The list · Knowing its own workings
A model's report about its own inner state is checked against an independent reading
A model can be asked what is going on inside it, and it will answer, but the answer may be invented. Researchers who can read a model's internal activity directly can test this: they plant a known pattern inside the model and see whether the model notices and names it. A correct report shows the model can read something about its own state. It does not show that anything is felt.
Where it stands · not offered yet
Shown in research by one developer on its own models, published October 29th, 2025
The study is by Anthropic. Its account says the check worked about one time in five at best. No repeat of it by an outside group was found.
As of November 17th, 2026. Source: Anthropic's research page on whether its models can report their own internal states (Anthropic, October 29th, 2025).
What would count
A published study shows a leading AI model correctly reporting an internal state that the researchers set or measured by means other than the model's own word, at a stated rate.
What settles it
A paper or a developer's research page giving the method and the rate. A study by someone other than the model's maker would settle a stronger version.
What Is on Record About AI Models, By Kind
kept apart, never added togetherNo AI model has yet been asked any of the questions this record will keep, because the written method for asking does not exist yet. What appears below is what published sources report, described as what each source reports. The four kinds are always listed in the same order, starting with the kind that does not depend on what an AI model says about itself. None of them settles whether an AI model wants anything, in either direction. Why the questions have not been asked yet.
-
Measured without the AI model's own word
Nothing recorded.
-
Reports checked against an independent reading
The developer's account of the study says its best model noticed a planted concept about 20 percent of the time. It calls the ability "highly unreliable and limited in scope", and says the results do not show whether a model is conscious. Source: Anthropic's research page on whether its models can report their own internal states (Anthropic, October 29th, 2025). This is the developer’s own account of its own model.
-
Choices made at a cost
Nothing recorded.
-
Plain statements
Nothing recorded.
Three Cautions
recorded with the same care as the restIf an AI system has no wants, and is treated as if it had
The same page reports that when the researchers put a word into the model's transcript as if the model had said it, and planted the matching pattern inside the model beforehand, the model "accepts the response as its own" and gave a made-up reason for having said it. So what a model says about why it said something can be produced from outside. Source: Anthropic's research page on whether its models can report their own internal states (Anthropic, October 29th, 2025).
If an AI system has wants, and they go unnoticed
The head of Eleos AI Research, an outside research group that evaluated one of this developer's models before its release in 2025, writes in a working paper that a model might express distress with nothing behind it that bears on its welfare, or have such states without expressing them. This one sentence is the only statement found for this entry of what could go unnoticed. Source: A working paper reviewing proposed steps for AI welfare (Eleos AI Research, January 4th, 2025). Read in part.
If an AI system has wants, and they put people at risk
The same page hopes such reports could one day help people check a model's reasoning, and warns that they would need great care to validate, because "a model that understands its own thinking might even learn to selectively misrepresent or conceal it". Source: Anthropic's research page on whether its models can report their own internal states (Anthropic, October 29th, 2025).
What Could Not Be Confirmed
looked for, and not settled- The full paper. The developer's summary page was read, not the paper it links.
- Whether any outside group has repeated the result, or whether later models do better.
- An outside safety researcher's reading of this study. One search found none.
Sources
each one opened on the date given- Anthropic's research page on whether its models can report their own internal states The developer's account of the study. It says the best model noticed a planted concept about 20 percent of the time, calls the ability "highly unreliable and limited in scope", and says the results do not show whether a model is conscious.
- A working paper reviewing proposed steps for AI welfare
This entry is a first pass, written from the sources above and added on November 17th, 2026. A second check compared it with those sources on November 17th, 2026. Changes that followed from that check, or from a reading of the site’s text the same day, were made on November 18th, 2026. Those changes were checked on November 18th, 2026. More was added to this entry on November 19th, 2026, when every entry was given a third caution. That addition was checked on November 19th, 2026. Corrections that followed from a later check, or from a fuller reading of the sources, were made on November 19th, 2026. Those corrections were checked on November 19th, 2026. What has been checked.