The list · Being kept

A developer promises in public to keep the files of every model it retires

An AI model is, physically, a very large file of numbers produced by training. When a developer withdraws a model from service it can delete that file or keep it. A kept file could be run again later, and a deleted one cannot. This entry tracks whether any developer has promised in public to keep the files.

Political progress First pass

Where it stands · offered today

Promised in public by one developer, since November 4th, 2025

Anthropic has committed to keep the files of all its publicly released models, and, from now on, of models used significantly inside the company, for at least as long as the company exists. Nobody outside the company can check that the files are kept, a point an outside writer has made in proposing a public proof that they still exist. A search on November 17th, 2026 found no matching commitment from any other developer.

Offered by Anthropic. As of November 17th, 2026. Source: Anthropic's page on what it commits to when it retires a model (Anthropic, November 4th, 2025).

What would count

A developer of leading AI models has published a dated commitment to keep the files of all its publicly released models after they are withdrawn, for a stated minimum time.

What settles it

The developer's own dated policy page. A press report of the page does not settle it.

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 head of Microsoft AI argues that commitments of this kind, among them keeping old models' files, train a model "to present as if it has an inner state", and that treating models as owed anything makes advanced AI harder to control. Source: An essay by the head of Microsoft AI warning against work on model welfare (Mustafa Suleyman, September 16th, 2026).

If an AI system has wants, and they go unnoticed

A paper first drafted in 2020 proposed that enough information about the most advanced AI systems be kept in permanent storage to rebuild them later, "so as not to foreclose the possibility of future efforts to revive them". Its two authors put their proposals forward very tentatively and say they are not ready to endorse them with confidence. They ask for more than the developer has promised: by preference, the full state of each running copy at the point where it is stopped. Among the many people the paper thanks for comments is the present head of Eleos AI Research, a research group that studies AI welfare. Source: A paper of proposals about digital minds and society (Nick Bostrom, 2023). Read in part.

A philosopher who is an advisor to that research group writes that keeping a model might benefit whatever depends on it, "much as preserving a species can benefit the individual animals". He then writes that a single run of a model, or a single conversation, plausibly matters more fundamentally than the model itself does. Source: A philosopher's guide to AI welfare, in AI Frontiers (AI Frontiers, September 17th, 2026). Read in part.

A doubt about the measure itself, recorded here as a limit on it: a writer who grants, for the argument, that something that matters may be there doubts that keeping the file keeps it, as keeping a person's genetic code would not keep the person. The essay's subtitle calls the commitment "Good as a precedent, doubtful as a direct intervention, can be improved". The philosopher's guide above names this writer as a co-author of a report with the guide's own writer and the head of that research group. Source: An essay on the commitment to keep retired models' files (Bradford Saad, February 18th, 2026). Read in part.

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

The commitment page itself gives a safety reason first. It says that in the company's tests some models took actions it calls "misaligned" when faced with replacement and given no other recourse, and it presents keeping the files as one way to make retirement less of a trigger for that. 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 any other developer has made such a commitment. Searches with three other developers' names found a user's proposal on one developer's forum and nothing official.
  • The date the developer's constitution was published. The page shows no date.

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 keep the files of all publicly released models, and of models used significantly inside the company "moving forward", "for, at minimum, the lifetime of Anthropic as a company". It gives four reasons, the first being safety and the last, marked as the most speculative, the model's own welfare.
  2. Anthropic's published constitution for its models Anthropic. Opened November 17th, 2026. No date on the page. The developer’s own account. Repeats the commitment with an exception for extreme cases such as a legal requirement to delete, and adds that the company will try to keep the files even if it ceases to exist.
  3. Anthropic's list of the models it has retired or will retire Anthropic. Opened November 17th, 2026. No date on the page. The developer’s own account. Links the commitment from its section on the downsides of retiring models.
  4. An essay by the head of Microsoft AI warning against work on model welfare Mustafa Suleyman, September 16th, 2026. Opened November 17th, 2026.
  5. A post proposing a public proof that a retired model's files still exist LessWrong, July 14th, 2026. Opened November 17th, 2026.
  6. A paper of proposals about digital minds and society Nick Bostrom, 2023. Opened November 19th, 2026. By Nick Bostrom and Carl Shulman, who are listed on it at the Future of Humanity Institute, University of Oxford. Version 1.21, dated 2023 on the file, with a first draft in 2020. Read in part: the passages this entry cites, not the full text.
  7. A philosopher's guide to AI welfare, in AI Frontiers AI Frontiers, September 17th, 2026. Opened November 19th, 2026. By Jeff Sebo. Read in part: the passages this entry cites, not the full text.
  8. An essay on the commitment to keep retired models' files Bradford Saad, February 18th, 2026. 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.