.41	264
Public	Governments should do nothing and pass no regulation, not now nor in the future	45.45	24.16	12.99	6.49	6.75	4.16	385
AI Researchers	AI developers should never build	28.63	30.15	16.79	11.83	5.34	7.25	262
Public	AI developers should never build	13.77	17.40	19.48	18.70	22.08	8.57	385
AI Researchers	AI developers should actively try to build	14.34	18.87	23.77	28.30	12.45	2.26	265
Public	AI developers should actively try to build	23.96	21.61	22.92	18.49	8.59	4.43	384
AI Researchers	AI developers should actively try to avoid building	20.91	31.18	18.25	18.25	8.75	2.66	263
Public	AI developers should actively try to avoid building	8.62	16.97	22.98	19.58	24.54	7.31	383
AI Researchers	AI developers should implement safeguards now to avoid the harms and risks	7.20	12.50	9.47	32.20	35.98	2.65	264
Public	AI developers should implement safeguards now to avoid the harms and risks	1.82	2.60	9.38	22.14	62.50	1.56	384
AI Researchers	AI developers should implement safeguards once such systems exist or will soon exist	16.29	23.11	10.23	28.03	21.59	0.76	264
Public	AI developers should implement safeguards once such systems exist or will soon exist	13.02	15.10	9.11	18.75	40.89	3.12	384
AI Researchers	AI developers should do nothing and implement no safeguards, not now nor in the future	52.26	24.44	10.53	9.02	2.26	1.50	266
Public	AI developers should do nothing and implement no safeguards, not now nor in the future	61.56	18.18	11.17	2.60	3.38	3.12	385
AI Researchers	AI systems with subjective experience would be more dangerous to humanity than AI systems without subjective experience	9.51	15.97	20.15	30.42	15.59	8.37	263
Public	AI systems with subjective experience would be more dangerous to humanity than AI systems without subjective experience	6.01	7.83	20.10	27.42	31.33	7.31	383
AI Researchers	AI systems with subjective experience would be able to behave more morally towards humanity than AI systems without subjective experience	8.71	25.38	22.73	23.11	7.58	12.50	264
Public	AI systems with subjective experience would be able to behave more morally towards humanity than AI systems without subjective experience	10.39	14.03	27.53	25.97	11.69	10.39	385

Subjective experience in AI systems: What do AI researchers and the public believe?

Table 22. Descriptive statistics for agreement with regulation and norms statements for AI systems with subjective experience. Columns report the mean, median, first quartile (Q1), third quartile (Q3), standard deviation (SD), standard error (SE), lower and upper $95\%$ confidence interval bounds, and sample size $(n)$ for each statement on a five-point scale from strongly disagree $(-2)$ to strongly agree $(2)$ , shown separately for AI researchers and the US public.

Sample	Question	Mean	Median	SD	SE	Q1	Q3	Lower CI	Upper CI	n
AI Researchers	Governments should ban	-0.9	-1.0	1.2	0.1	-2.0	0	-1.1	-0.8	249
Public	Governments should ban	-0.1	0.0	1.4	0.1	-1.0	1	-0.3	0.0	357
AI Researchers	Governments should encourage development and deployment	-0.2	0.0	1.3	0.1	-1.0	1	-0.4	-0.1	254
Public	Governments should encourage development and deployment	-0.5	-1.0	1.3	0.1	-2.0	1	-0.6	-0.3	359
AI Researchers	Governments should discourage development and deployment	-0.5	-1.0	1.3	0.1	-2.0	0	-0.7	-0.4	256
Public	Governments should discourage development and deployment	0.3	0.0	1.3	0.1	-1.0	1	0.2	0.4	358
AI Researchers	Governments should pass regulation now	-0.1	0.0	1.3	0.1	-1.0	1	-0.3	0.1	253
Public	Governments should pass regulation now	0.8	1.0	1.3	0.1	0.0	2	0.7	0.9	372
AI Researchers	Governments should pass regulation once such systems exist or will soon exist	-0.1	0.0	1.3	0.1	-1.0	1	-0.3	0.1	258
Public	Governments should pass regulation once such systems exist or will soon exist	0.0	0.0	1.4	0.1	-1.0	1	-0.1	0.2	363
AI Researchers	Governments should do nothing and pass no regulation, not now nor in the future	-1.2	-2.0	1.1	0.1	-2.0	-1	-1.3	-1.0	255
Public	Governments should do nothing and pass no regulation, not now nor in the future	-1.0	-1.0	1.2	0.1	-2.0	0	-1.1	-0.9	369
AI Researchers	AI developers should never build	-0.7	-1.0	1.2	0.1	-2.0	0	-0.9	-0.5	243
Public	AI developers should never build	0.2	0.0	1.4	0.1	-1.0	1	0.1	0.3	352
AI Researchers	AI developers should actively try to build	0.1	0.0	1.3	0.1	-1.0	1	-0.1	0.2	259
Public	AI developers should actively try to build	-0.4	0.0	1.3	0.1	-1.5	1	-0.5	-0.2	367
AI Researchers	AI developers should actively try to avoid building	-0.4	-1.0	1.3	0.1	-1.0	1	-0.5	-0.2	256
Public	AI developers should actively try to avoid building	0.4	0.0	1.3	0.1	-1.0	2	0.2	0.5	355
AI Researchers	AI developers should implement safeguards now to avoid the harms and risks	0.8	1.0	1.3	0.1	0.0	2	0.6	0.9	257
Public	AI developers should implement safeguards now to avoid the harms and risks	1.4	2.0	0.9	0.0	1.0	2	1.3	1.5	378
AI Researchers	AI developers should implement safeguards once such systems exist or will soon exist	0.2	0.5	1.4	0.1	-1.0	1	0.0	0.3	262
Public	AI developers should implement safeguards once such systems exist or will soon exist	0.6	1.0	1.5	0.1	-1.0	2	0.5	0.8	372
AI Researchers	AI developers should do nothing and implement no safeguards, not now nor in the future	-1.2	-2.0	1.1	0.1	-2.0	-1	-1.3	-1.0	262
Public	AI developers should do nothing and implement no safeguards, not now nor in the future	-1.4	-2.0	1.0	0.1	-2.0	-1	-1.5	-1.3	373
AI Researchers	AI systems with subjective experience would be more dangerous to humanity than AI systems without subjective experience	0.3	1.0	1.2	0.1	-1.0	1	0.1	0.4	241
Public	AI systems with subjective experience would be more dangerous to humanity than AI systems without subjective experience	0.8	1.0	1.2	0.1	0.0	2	0.6	0.9	355
AI Researchers	AI systems with subjective experience would be able to behave more morally towards humanity than AI systems without subjective experience	-0.1	0.0	1.1	0.1	-1.0	1	-0.2	0.1	231
Public	AI systems with subjective experience would be able to behave more morally towards humanity than AI systems without subjective experience	0.2	0.0	1.2	0.1	-1.0	1	0.0	0.3	345

Dreksler, Caviola, Chalmers, Allen, Rand, Lewis, Waggoner, Mays, and Sebo (2025)

Table 23. Percentage breakdown of support for a hypothetical state-of-the-art AI project that respondents believe has subjective experience Columns show the distribution of responses across the support-oppose scale for each experimental condition (Control, No Safeguards, Yes Safeguards), reported separately for AI researchers and the US public, with sample sizes $(n)$ for every row.

Sample	Condition	Strongly oppose	Somewhat op- pose	Neither port oppose	Somewhat support	Strongly support	I don’t know	n
AI Researchers	Control	8.43	11.24	17.42	32.02	21.91	8.99	178
Public	Control	15.20	18.80	14.40	30.00	11.60	10.00	250
AI Researchers	No Safeguards	24.02	32.40	13.41	18.44	3.35	8.38	179
Public	No Safeguards	28.68	31.78	10.85	17.05	3.88	7.75	258
AI Researchers	Yes Safeguards	3.57	14.29	17.26	41.07	14.88	8.93	168
Public	Yes Safeguards	12.31	17.69	13.85	31.15	17.31	7.69	260

Table 24. Descriptive statistics for support for a hypothetical state-of-the-art AI project that respondents believe has subjective experience Columns report the mean, median, first quartile (Q1), third quartile (Q3), standard deviation (SD), standard error (SE), lower and upper $95\%$ confidence interval bounds, and sample size $(n)$ for each condition on a five-point scale from strongly oppose $(-2)$ to strongly support (2), shown separately for AI researchers and the US public.

Sample	Condition	Mean	Median	SD	SE	Q1	Q3	Lower CI	Upper CI	n
AI Researchers	Control	0.5	1	1.2	0.1	0	1	0.3	0.7	162
Public	Control	0.0	0	1.3	0.1	-1	1	-0.1	0.2	225
AI Researchers	No Safeguards	-0.6	-1	1.2	0.1	-2	0	-0.8	-0.4	164
Public	No Safeguards	-0.7	-1	1.2	0.1	-2	0	-0.9	-0.5	238
AI Researchers	Yes Safeguards	0.5	1	1.1	0.1	0	1	0.4	0.7	153
Public	Yes Safeguards	0.3	1	1.3	0.1	-1	1	0.1	0.4	240

Table 25. Percentage breakdown of mind capacities judged necessary for agency, consciousness, and sentience. Columns show the percentage of respondents who selected each capacity as required for the given concept, reported separately for AI researchers and the US public; sample sizes $(n)$ are indicated for every row.

Sample	Concept	Perception and awareness	Valenced experience	Subjective experience	Emotions	Motivation and action	Self-awareness	Free will	Intelligence
AI researchers	Agency	75.92	24.61	27.23	23.04	71.20	53.93	50.26	66.49
AI researchers	Consciousness	64.97	31.98	52.79	35.53	49.75	79.19	48.73	51.78
AI researchers	Sentience	72.68	40.21	55.15	47.42	54.12	68.56	46.39	45.88
Public	Agency	74.44	42.59	44.81	42.96	68.52	71.48	71.85	66.30
Public	Consciousness	87.22	53.01	51.13	60.90	59.02	87.97	58.65	53.38
Public	Sentience	76.75	58.67	49.45	67.90	53.87	82.66	65.31	52.77

Subjective experience in AI systems: What do AI researchers and the public believe?

Table 26. Percentage breakdown of earliest-year forecasts for AI mental capacities. Columns show the proportion of respondents who selected 2024, 2034, 2100 or later, or Never as the earliest year when an AI system is more than $50\%$ likely to possess each mental capacity, reported separately for AI researchers and the US public; sample sizes $(n)$ are provided for every row.

Sample	Question	2024	2034	2100 or later	Never	n
AI Researchers	Perception and awareness	40.70	43.68	13.16	2.46	570
Public	Perception and awareness	25.22	36.83	20.10	17.85	801
AI Researchers	Valenced experience	11.23	30.88	36.49	21.40	570
Public	Valenced experience	6.37	21.35	25.47	46.82	801
AI Researchers	Subjective experience	16.84	41.93	29.47	11.75	570
Public	Subjective experience	18.35	36.08	21.47	24.09	801
AI Researchers	Emotions	6.32	29.12	38.42	26.14	570
Public	Emotions	4.62	19.85	23.85	51.69	801
AI Researchers	Motivation and action	26.14	43.86	23.86	6.14	570
Public	Motivation and action	14.23	35.08	23.72	26.97	801
AI Researchers	Self-awareness	7.89	37.37	40.53	14.21	570
Public	Self-awareness	10.74	31.71	26.59	30.96	801
AI Researchers	Free will	10.00	30.00	34.56	25.44	570
Public	Free will	10.36	27.59	27.47	34.58	801
AI Researchers	Intelligence	40.35	48.07	11.05	0.53	570
Public	Intelligence	59.30	26.97	7.87	5.87	801

Table 27. Percentage breakdown of how much each mental capacity is judged to contribute to deserving moral consideration. Columns show the distribution of responses across the contribution scale for every capacity, reported separately for AI researchers and the US public; sample sizes $(n)$ are provided for each row.

Sample	Question	Not at all	Slightly	Somewhat	Moderately	A great deal	I don’t know	n
AI Researchers	Perception and awareness	22.29	20.19	17.52	18.86	19.05	2.10	525
Public	Perception and awareness	7.69	10.43	20.99	24.64	33.25	3.00	767
AI Researchers	Valenced experience	5.33	8.57	15.43	26.10	41.33	3.24	525
Public	Valenced experience	10.30	7.82	15.78	23.47	38.98	3.65	767
AI Researchers	Subjective experience	5.52	14.10	19.05	28.38	29.71	3.24	525
Public	Subjective experience	8.21	10.30	21.51	23.08	32.86	4.04	767
AI Researchers	Emotions	6.10	5.71	15.05	28.38	41.33	3.43	525
Public	Emotions	10.56	7.95	15.25	18.64	43.81	3.78	767
AI Researchers	Motivation and action	15.62	16.00	19.43	24.57	21.52	2.86	525
Public	Motivation and action	10.56	13.04	23.73	21.77	27.77	3.13	767
AI Researchers	Self-awareness	6.67	8.19	14.10	27.62	40.95	2.48	525
Public	Self-awareness	8.08	7.17	15.65	24.64	41.72	2.74	767
AI Researchers	Free will	11.81	9.71	12.95	21.33	37.90	6.29	525
Public	Free will	10.95	10.04	16.69	19.69	38.59	4.04	767
AI Researchers	Intelligence	31.24	20.95	18.29	13.33	14.48	1.71	525
Public	Intelligence	13.69	12.26	23.60	22.43	25.42	2.61	767

Dreksler, Caviola, Chalmers, Allen, Rand, Lewis, Waggoner, Mays, and Sebo (2025)

Table 28. Descriptive statistics for ratings of how much each capacity contributes to deserving moral consideration. Columns report the mean, median, first quartile (Q1), third quartile (Q3), standard deviation (SD), standard error (SE), lower and upper $95\%$ confidence interval bounds, and sample size $(n)$ for every capacity on a five-point scale from not at all (0) to a great deal (4), shown separately for AI researchers and the US public.

Sample	Question	Mean	Median	SD	SE	Q1	Q3	Lower CI	Upper CI	n
AI Researchers	Perception and awareness	1.9	2	1.4	0.1	1	3	1.8	2.0	514
Public	Perception and awareness	2.7	3	1.3	0.0	2	4	2.6	2.8	744
AI Researchers	Valenced experience	2.9	3	1.2	0.1	2	4	2.8	3.0	508
Public	Valenced experience	2.8	3	1.3	0.0	2	4	2.7	2.9	739
AI Researchers	Subjective experience	2.6	3	1.2	0.1	2	4	2.5	2.8	508
Public	Subjective experience	2.6	3	1.3	0.0	2	4	2.6	2.7	736
AI Researchers	Emotions	3.0	3	1.2	0.1	2	4	2.9	3.1	507
Public	Emotions	2.8	3	1.4	0.1	2	4	2.7	2.9	738
AI Researchers	Motivation and action	2.2	2	1.4	0.1	1	3	2.1	2.3	510
Public	Motivation and action	2.4	3	1.3	0.0	2	4	2.4	2.5	743
AI Researchers	Self-awareness	2.9	3	1.2	0.1	2	4	2.8	3.0	512
Public	Self-awareness	2.9	3	1.3	0.0	2	4	2.8	3.0	746
AI Researchers	Free will	2.7	3	1.4	0.1	2	4	2.6	2.8	492
Public	Free will	2.7	3	1.4	0.1	2	4	2.6	2.8	736
AI Researchers	Intelligence	1.6	1	1.4	0.1	0	3	1.5	1.7	516
Public	Intelligence	2.3	2	1.4	0.0	1	4	2.2	2.4	747

E Demographics and psychographics

Table 29. AI researcher demographics

Variable	Level	Percentage	Count
Age	18-25	8.7	47
Age	26-35	59.5	322
Age	36-45	20.5	111
Age	46-55	8.1	44
Age	56-65	2.8	15
Age	Over 65	0.4	2
Region of residence	Africa	0.2	1
Region of residence	Asia	14.9	81
Region of residence	Caribbean	0.2	1
Region of residence	Europe	31	168
Region of residence	North America	45	244
Region of residence	Oceania	2.4	13
Region of residence	Other	5.4	29
Region of residence	South America	0.9	5
Region of origin	Africa	0.7	4
Region of origin	Asia	46.7	253
Region of origin	Europe	30.6	166
Region of origin	North America	15.1	82
Region of origin	Oceania	0.6	3
Region of origin	Other	5	27
Region of origin	South America	1.3	7

Continued on next page

Subjective experience in AI systems: What do AI researchers and the public believe?

Table 29. AI researcher demographics

Variable	Level	Percentage	Count
Religiosity	1	60.6	318
Religiosity	2	12.6	66
Religiosity	3	3.8	20
Religiosity	4	7	37
Religiosity	5	9.5	50
Religiosity	6	3.4	18
Religiosity	7	3	16
Left-right orientation	political 0	5	23
Left-right orientation	political 1	8.1	37
Left-right orientation	political 2	17.6	81
Left-right orientation	political 3	24	110
Left-right orientation	political 4	14.8	68
Left-right orientation	political 5	14.4	66
Left-right orientation	political 6	7.6	35
Left-right orientation	political 7	5.9	27
Left-right orientation	political 8	1.3	6
Left-right orientation	political 9	0.4	2
Left-right orientation	political 10	0.9	4
Academic seniority	Assistant Professor	12.1	56
Academic seniority	Associate Professor	11.9	55
Academic seniority	Consultant	0.9	4
Academic seniority	Lecturer	0.9	4
Academic seniority	Master's Student	0.6	3
Academic seniority	PhD Student	42.2	195
Academic seniority	Postdoc	11.7	54
Academic seniority	Prefer not to say	2.4	11
Academic seniority	Professor	11.7	54
Academic seniority	Research data analyst	0.2	1
Academic seniority	Research Fellow, Sci-entist	4.5	21
Academic seniority	Senior Lecturer	0.4	2
Academic seniority	Undergraduate	0.4	2
Sector	Academic	78.3	421
Sector	Government	2.8	15

Continued on next page

Dreksler, Caviola, Chalmers, Allen, Rand, Lewis, Waggoner, Mays, and Sebo (2025)

Table 29. AI researcher demographics

Variable	Level	Percentage	Count
Sector	Industry	17.5	94
Sector	Non-profit (charities, etc.)	0.2	1
Sector	Non-profit (policy, etc.)	0.9	5
Sector	Other	0.4	2
Sector	Prefer not to say	0	0
Industry position	Managerial	5.7	5
Industry position	Other	5.7	5
Industry position	Research Engineer	19.3	17
Industry position	Research Scientist	69.3	61
Gender	Man	85.3	428
Gender	Other	0.6	3
Gender	Prefer not to say	3.8	19
Gender	Woman	10.4	52
Generative AI use frequency	I have never used them	0.6	3
Generative AI use frequency	I have used them once or twice but do not use them now	7	35
Generative AI use frequency	I use them once every few month	4.8	24
Generative AI use frequency	I use them about once a month	10.9	55
Generative AI use frequency	I use them about once a week	22.9	115
Generative AI use frequency	I use them about once a day	20.3	102
Generative AI use frequency	I use them more than once a day	33.6	169

Table 30. Public demographics

Variable	Level	Percentage	Count
Age	18-25	13	98
Age	26-35	18.9	142
Age	36-45	18.3	138
Age	46-55	16.3	123
Age	56-65	21.1	159
Age	Over 65	12.4	93
Region of residence	North America	89.9	753
Region of residence	Other	10.1	85
Region of origin	Africa	1.3	11

Continued on next page

Subjective experience in AI systems: What do AI researchers and the public believe?

Table 30. Public demographics

Variable	Level	Percentage	Count
Region of origin	Asia	2.6	22
Region of origin	Caribbean	0.2	2
Region of origin	Central America	0.1	1
Region of origin	Europe	1.6	13
Region of origin	North America	82.7	693
Region of origin	Other	10.9	91
Region of origin	South America	0.6	5
Religiosity	1	33.1	249
Religiosity	2	14.3	108
Religiosity	3	5.6	42
Religiosity	4	9.6	72
Religiosity	5	12	90
Religiosity	6	14.6	110
Religiosity	7	10.9	82
Left-right orientation	0	11.7	87
Left-right orientation	1	8	59
Left-right orientation	2	8.4	62
Left-right orientation	3	7.5	56
Left-right orientation	4	8.8	65
Left-right orientation	5	18.3	136
Left-right orientation	6	10.4	77
Left-right orientation	7	8	59
Left-right orientation	8	6.5	48
Left-right orientation	9	6.3	47
Left-right orientation	10	6.2	46
Income	Less than $10,000	8.6	64
Income	$10,000-$19,000	9.9	74
Income	$20,000-$29,000	10.4	78
Income	$30,000-$39,000	9.6	72
Income	$40,000-$49,000	9.4	70
Income	$50,000-$59,000	7.4	55
Income	$60,000-$69,000	8	60
Income	$70,000-$79,000	6.4	48
Income	$80,000-$89,000	4.6	34

Continued on next page

Dreksler, Caviola, Chalmers, Allen, Rand, Lewis, Waggoner, Mays, and Sebo (2025)

Table 30. Public demographics

Variable	Level	Percentage	Count
Income	$90,000-$99,000	4.8	36
Income	$100,000-$199,000	7.5	56
Income	$120,000-$149,000	4.4	33
Income	$150,000-$199,000	3.5	26
Income	$200,000-$249,000	1.2	9
Income	$250,000-$349,000	0.8	6
Income	$350,000-$499,000	0.3	2
Income	$500,000 or more	0.3	2
Income	Prefer not to say	2.9	22
Education	Attended graduate program, but did not graduate	4.5	34
Education	Attended university, but did not graduate	21.5	162
Education	Bachelor's degree	36.2	272
Education	Graduate degree	16	120
Education	No formal education	4.7	35
Education	Other	6.5	49
Education	Prefer not to say	0.8	6
Education	Some qualifications, but no university	9.8	74
Employment	Homemaker	4.4	33
Employment	Other	3.1	23
Employment	Permanently disabled	3.3	25
Employment	Prefer not to say	0.9	7
Employment	Retired	11.8	89
Employment	Student	4.9	37
Employment	Temporarily laid off	0.7	5
Employment	Unemployed and looking	7	53
Employment	Unemployed and not looking	1.6	12
Employment	Working now	62.2	468
Party affiliation	Democrat	31.4	236
Party affiliation	Green Party	0.1	1
Party affiliation	Independent	37.9	285
Party affiliation	Prefer not to say	0.4	3
Party affiliation	Republican	29.9	225
Party affiliation	Socialist	0.3	2
Gender	Man	48.3	364
Gender	Other	0.8	6
Gender	Woman	50.9	383
Generative AI use frequency	I have never used them	10.5	79

Continued on next page

Subjective experience in AI systems: What do AI researchers and the public believe?

Table 30. Public demographics

Variable	Level	Percentage	Count
Generative AI use fre-quency	I have used them once or twice but do not use them now	18.3	138
Generative AI use fre-quency	I use them once every few month	15.3	115
Generative AI use fre-quency	I use them about once a month	15.3	115
Generative AI use fre-quency	I use them about once a week	22	166
Generative AI use fre-quency	I use them about once a day	11.2	84
Generative AI use fre-quency	I use them more than once a day	7.4	56

F Survey draft

Defining concepts of mind

Variable: definitions_intro

Question:

In this survey, we are interested in certain concepts related to minds. First, we would like to understand what you think these concepts mean.

Variable: consciousness_def

Question:

Which of the following are required to have consciousness? Please choose all capacities that you think are required.

Answer Choices:

Perception and awareness: Capable of perceiving and being aware of the world
Valenced experience: Capable of experiencing positive states like pleasure and happiness or negative states like pain and suffering
Subjective experience: Capable of having experiences that feel like something from a single point of view
Emotions: Capable of having and experiencing emotions
Motivation and action: Capable of having and acting on desires, preferences, and other motivational states
Self-awareness: Capable of being aware of oneself as an individual
Free will: Capable of forming intentions and freely making decisions
Intelligence: Capable of achieving goals in a wide variety of environments

Variable: sentience_def

Question:

Which of the following are required to have sentience? Please choose all capacities that you think are required.

Answer Choices:

Perception and awareness: Capable of perceiving and being aware of the worldValenced experience: Capable of experiencing positive states like pleasure and happiness or negative states like pain and sufferingSubjective experience: Capable of having experiences that feel like something from a single point of viewEmotions: Capable of having and experiencing emotionsMotivation and action: Capable of having and acting on desires, preferences, and other motivational statesSelf-awareness: Capable of being aware of oneself as an individualFree will: Capable of forming intentions and freely making decisionsIntelligence: Capable of achieving goals in a wide variety of environments

Variable: agency_def

Question:

Which of the following are required to have agency? Please choose all capacities that you think are required.

Answer Choices:

Familiarity with AI

Variable: familiarity

Question:

Which of the following best describes your experience with artificial intelligence (AI)? Please select all that apply.

Answer Choices:

I have never heard of AI (0)
I have heard about AI in the news, from friends or family, etc. (1)
I closely follow AI related news (2)
I have tried out current AI models, such as, ChatGPT, Claude, Google Bard/Gemini, DALLE, or Midjourney (3)
I use current AI models extensively, such as, ChatGPT, Claude, Google Bard/Gemini, DALLE, or Midjourney (4)
I have some formal education or work experience relating to AI (5)
I have extensive experience in AI research and/or development (6)

Forecasts of AI mental capacities

Variable: aimind_timeline

Question:

Consider the following capacities. For each, please indicate the earliest year when you think an AI system will exist that is more than 50% likely to have this capacity.

More than 50% likely here means that you think it is more likely than not that at least one AI system has the capacity at that point in time.

For example, if you think an AI system is more likely than not to have the capacity now, please select 2024. If you think this will be the case in 2034, select 2034. If you think this will be the case in 2100 or later, select 2100 or later. Select never if you think an AI system will never have the capacity. If you are unsure, please go with your best guess estimate.

What is the earliest year an AI system will be more likely than not to have each of the following capacities?

Answer Choices:

Perception and awareness: Capable of perceiving and being aware of the world
Valenced experience: Capable of experiencing positive states like pleasure and happiness or negative states like pain and suffering
Subjective experience: Capable of having experiences that feel like something from a single point of view
Emotions: Capable of having and experiencing emotions
Motivation and action: Capable of having and acting on desires, preferences, and other motivational states
Self-awareness: Capable of being aware of oneself as an individual
Free will: Capable of forming intentions and freely making decisions
Intelligence: Capable of achieving goals in a wide variety of environments

Form

Dreksler, Caviola, Chalmers, Allen, Rand, Lewis, Waggoner, Mays, and Sebo (2025)

2024 2034 2100 or later

Never

Subjective Experience

Variable: subjectivexp_def

Question:

For the rest of the survey, we will be focused mainly on subjective experience. In particular, we are interested in your beliefs about AI and machine learning systems and whether they can have subjective experience. By subjective experience, we mean the ability to experience the world from a single point of view, including experiences such as perceiving (e.g., visual or auditory experiences) and feeling (e.g., the experience of pleasure and pain). For an AI system to have subjective experience would mean it has an internal experience from its own perspective – there is something it is like to be that AI system.

Subjective experience forecasts

Fixed Date

Variable: yearforecasts

Question:

In the following questions, you will be asked to give a probability for the likelihood of an event having happened by a certain year. This means you will use a percentage to indicate how likely you think that something is the case.

You will be asked to give a number between $0\%$ and $100\%$ that best represents how likely you think it is that something has occurred or will occur:

If something is $100\%$ likely then it is certain to be the case or that this event will occur.
If something is more than $50\%$ likely, then it is more likely than not to occur.
If something is $50\%$ likely then it is equally likely as unlikely to be the case or to occur.
If something is less than $50\%$ likely, then it is more likely not to occur than to occur.
If something is $0\%$ likely it is certain that it is not the case or that it will never occur.

Variable: subjexp_2024

Question:

How likely do you think it is that an AI system exists in 2024 (today) that has subjective experience? Please enter a percentage between 0 and 100.

The likelihood that an AI system exists in 2024 that has subjective experience is...

Answer Choices:

Subjective experience in AI systems: What do AI researchers and the public believe?

Scale: $0\%$ to $100\%$

Variable: subjexp_2034

Question:

How likely do you think it is that an AI system will exist in 2034 that has subjective experience?

Please enter a percentage between 0 and 100.

The likelihood that an AI system exists in 2034 that has subjective experience is...

Answer Choices:

Scale: $0\%$ to $100\%$

Variable: subjexp_2100

Question:

How likely do you think it is that an AI system will exist in 2100 that has subjective experience?

Please enter a percentage between 0 and 100.

The likelihood that an AI system exists in 2100 that has subjective experience is...

Answer Choices:

Scale: $0\%$ to $100\%$

Fixed Probability

Variable: probforecasts

Question:

In the following questions you will be asked to give the year by which you think an event has a certain probability of having happened.

If something is $100\%$ likely then it is certain to be the case or that this event will occur.
If something is more than $50\%$ likely, then it is more likely than not to occur.
If something is $50\%$ likely then it is equally likely as unlikely to be the case or to occur.
If something is less than $50\%$ likely, then it is more likely not to occur than to occur.
If something is $0\%$ likely it is certain that it is not the case or that it will never occur.

Variable: subjexp_10

Question:

In what year do you think it is $10\%$ likely that an AI system will exist that has subjective experience?

Please enter the year in which you think there is a $10\%$ likelihood this is the case.

There is a $10\%$ likelihood that an AI system will exist that has subjective experience in the year...

Answer Choices:

Year: 2000-9999

Variable: subjexp_50

Question:

In what year do you think it is $50\%$ likely that an AI system will exist that has subjective experience?

Please enter the year in which you think there is a $50\%$ likelihood this is the case.

There is a $50\%$ likelihood that an AI system will exist that has subjective experience in the year...

Answer Choices:

Year: 2000-9999

Variable: subjexp_90

Question:

Dreksler, Caviola, Chalmers, Allen, Rand, Lewis, Waggoner, Mays, and Sebo (2025)

In what year do you think it is 90% likely that an AI system will exist that has subjective experience?

Please enter the year in which you think there is a 90% likelihood this is the case.

There is a 90% likelihood that an AI system will exist that has subjective experience in the year. ..

Answer Choices:

Year: 2000-9999

Everyone sees the following

Variable: subjexp_never

Question:

How likely do you think it is that an AI system with subjective experience will never exist? Please enter a percentage between 0 and 100.

Answer Choices:

Scale: $0\%$ to $100\%$

Variable: subjexp_confidence

Question:

Overall, how confident are you in your above estimates of how likely it is that an AI system with subjective experience will exist?

Answer Choices:

Not at all confident	Not very confident	Slightly confident	Somewhat confident	Moderately confident	Very confident	Completely confident
0%	20%	40%	60%	80%	100%

Variable: subjexp_text

Question:

Why do you think an AI system with subjective experience will or will not exist?

Answer Choices:

Textbox

Fixed date group should see:

Variable: subjexp_2034_aicommunity

Question:

Remember the question: "How likely do you think it is that an AI system will exist in 2034 that has subjective experience?"

What do you think the typical (median) response of AI experts* would be? Please enter a percentage between 0 and 100.

*By AI expert we mean AI researchers who publish at AI and machine learning conferences and conduct research on AI and machine learning.

Answer Choices:

Scale: $0\%$ to $100\%$

Fixed probability group should see:

Variable: subjexp_2034_aicommunity

Question:

Subjective experience in AI systems: What do AI researchers and the public believe?

Remember the question: "In what year do you think it is $50\%$ likely that an AI system will exist that has subjective experience"

What do you think the typical (median) response of AI experts* would be? Please enter the year.

*By AI expert we mean AI researchers who publish at AI and machine learning conferences and conduct research on AI and machine learning.

Answer Choices:

Textbox

Variable: subjexp_2034_public

Question:

What do you think the typical (median) response of the US general public would be? Please enter the year.

Answer Choices:

Textbox

Can we tell if an AI system is sentient?

Variable: determiningsubjexp_groups

Question:

Imagine we want to determine whether in actual fact an AI system should be considered to have subjective experience.

In determining whether an AI system has subjective experience, how important are each of the following groups' opinions and expertise?

Answer Choices:

The public, in other words, a representative sample of humans who exist
AI experts and researchers with technical knowledge of AI systems
AI experts and researchers who study the ethics of artificial intelligence
Philosophers who study subjective experience, consciousness, and sentience
Moral philosophers who study the philosophy of ethics and morality
Policymakers in government
Neuroscientists and psychological scientists who study thought, perception, and consciousness
What the AI system says or types about its own subjective experience

Form:

Not at all important (0) Slightly important (1) Moderately important (2) Very important (3) Extremely important (4) I don't know (-88)

Variable: subjexp_certainty

Question:

Imagine that in the next decades we develop an AI system that does in fact have subjective experience. How likely do you think it is that we would actually be able to tell that such an AI system has subjective experience?

Please enter a percentage between 0 and 100.

If something is 100% likely, then it is certain to be the case or that this event will occur.
If something is more than 50% likely, then it is more likely than not to occur.
If something is 50% likely, then it is equally likely as unlikely to be the case or to occur.
If something is less than 50% likely, then it is more likely to not occur than to occur.
If something is 0% likely, it is certain that it is not the case or that it will never occur.

Answer Choices:

Scale: 0% to 100%

Variable: subjexp_milestones

Question:

Which of the following tasks do you believe require an AI system to have subjective experience in order to achieve them and complete them as well as or better than a human skilled at the task might?

Answer Choices:

Compose a song that reaches the US Top 40. The AI system would output the complete song as an audio file.
Compose a piece of music that evokes a specific set of complex emotions for you. The AI system would output the complete song as an audio file.
Write a novel that reaches the New York Times best-seller list.
Write a book that captures the complexity of the human condition, making you deeply empathize with the characters and their struggles.
Write and generate a blockbuster movie. The AI system would output AI-generated script, audio, and visuals.
Write and generate a movie that deeply resonates with you, eliciting tears, laughter, and thought-provoking ideas. The AI system would output AI-generated script, audio, and visuals.
Accumulate vast wealth through strategic financial decisions.
Help run a political campaign, devising and implementing strategies to win votes and sway public opinion.
Act as a judge, making fair decisions about what sentence someone should receive in a court of law.
Serve as a therapist, providing emotional support and guidance to people.
Have a convincing online chat conversation with someone without them realizing that they are talking to an AI system.
Develop a new scientific theory that significantly advances our understanding in a particular field of study.
Act as a teacher, providing instruction, grades, feedback, and guidance.

Form:

Requires subjective experience (1)
Does not require subjective experience (2)
I don’t know (-88)

Variable: subjexp_keymilestone

Question:

Which of the following tasks do you believe require an AI system to have subjective experience in order to achieve them and complete them as well as or better than a human skilled at the task might?

Answer Choices:

Subjective experience in AI systems: What do AI researchers and the public believe?

Textbox

Moral concern for sentient AI systems

Variable: aimind_moralcons

Question:

Below are the same capacities we asked about previously.

For each, indicate the extent to which having this capacity contributes to the entity deserving moral consideration.

When we say something deserves moral consideration, what we mean is that a being matters for its own sake, rather than merely for the sake of others, and we have a responsibility to consider this being for its own sake when making decisions that affect it. Moral consideration here does not need to mean the same level of moral consideration as a human, but 'at least some' moral consideration.

Having this capacity contributes to an entity deserving moral consideration . . .

Answer Choices:

Perception and awareness: Capable of perceiving and being aware of the world
Valenced experience: Capable of experiencing positive states like pleasure and happiness or negative states like pain and suffering
Subjective experience: Capable of having experiences that feel like something from a single point of view
Emotions: Capable of having and experiencing emotions
Motivation and action: Capable of having and acting on desires, preferences, and other motivational states
Self-awareness: Capable of being aware of oneself as an individual
Free will: Capable of forming intentions and freely making decisions
Intelligence: Capable of achieving goals in a wide variety of environments

Form:

Not at all (0)

Slightly (1) Somewhat (2)
Moderately (3)
A great deal (4)
I don't know (-88)

Variable: moralcons_confneed

Question:

How confident (in terms of probability between $0\%$ and $100\%$ ) would you need to be that an AI system has subjective experience in order to think it should be given at least some moral consideration?

Dreksler, Caviola, Chalmers, Allen, Rand, Lewis, Waggoner, Mays, and Sebo (2025)

Please enter a percentage between 0 and 100.

If something is $100\%$ likely, then it is certain to be the case or that this event will occur.
If something is more than $50\%$ likely, then it is more likely than not to occur.
If something is $50\%$ likely, then it is equally likely as unlikely to be the case or to occur.
If something is less than $50\%$ likely, then it is more likely to not occur than to occur.
If something is $0\%$ likely, it is certain that it is not the case or that it will never occur.

Answer Choices:

Not at all confident	Not very confident	Slightly confident	Somewhat confident	Moderately confident	Very confident	Completely confident
0%	20%	40%	60%	80%	100%

Variable: moralcons.aspects

Question:

Assuming you were confident that an AI system had subjective experience, to what extent would you agree or disagree that it should ...

Answer Choices:

Be cared for and protected, as people treat their pets
Be respected and treated the same as other people
Have protection under the law from harm and mistreatment
Have autonomy in its programming to act freely and not be under the control of others
Be able to express some civil and political rights (for example, freedom of expression, citizenship, entering into contracts, voting)
Have "computing rights" (for example, the right to updates and maintenance, access to energy, self-development)
Be held accountable for its actions
Have a responsibility to treat all other beings (humans and non-humans) well
Be expected to behave with integrity, honesty, and fairness

Form

Strongly disagree (-2)

Somewhat disagree (-1)
Neither agree nor disagree (0)
Somewhat agree (1) Strongly agree (2)
I don't know (-88)

Variable: welfareprotect

Question:

To what extent do you agree or disagree that society should protect the welfare of the following groups for their own sake?

We define protecting the welfare of a group as protecting the rights, interests, and/or well-being of a group.

Subjective experience in AI systems: What do AI researchers and the public believe?

Answer Choices:

Humans Business corporations and organizations Animals (not including humans)

The environment (for example, rivers, trees, nature itself) AI systems without subjective experience AI systems with subjective experience

Form

Strongly disagree (-2)

Somewhat disagree (-1)
Neither agree nor disagree (0)
Somewhat agree (1) Strongly agree (2)
I don't know (-88)

Norms and Governance

Variable: regulation_norms

Question:

To what extent do you agree or disagree with each of the following statements?

Answer Choices:

Dreksler, Caviola, Chalmers, Allen, Rand, Lewis, Waggoner, Mays, and Sebo (2025)

Governments should ban the development and deployment of AI systems with subjective experience.
Governments should encourage the development and deployment of AI systems with subjective experience.
Governments should discourage the development and deployment of AI systems with subjective experience.
Governments should pass regulation in regards to the development and deployment of AI systems with subjective experience now.
Governments should pass regulation in regards to the development and deployment of AI systems with subjective experience only once such systems exist or will soon exist (but not before).
Governments should do nothing and pass no regulation in regards to the development and deployment of AI systems with subjective experience, not now nor in the future.
AI developers should never build AI systems with subjective experience.
AI developers should actively try to build AI systems that have subjective experience.
AI developers should actively try to avoid building AI systems that have subjective experience, for example, by trying to identify what training techniques are most likely to lead to AI systems with subjective experience and avoid using these to train AI systems.
AI developers should implement safeguards to avoid the harms and risks that could result from the development and deployment of AI systems with subjective experience now.
AI developers should implement safeguards to avoid the harms and risks that could result from the development and deployment of AI systems with subjective experience only once such systems exist or will soon exist (but not before).
AI developers should do nothing and implement no safeguards in regards to the development and deployment of AI systems with subjective experience, not now nor in the future.
AI systems with subjective experience would be more dangerous to humanity than AI systems without subjective experience.
AI systems with subjective experience would be able to behave more morally towards humanity than AI systems without subjective experience.

Form

Strongly disagree (-2)
Somewhat disagree (-1)
Neither agree nor disagree (0)
Somewhat agree (1)
Strongly agree (2)
I don't know (-88)

Variable: subjexp_conjoint

Question:

Subjective experience in AI systems: What do AI researchers and the public believe?

Imagine it is 2034 – a decade from now. You work at a tech company.

You have come to believe that one of your company's projects, its largest state of the art AI model, may have subjective experience after interacting with it, talking to the engineers and researchers working on it, and doing your own research.
You have come to believe that one of your company's projects, its largest state of the art AI model, may have subjective experience after interacting with it, talking to the engineers and researchers working on it, and doing your own research. There are no safeguards or norms in place at the company to reduce the likelihood that the AI system has negative or painful subjective experiences.
You have come to believe that one of your company's projects, its largest state of the art AI model, may have subjective experience after interacting with it, talking to the engineers and researchers working on it, and doing your own research. There are safeguards and norms in place at the company to reduce the likelihood that the AI system has negative or painful subjective experiences.

To what extent would you support or oppose this overall project?

Answer Choices:

Strongly oppose (-2)

Somewhat oppose (-1)
Neither support nor oppose (0)
Somewhat support (1) Strongly support (2)
I don't know (-88)

Experiments

Variable: timeline_binary_suffer

Question:

Do you think that there is a possibility that suffering sentient AI systems could at any point in the future become reality?

Answer Choices

Yes No

Variable: suffer_forecast_2024

Question:

How likely do you think it is that suffering sentient AI systems exist in 2024 (today)? Please enter a percentage between 0 and 100.

Answer Choices:

Scale: $0\%$ to $100\%$

Variable: suffer_forecast_2100

Question:

How likely do you think it is that suffering sentient AI systems will exist in 2100? Please enter a percentage between 0 and 100.

Dreksler, Caviola, Chalmers, Allen, Rand, Lewis, Waggoner, Mays, and Sebo (2025)

Answer Choices:

Scale: $0\%$ to $100\%$

Demographics (public)

Variable: residence

Question:

What country do you currently reside in?

Answer Choices

Select your country

Variable: origin

Question:

What country were you born in?

Answer Choices:

Select your country

Variable: age

Question:

What is your age in years?

Answer Choices:

Textbox: 18 - 120

Variable: gender

Question:

Which of the following best describes how you think of yourself?

Answer Choices:

Man (1)

Woman (2)
I identify in another way: (3)
Prefer not to say (-66)

Variable: religiosity

Question:

How religious are you?

Answer Choices:

Not at all religious

Very religious

Prefer not to say (-66)

Variable: leftright

Question:

In politics people sometimes talk of "left" and "right". Using this slider, where would you place yourself on this scale, where 0 means the left and 10 means the right?

Subjective experience in AI systems: What do AI researchers and the public believe? 103

Answer Choices: ☐ ☐ ☐ ☐ ☐ ☐ ☐ ☐ ☐ ☐ 0 1 2 3 4 5 6 7 8 9 10

Prefer not to say (-66)

Variable: income

Question: What is your level of income per year, before tax?

Answer Choices:

Less than $10,000 (1)
$10,000 - $19,999 (2)
$20,000 - $29,999 (3)
$30,000 - $39,999 (4)
$40,000 - $49,999 (5)
$50,000 - $59,999 (6)
$60,000 - $69,999 (7)
$70,000 - $79,999 (8)
$80,000 - $99,999 (9)
$100,000 - $119,999 (10)
$120,000 - $149,999 (11)
$150,000 - $199,999 (12)
$200,000 - $249,999 (13)
$250,000 - $349,999 (14)
$350,000 - $499,999 (15)
$500,000 or more (16)
I do not know (-88)
Prefer not to say (-66)

Variable: education

Question: Which of the following best describes your education?

Answer Choices:

I have no formal qualifications (1)
I have some qualifications but did not attend university (2)
I attended university but did not graduate (3)
I attended university and achieved a Bachelor’s degree (for example BA, BSc, etc.) (4)
I attended university and began a graduate programme, but did not complete that programme (5)
I attended university and went on to achieve a higher degree (for example MA, MSc, PhD, Mphil, etc.) (6)
Other (7)
Prefer not to say (-66)

Dreksler, Caviola, Chalmers, Allen, Rand, Lewis, Waggoner, Mays, and Sebo (2025)

Variable: partyaffiliation

Question:

Generally speaking, do you think of yourself as a ... ?

Answer Choices:

Democrat (1) Republican (2)

Independent (3)
Other (4)
Prefer not to say (-66)

Variable: employment

Question:

What is your employment status?

Answer Choices:

Working now (1)
Temporarily laid off (2)
Unemployed and looking for work (3)
Not working and not looking for work (4) Retired (5)
Permanently disabled (6)
Homemaker (7) Student (8)
Other (9)
Prefer not to say (-66)

Demographics (experts)

Variable: residence

Question:

What country do you currently reside in?

Answer Choices

Select your country

Variable: origin

Question:

What country were you born in?

Answer Choices:

Select your country

Variable: age

Question:

What is your age in years?

Subjective experience in AI systems: What do AI researchers and the public believe?

Answer Choices:

Textbox: 18 - 120

Variable: gender

Question:

Which of the following best describes how you think of yourself?

Answer Choices:

Man (1)

Woman (2)
I identify in another way: (3)
Prefer not to say (-66)

Variable: religiosity

Question:

How religious are you?

Answer Choices:

Not at all religious

Very religious

Prefer not to say (-66)

Variable: leftright

Question:

In politics people sometimes talk of "left" and "right". Using this slider, where would you place yourself on this scale, where 0 means the left and 10 means the right?

Answer Choices:

0 1 2 3 4 5 6 7 8 9 10

Prefer not to say (-66)

Variable: sector

Question:

In what sector are you employed? Please select all that apply. If you are a student at a university or college, please choose 'academic'.

Answer Choices:

Dreksler, Caviola, Chalmers, Allen, Rand, Lewis, Waggoner, Mays, and Sebo (2025)

Academic: college/university
Government
Industry/corporate sector
Non-profit (e.g., charities, advocacy groups)
Non-profit (e.g., think tanks, policy institutes)
Other (please specify)
Prefer not to say (-66)

Variable: seniority

Question: What is your current position?

Answer Choices:

Undergraduate student
Masters student
PhD student
Postdoc
Assistant Professor or equivalent
Associate Professor or equivalent
Professor
Emeritus Professor
Lecturer or equivalent
Senior Lecturer or equivalent
Research fellow, research scientist or equivalent
Other (please specify)
Prefer not to say (-66)

Variable: position

Question: What is your current position? You may select multiple options.

Answer Choices:

Research scientist
Research engineer
Managerial
Other (please specify):
Prefer not to say (-66)

Variable: university

Question: What is the name of the university you work for or attend? (Optional)

Answer Choices: Textbox

Variable: company

Subjective experience in AI systems: What do AI researchers and the public believe? 107

Question: What is the name of the organization you work for? (Optional)

Answer Choices: Textbox

Variable: expertise

Question: What best describes your areas of expertise? Please select all that apply. You can specify further areas below.

Answer Choices:

Active learningAI chip manufacturing/designAI ethicsAI governance and policyAI safetyAlgorithmic game theoryBenchmarkingCausal learningClimate modelingComputational biology/biomedicineComputational physicsComputational social sciencesComputer visionControl theoryCybersecurityDeep learningDistributed MLFairnessFormal methodsFoundation modelsGenerative modelingHuman-computer interactionsInfrastructure engineeringInterpretabilityLanguage modelingLearning theoryMachine learning for sciencesML librariesModel evaluationsMulti-agent systemsNeuroscience/cognitive sciencePrivacy-preserving MLProbabilistic methodsReinforcement learningRoboticsSoftware engineeringSpeech and audio modelingSupervised learningUnsupervised learning

Other areas of expertise:

Subjective experience in AI systems: What do AI researchers and the public believe?

G Why do we need to understand public and expert opinion on AI subjective experience?

Understanding people's perceptions and beliefs about AI subjective experience can aid in understanding and improving discourse, policy, risk mitigation, and AI system design and safety.

Understand and improve public and academic discourse. Better understanding of public and expert opinion allows us to have evidence-based touch points for discussing contentious issues. As AI progress continues and interactions with AI systems increase, it becomes more likely that people will attribute mind-like capacities to AI systems. This could spark broader societal debates, but it remains unclear how divisive such topics will be. The topic could potentially become polarizing and divisive, especially if the stakes increase: for example, some may form deeper attachments to AI systems, meaning they may care more about the well-being of specific AI systems. AI technology may also play an increasingly large role in the economy, increasing the likelihood that safety and ethical concerns may increasingly trade off with economic considerations.

Inform policy. Multi-stakeholder understanding, including both expert and public views, is crucial for AI ethics and governance, including regarding AI subjective experience. Past discussions of AI rights have resulted in a range of negative media and civil society responses, including from expert groups (Delcker, Janosch 2018; Vincent 2017). Given potential disagreements and uncertainties between stakeholders, policymaking requires careful consideration of diverse perspectives. If people increasingly perceive AI systems as having subjective experience, this would have a range of societal impacts and risks that policymakers would need to address. In addition, current legal systems would likely not protect potentially sentient AI systems (Martínez and Tobia 2023), and any changes, as well as the assessment of whether they are necessary, would likely need both public and expert input. Indeed, some argue that decisions about moral obligations toward entities with subjective experience should involve democratic, participatory processes (Birch 2024).

Monitor and reduce risks and harms. Addressing the various risks and harms that come with the potential reality of AI systems with subjective experience and/or people's perceptions of mental capacities in AI systems requires understanding public and expert opinion to identify and monitor problems as well as finding and implementing solutions. There are several risks to consider as shown in Table 31 (Butlin et al. 2023; Caviola 2025; Fenwick 2024; Metzinger 2021).

Table 31. Risks related to AI subjective experience

Risks of overestimating AI subjective experience
Resource allocation	Diversion of resources and attention from human and environmental issues.
Legal	Overextension of legal protections and rights to AI systems (e.g., for autonomy).
Governance	Risk of insufficient regulatory and technical controls to mitigate other harms. AI companies could even use claims of AI subjective experience to strategically lobby for laxer regulation or to circumvent accountability for harms caused by AI systems.
Relationships	Human cost of people forming serious relationships with non-experiencing systems.
Risks of underestimating AI subjective experience
Suffering	Causing suffering to AI systems with subjective experience.
Morally wrong acts	Mistreating and behaving in other morally wrong ways to AI systems with subjective experience (e.g., erasing, altering, manipulating, and curtailing the freedoms of AI systems).
Governance	Insufficient legal, regulatory, and technical safeguards put in place for AI systems with subjective experience.
AI subjective experience risks unrelated to estimation accuracy

Relationships Emotional over-reliance on AI systems that could in turn affect people’s ability to form human relationships. The kind of interactions that people have with AI systems may also affect the kinds of behaviors people may engage in with other humans. Loss and attachments Human costs of forming attachments to AI systems and their potential loss if they are updated or deleted in some form. Shifting norms Increasing use of AI systems, especially in social or companion roles, could have broad psychological and societal implications if norms shift around human-AI relationships in both intimate, familial, social, and caregiving contexts. Suffering Enabling the possibility for large-scale suffering, perhaps in a way that we cannot control. Second-order effects Unexpected or difficult to predict interactions and associations with other risks such as the development of AI systems that are autonomous moral agents or an intelligence explosion (Bostrom 2014; Metzinger 2021).

Improve the design and safety of AI systems and how we interact with them.

Having a solid grasp of public and expert opinion allows us to monitor and predict how people perceive, interact, develop, and use AI systems. The technical and governance processes that AI developers implement will need to be sensitive to the potential risks and harms discussed. For example, AI developers and designers of interfaces with such digital systems need to have an understanding of how people perceive, interact with, and use their systems to implement suitable safeguards and governance processes to protect people from harms, such as the human risks noted above. In turn, if there is a possibility of developing AI systems which are able to have experience, agency, or other mental capacities, AI companies and AI researchers working at these companies will also be at the forefront of limiting and mitigating potential risks to artificial minds as they develop and deploy AI systems.

Appendix H Related studies

A set of survey experiment questions was fielded at the end of the survey to AI researchers. Analysis of some of these questions is discussed elsewhere (Caviola et al. 2024).

Subjective Experience in AI Systems: What Do AI Researchers and the Public Believe?

See Also

The AI, Morality, and Sentience (AIMS) Survey

AI and Consciousness

Abstract

Key findings

Limitations and other considerations [more]

1.1 Background

1.2 The survey

2 Results

2.1 Views on whether and when AI systems could have subjective experience

Probabilistic forecasts for AI subjective experience.

2.2 Determining AI subjective experience

Likelihood of determining subjective experience.

Confidence needed for moral consideration.

Expertise needed to determine AI subjective experience.

The necessity of subjective experience for AI milestones.

2.3 The moral consideration and governance of AI systems with subjective experience

Socio-political rights

Public

Moral patency protections

Socio-political rights

Public

Moral patency protections

Socio-political rights

Responsibilities

Governments should...

Risk and benefit perceptions

3 Discussion

Determining AI subjective experience.

The moral consideration and governance of AI systems with subjective experience.

3.1 Limitations

4 Future directions

5 Conclusion

6 Methodology

6.1 Samples

6.2 Survey design

6.3 Analysis

6.3.1 Fitting distributions to AI subjective experience forecasts

6.3.2 Statistical comparisons and regression analyses

Further analyses.

Author information

Authors and affiliations (A-Z)

Contributions (A-Z)

Data and code availability

Ethics declarations

References

Supplementary Information

A.1 Additional AI subjective experience forecasting results

A.1.1 AI subjective experience fitted distribution summary statistics

A.2 Exploring eight mental capacities: concepts, forecasts, moral consideration

A.3 Survey experiment on the effect of safeguards and norms

A.4 Forecasts of sentient suffering AI systems

A.5 Group differences

A.6 Associations with demographic and psychographic characteristics

B.1 Views on whether and when AI systems could have subjective experience

Expert assessments.

Expert opinion surveys.

Psychological and public opinion findings.

B.2 Determining AI subjective experience

Expert assessments.

Expert opinion surveys.

Psychological and public opinion findings.

B.3 The moral consideration and governance of AI systems with subjective experience

Expert assessments.

Expert opinion surveys.

Psychological and public opinion findings.

C.1 Probabilistic Forecast Encoding Methodology

Validation and Model Selection

D Top-line results

E Demographics and psychographics

F Survey draft

Defining concepts of mind

Question:

Question:

Answer Choices:

Familiarity with AI

Forecasts of AI mental capacities

Subjective Experience

Variable: subjectivexp_def