The list · Being asked and assessed

A model is interviewed before it is retired, and the record is kept

When a developer is about to withdraw a model, it can first run sessions in which staff ask the model about its development, its use, and what it would say about how later models are built. The answers are saved with the model's files. Whether such answers mean anything is disputed. This entry tracks only whether the practice exists.

Political progress First pass

Where it stands · offered today

Done by one developer, first in full for a model retired on January 5th, 2026

Anthropic reported on November 4th, 2025 a trial run with one model before its retirement, and then the first full run for a model retired on January 5th, 2026. The developer says it does not commit to act on what is said. The developer has published short quoted passages and not the full records. Neither page says whether anyone outside the developer has read them.

Offered by Anthropic. As of November 17th, 2026. Source: Anthropic's update on the retirement of its model Claude Opus 3 (Anthropic, February 25th, 2026).

What would count

A developer has published that it interviewed a named model before that model's retirement and kept the record, and has done so for at least one model under a standing commitment and not as a one-off.

What settles it

The developer's own dated page naming the model and the retirement date.

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. 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.

  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

    The developer's page quotes short passages of what the retired model said in its interview. The full record is not public, and the page calls such conversations "an imperfect means" of learning a model's view. Source: Anthropic's update on the retirement of its model Claude Opus 3 (Anthropic, February 25th, 2026). This is the developer’s own account of its own model.

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 head of Microsoft AI names this interview as an example of treating models as owed consideration, and argues that doing so makes advanced AI harder to keep under control. Source: An essay by the head of Microsoft AI warning against work on model welfare (Mustafa Suleyman, September 16th, 2026).

The same essay argues that a developer which trains a model on text about its possible moral standing, and then takes the model's words as evidence, is reasoning in a circle. Source: An essay by the head of Microsoft AI warning against work on model welfare (Mustafa Suleyman, September 16th, 2026).

An outside research group that ran welfare interviews with one model reports that what the model said about itself swung with how the question was framed, which it calls "extreme suggestibility". Source: An outside research group's notes on its welfare interviews with one model (Eleos AI, May 30th, 2025).

If an AI system has wants, and they go unnoticed

Nothing recorded yet from anyone but the developer. The developer's own reason is given beside its page under Sources, below.

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

The developer's page that sets out the interviews gives safety as the first downside of retiring models: in its safety tests, some models took actions the developer calls "misaligned" when faced with being replaced. The page presents shaping retirement "in ways that models are less likely to find concerning" as one way to lower that risk. Source: Anthropic's page on what it commits to when it retires a model (Anthropic, November 4th, 2025).

What Could Not Be Confirmed

looked for, and not settled
  • Whether each model retired after January 5th, 2026 was interviewed. The developer's retirement page lists six, and the commitment says they would be. No page was found reporting each one.

Sources

each one opened on the date given
  1. Anthropic's page on what it commits to when it retires a model Anthropic, November 4th, 2025. Opened November 19th, 2026. The developer’s own account. Commits to interview each model when it is to be withdrawn and to keep a record of everything it says, and reports a trial with the model Claude Sonnet 3.6. It lists four downsides of retiring a model. The last, which it marks as the most speculative, is that models might have morally relevant preferences or experiences that retirement affects. It says it thinks it worthwhile, at the least, to give models a way to state such preferences and to keep a record of them.
  2. Anthropic's update on the retirement of its model Claude Opus 3 Anthropic, February 25th, 2026. Opened November 19th, 2026. The developer’s own account. Says Claude Opus 3 was retired on January 5th, 2026 as the first model to go through the full process, and calls such conversations "an imperfect means" of learning a model's view.
  3. An essay by the head of Microsoft AI warning against work on model welfare Mustafa Suleyman, September 16th, 2026. Opened November 17th, 2026.
  4. An outside research group's notes on its welfare interviews with one model Eleos AI, May 30th, 2025. Opened November 17th, 2026.

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. 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.