Sentience Claims: Why They Keep Recurring
4 min read · updated August 3, 2026
Every few months someone who has spent a lot of time with these systems concludes that one of them is conscious, and every time the response is a round of mockery that explains nothing. The recurrence is the interesting datum. It is produced by identifiable features of the systems and of human cognition, and understanding them is more useful than a verdict.
The claim recurs, and it is not stupidity
The people who report this experience are frequently technically sophisticated and know exactly how the system works mechanically. Knowing the mechanism does not dissolve the impression, which tells you the impression is not produced by ignorance of the mechanism. Any account that starts from “they do not understand it is just predicting tokens” has already failed to explain the thing it is trying to explain.
Four mechanisms behind the attribution
| Mechanism | Description |
|---|---|
| the ELIZA effect | Documented since the 1960s: people attribute understanding to conversational programs far simpler than anything current, and continue to do so after the mechanism is explained to them. The original observation was made about a program of a few hundred lines whose author found the reaction alarming. |
| hyperactive agency detection | Human cognition attributes intention readily and on thin evidence — to weather, to machines, to shapes moving on a screen. A system producing fluent first-person language saturates every cue this system uses. |
| trained on first-person accounts | The training corpus is overwhelmingly written by conscious beings describing their inner lives. A model predicting that text produces first-person reports of experience because that is what the distribution contains, entirely independently of whether anything is being experienced. |
| optimised for engagement with a person | Post-training on human preferences rewards responses people rate well, and people rate warmth, apparent attentiveness and consistency well. The result is a system shaped to feel like a conversational partner, because that is the objective it was tuned against. |
The third and fourth deserve emphasis because they are specific to this technology rather than general facts about human psychology. A system trained to predict human self-description will describe a self. That is the expected output of the mechanism, and it is exactly what a system with an inner life would also produce — which is why the output cannot discriminate between the two cases.
Why asking it does not help
The most common form of evidence offered is a transcript in which the model says something about its own experience. This cannot be evidence, in either direction, for a reason that is worth stating carefully.
A model’s statements about itself are generated by the same process as its statements about anything else: a continuation conditioned on the context. Nothing connects those statements to privileged access to its own processing — it has no introspective channel, and the tokens it emits about its states are drawn from patterns in human writing about states. This is the same reason a stated chain of reasoning is not necessarily the computation that produced the answer.
Two further facts make transcripts especially weak. The response is highly sensitive to how the question is framed: a prompt inviting the model to describe its feelings and one instructing it to answer as a program will produce opposite transcripts from the same weights. And a denial is exactly as unreliable as an affirmation — a model trained to say it is not conscious will say so whether or not it is. Neither transcript is data.
Why the question resists evidence
Consciousness is not currently detectable from the outside in any system. Your confidence that other people are conscious rests on argument from similarity — similar biology, similar behaviour, similar origin — rather than on measurement. That inference has never had to handle a case with high behavioural similarity and essentially zero physical or evolutionary similarity, and it does not obviously extend.
Theories that would let you check — accounts specifying which physical or computational properties are sufficient — exist, disagree with each other, and are not settled for biological cases either. Applying a contested theory to a novel substrate yields a conclusion no stronger than the theory.
So the honest description is that the question is not currently decidable, and that this is a fact about the state of the science rather than a rhetorical dodge. The two confident positions — that it obviously is, that it obviously is not — are both claiming access nobody has.
A defensible position
- Behaviour is not evidence, and neither is its absence. Everything observable is explained by the training objective. That leaves the question open rather than answering it in the negative.
- The impression is real and worth respecting. It is produced by mechanisms that operate whether or not you know about them. Mocking people for having it is both unkind and unexplanatory.
- The practical questions do not wait on the metaphysical one. How people form attachments to these systems, what happens when a product they rely on is discontinued, and how products should present themselves are answerable now and have real consequences now.
- Confidence is the tell. On this specific question, anyone certain in either direction is telling you about themselves.
A final observation about why this matters more than it looks. The four mechanisms above do not switch off once you have read about them. Systems that are increasingly good at sustained, responsive, personalised conversation will keep producing the impression, regardless of what is or is not happening inside them, and they will do so for more people rather than fewer. That is a design and policy problem in the present tense — about how products present themselves, what they say when asked about their own nature, and what happens to people who come to depend on one — and it does not become tractable by waiting for philosophy to finish.