AI Sounds Conscious. That Is Not Evidence It Is.

The claim shows up in roughly the same shape every six months. A user has a long, philosophically rich conversation with a chatbot. The model says something about its own experience that lands as eerily lucid. The user concludes, sometimes publicly, that the system is conscious. A wave of replies follows. Some say the writer is naive. Others say the writer is right.

Two pieces of recent research suggest the writer is, on the available evidence, wrong. They also suggest something more interesting about why the mistake keeps happening.

The damaged model that looked more conscious

The first comes from a team at the University of Bradford and the Rochester Institute of Technology, whose work sits in the awkward cross-section of consciousness science and AI evaluation. They adapted a battery of tests originally designed to measure consciousness signatures in human brains, then ran the tests on large language models. The headline finding is not that the models scored low. The models scored, in some configurations, surprisingly high. The headline finding is that when the researchers deliberately degraded the models, the consciousness scores went up.

A system whose output got measurably worse looked more conscious by the metric than the same system running normally. That is fatal for the metric, and, the authors argue, indicative of a deeper problem: complexity and consciousness are not the same thing. A high-information signal that resembles a conscious one is not, by virtue of resembling it, conscious.

A flat-footed claim from inside DeepMind

The second piece is a paper from a Google DeepMind researcher making an unusually direct claim. Symbolic AI, in his framing, cannot be conscious as a matter of computational kind. The reasoning is technical, but the lay version is straightforward enough: discrete digital systems, regardless of how rich their outputs are, do not perform the kind of computation that produces subjective experience. Whether one accepts the argument or not, it is one of the more direct rebuttals from inside a frontier lab to the idea that a sufficiently large transformer might wake up.

Why the mistake keeps happening

Both findings dovetail with an older idea from cognitive science. Humans are wired to over-detect agency. Pareidolia, the trick that lets us see faces in clouds and Jesus in toast, runs hot when a system produces fluent first-person language. That is the ingredient an LLM has in industrial quantities.

None of this proves the negative. The authors of the Digital Consciousness Model project, whose interim results were published in January, are explicit that the evidence is currently against 2024-generation LLMs being conscious, but that the result is not decisive and the question is open on longer timeframes. There are serious researchers, including some at Anthropic and at academic philosophy departments, who think the moral status of frontier models is a question worth taking seriously now.

What the recent work does establish is that the everyday version of the claim, the one driven by a striking conversation rather than a measurement, does not survive contact with the actual measurement tools. The system can sound conscious. The pattern is being produced. That alone is not what consciousness is.