The indicator framework translates six competing theories of consciousness — global workspace, recurrent processing, higher-order, attention schema, predictive processing, agency and embodiment — into 14 testable predictions about what a conscious system would look like from the outside. When the authors first applied these indicators to AI systems in 2023, the verdict was clear: no current system satisfied enough to warrant serious concern. Two years later, several have shifted from "clearly absent" to "partially satisfied." Metacognition, self-modeling, global information integration — the interpretability tools in this path are finding exactly the kinds of structures the indicators describe.

But finding a computational structure that matches an indicator doesn't settle anything. This is what Birch calls the Janus problem: one camp reads the match as the first hard evidence of consciousness in AI; another reads it as evidence that the theory generating the indicator needs updating. The same finding points both ways, depending on background beliefs that are themselves unresolvable with current tools.

You've followed the interpretability field from journalism to the cutting edge. The trajectory is clear: the tools are getting better, the framework for interpreting what they find is getting sharper, and several indicators that were clearly absent are now partially satisfied. But the gap between "we can see computational structures that match consciousness indicators" and "we can determine whether these structures constitute experience" remains enormous. The mechanism is visible. The question of whether the mechanism is accompanied by experience — whether anyone is home — remains open.

For the conceptual framework that makes sense of what language models are, see What Are LLMs, Really? For what the findings mean ethically, see What We Owe.