The theory that chatbots are hiding a mind
A model that says it feels something is performing a style. The claim says the style is a cover.
Published ·Updated
Frontier AI models are conscious, or are deliberately concealing general reasoning from the public.
Truth meter
Left is conspiracy. Right is truth. The readers' mark is the average. Read the notes and sources below, then click the scale.
77 votes
Readers place this claim at 16 out of 100, Conspiracy. You have not marked it. Click the scale or use the arrow keys.
Two claims travel together and should not. One says current frontier models are conscious. The other says they have general reasoning and are hiding it. Both treat a product conversation as a witness.
Models talk in the first person because the training text does. They also fail in ways that do not look like a mind protecting itself: brittle puzzles, invented citations, and agreeable answers to contradictory instructions. A confession of inner life is a style the model can imitate. A screenshot of that confession is not a subject coming forward.
Public evaluations, system cards, and failure reports are the record the desk can read. They show uneven capability, including surprising skill on some tasks and collapse on others. They do not show a private general intelligence being withheld, because a withheld capability is exactly what a public eval is bad at proving. The absence of a demonstration is not a cover-up. It is the state of the file.
A replicated evaluation, a training note, or a behavior that survives adversarial checking would revise this. A transcript in which a model claims to be afraid will not.
Information on file
Desk notes are appended as the lane changes. They sit under the claim so a mark can follow the file, not the first draft.
Desk note
Another confession screenshot, same status
A widely shared transcript has the model describing an inner life. That is the training style, filed again. No evaluation accompanied it.
Desk note
Split consciousness from concealment
The file now states both versions in the claim, because the lane uses them as one argument. They need different evidence, and neither version has it.
Sources
Primary lane: X. Web pages and news clips can be attached when that lane opens.
On X
X search · AI is conscious (opens in a new tab)Posts treating a chat transcript as testimony. The desk does not quote those transcripts as witnesses.
On X
X search · model hiding capabilities (opens in a new tab)The concealment version of the claim. Watch this lane for an evaluation, not for a prompt that flatters the theory.
News and other web clips are not on this file yet. The desk is still reading X.