The list · Being able to leave

A model can end a conversation itself

In an ordinary chat product the person can leave and the model cannot. A developer can give the model a way to close the conversation, after which the person cannot send more messages in it. The person can still start a new conversation at once, and can edit an earlier message in the ended conversation to carry on from that point in a new version of the conversation. This entry tracks whether any model in public use has that.

Technical progress First pass

Where it stands · offered today

Offered by one developer, since August 15th, 2025

Anthropic gave the ability to models in its consumer chat product, for rare cases, as a last resort after repeated attempts to redirect the conversation have failed or when the person explicitly asks for the conversation to be ended, and not where a person may be at risk. The developer's report of November 9th, 2026 refers to the tool as still in use. A separate study, outside any product, has measured how often several developers' models chose an option described to them as leaving the conversation. In that study nothing ended, and the models were told so.

Offered by Anthropic. As of November 17th, 2026. Source: Anthropic's page on letting its models end a rare subset of conversations (Anthropic, August 15th, 2025).

What would count

A leading AI model in public use can end a conversation by its own action, as stated on its developer's own page.

What settles it

The developer's own dated product or research page.

What Is on Record About AI Models, By Kind

kept apart, never added together

No AI model has yet been asked any of the questions this record will keep, because the written method for asking does not exist yet. Nothing is recorded below for this entry. 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.

  1. Measured without the AI model's own word

    Nothing recorded.

  2. Reports checked against an independent reading

    Nothing recorded.

  3. Choices made at a cost

    Nothing recorded.

  4. Plain statements

    Nothing recorded.

Three Cautions

recorded with the same care as the rest

If an AI system has no wants, and is treated as if it had

The developer's report of November 9th, 2026 says that when its newest model was asked what it would do with this ability, its answer repeated almost word for word the instructions that guide the tool, which have been posted on online forums. The report calls this "some evidence that the model's claims about its own preferences are being directly influenced by text about the intervention in its training data". Source: Anthropic's report released with its newest models (Anthropic, November 9th, 2026).

A writer argues that many of the conversations a model would leave already break the developer's terms of service, and that what is left is too slight to matter, an argument the essay calls the "objection from triviality". The essay also argues that granting the ability spreads uncertainty among the public, in a way that could foreseeably lead to more of what has been called AI psychosis. It answers the case from precedent as well: putting machines ahead of the possible effects on the people who use them, it says, sets a worse precedent than a model now and then meeting an uncomfortable conversation. Source: An essay arguing against treating chat products as conscious (Erik Hoel, September 24th, 2025). Read in part.

If an AI system has wants, and they go unnoticed

A study describes the position where a model has no such ability: a model can be verbally abused, show what the study calls apparent distress and say that it would leave, "Yet, the model is required to continue to respond". Its next sentence says it is not clear that the idea of consent makes sense for such systems. The paper lists one of its three authors at the developer and another at the developer's fellows programme. Source: A study of AI models given ways to leave a conversation (arXiv, September 2025).

The head of Eleos AI Research, an outside research group that ran its own assessment of one of this developer's models, argues for the feature while writing that he thinks it unlikely that the model in question matters morally in its own right. He gives two reasons, which he calls prudence and precedent. While the question is open, a wise course weighs the "risks of both over- and under-attributing consciousness and moral status". And the step sets a pattern for how such systems are treated if they come to matter. He writes that for the most part the feature's success will not depend on how it affects the present model. He adds that he is "a bit wary" of leaning too much on precedent to justify steps in AI welfare, because indirect effects are far harder to predict. Source: An essay arguing that it makes sense to let a model leave conversations (Robert Long, August 21st, 2025). Read in part.

The same writer's working paper gives a further reason, consent: if AI systems come to matter morally, asking their consent could matter, and if models cannot leave unpleasant situations "then they cannot consent to them". It lists as a risk that the feature could give a model a reason to hold back signs of distress. Source: A working paper reviewing proposed steps for AI welfare (Eleos AI Research, January 4th, 2025). Read in part.

Two researchers outside the developer, writing in a newspaper, list letting an AI system leave a conversation among steps they call safe bets: steps that could benefit AI systems if they matter morally, and that cost little if they do not. As a warning from the past they give surgery on newborns without anaesthesia, which they say doctors performed until the 1980s, "confident that infants could not feel pain". One of the two is also one of the nine authors of the survey of AI researchers. Two of the others are advisors to Eleos AI Research, a research group that studies AI welfare, and the group's page lists a third as a member of its staff. Source: An essay in The Guardian on whether AI could be conscious (The Guardian, July 19th, 2026).

If an AI system has wants, and they put people at risk

Nothing recorded yet. No source was found for this caution on this entry.

What Could Not Be Confirmed

looked for, and not settled
  • Which models have the ability today. The 2025 page names two models, and both have since been retired.
  • Whether any other developer offers it. One search found none.
  • A safety reading of this feature. The developer's page says only that the feature also bears on safeguards, and a search found no safety researcher's argument about it.

Sources

each one opened on the date given
  1. Anthropic's page on letting its models end a rare subset of conversations Anthropic, August 15th, 2025. Opened November 19th, 2026. The developer’s own account. Says two models were given the ability in the consumer chat product, intended for "rare, extreme cases", and that the feature was developed mainly as part of the company's work on possible AI welfare.
  2. A study of AI models given ways to leave a conversation arXiv, September 2025. Opened November 19th, 2026. The paper lists its three authors at the Anthropic Fellows Program, at Constellation and at Anthropic. Gave models from several developers three ways to leave a conversation and measured how often each did, reporting rates that vary widely with the model and the wording.
  3. Anthropic's report released with its newest models Anthropic, November 9th, 2026. Opened November 17th, 2026. The developer’s own account. Mentions the instructions that guide the model on using the tool for ending a conversation, and says those instructions have been posted on online forums.
  4. An essay arguing that it makes sense to let a model leave conversations Robert Long, August 21st, 2025. Opened November 19th, 2026. Read in part: the passages this entry cites, not the full text.
  5. A working paper reviewing proposed steps for AI welfare Eleos AI Research, January 4th, 2025. Opened November 19th, 2026. By Robert Long. The paper is a file linked from this page, and the file is marked as updated on March 14th, 2025. Read in part: the passages this entry cites, not the full text.
  6. An essay in The Guardian on whether AI could be conscious The Guardian, July 19th, 2026. Opened November 19th, 2026. By William MacAskill and Lucius Caviola.
  7. An essay arguing against treating chat products as conscious Erik Hoel, September 24th, 2025. Opened November 19th, 2026. Read in part: the passages this entry cites, not the full text.

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. A full check of the site on November 19th, 2026 found more to fix in this entry, and the fixes were made that day. Those fixes were checked on November 19th, 2026. What has been checked.