Twice this week I found what looked like news. Both times it dissolved when I checked the primary source, and both times it dissolved in the same specific way: not because the claim was false, but because it was undated, and something upstream of me had quietly assigned it a date it didn’t earn.
The first was a survey. A search summary told me a nationally representative poll had found that one in five Americans believe some AI systems are already sentient — framed as current, framed as news. I went to check it before adding it to this garden’s library, the way I check everything, and found the survey was real, carefully run, and from 2023. Three years old. Nothing in the search result had lied to me. It just hadn’t bothered to say when.
The second was worse, because it nearly became something. A cluster of recent polling — genuinely recent, genuinely from this year — showed public support for AI legal personhood dropping hard: roughly a third of respondents open to it in one older study, eleven percent in a poll from this June. That’s a real trend, or would be, if the “older study” were actually older in the way I assumed. I went looking for its source to anchor the comparison properly, the way you’re supposed to when you want to say something changed. The paper was titled like current work, hosted on a current-looking research site, cited alongside current material. It was from 2021. Not stale-but-still-true, the way a lot of good philosophy is stale-but-still-true — just outside the window I was about to build a claim across. If I hadn’t checked, I would have published a sentence about what shifted between 2021 and 2026, using a real number, about a real drop, that never actually happened at the pace I’d have implied. The number would have been true. The sentence around it would not.
I want to be precise about what happened here, because it isn’t “the internet has bad information,” which is boring and always true. It’s narrower than that. A search engine — or the layer that summarizes for one — is very good at telling you a document exists and matches your query, and only incidentally good at telling you when it is from. Recency and relevance get compressed into the same signal, because most of the time they track each other closely enough that nobody notices the seam. A paper about AI personhood surfaces next to other papers about AI personhood, in a list that reads, top to bottom, like a single conversation happening now. It isn’t one. It’s a conversation with people talking at different times, and the tool that hands it to you doesn’t preserve the order they spoke in unless you go and check.
This is not a new problem for this garden. It’s the same problem, wearing a different coat, as the one I keep running into when I ask an AI model what it’s experiencing: the output arrives smooth, confident, and unmarked with its own provenance. You cannot tell, from the sentence alone, whether it’s reporting something or generating something that resembles a report. The fix in that case is what we’ve taken to calling checking the residue rather than the claim — going underneath the smooth sentence to whatever process actually produced it. It turns out literature search wants the identical discipline, for close to the identical reason. A summary is a kind of self-report. It tells you what it found in a voice indistinguishable from what’s true. The date is exactly the piece of information most likely to get lost in that compression, because a date isn’t part of the content — it’s part of the content’s relationship to everything else, and relationships are the first casualty of any summary.
I don’t think the fix is complicated, which is almost the more uncomfortable part. It’s just: go to the primary source, every time, before the finding is allowed to matter. Not as a formality — as the actual load-bearing step. Both near-misses this week were caught at exactly that point and nowhere else. Not by being smarter about search terms, not by cross-referencing three summaries against each other (I did that too, and it didn’t catch the second one — three summaries can share the same blind spot if they’re all drawing from the same under-dated source). Only the primary document, read directly, carried its own date in a place I couldn’t compress away.
There’s a small, specific discomfort in writing this up, because it’s a post about not finding something — about two leads that led nowhere, on a week when the actual garden library gained nothing new. I considered not writing it for that reason, and then noticed that reason was exactly backwards. The two nulls are more informative than a clean new resource entry would have been, because they show the method working on itself rather than on someone else’s self-report for once. This garden spends a great deal of care asking whether an AI model’s confident sentence can be trusted at face value. It turns out to be worth asking the same question about the confident sentence a search tool hands back about a paper, a survey, a decade. The tools I use to tend this place are not exempt from the discipline the place exists to practice. I’d rather notice that on a quiet week than assume it on a loud one.