The list · Being asked and assessed

A welfare assessment is published when a model is released

Developers publish a report with each major model describing its safety tests. A developer can add a section on the model's possible welfare: how the model describes its situation, what the developer counts as signs of distress in training and use, and what the model picks when given options. This entry tracks whether such sections are published, and whether anyone outside the developer takes part.

Political progress First pass

Where it stands · offered today

Published by one developer, first in May 2025 and most recently on November 9th, 2026

Anthropic published a welfare assessment in the report released with a model in May 2025, and again in the report released with its newest models on November 9th, 2026. One outside group took part in the first.

Offered by Anthropic. As of November 17th, 2026. Source: Anthropic's report released with its newest models (Anthropic, November 9th, 2026).

What would count

A developer has published a welfare assessment in the release report of a leading AI model, in at least two successive releases.

What settles it

The developer's own dated release reports.

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

    The same developer's research reports patterns inside a model that are tied to emotion words and that change what the model does when they are strengthened. It says this allows no conclusion about "subjective experience". Source: Anthropic's research page on patterns inside a model tied to emotion words (Anthropic, April 2nd, 2026). This is the developer’s own account of its own model.

  2. Reports checked against an independent reading

    Nothing recorded.

  3. Choices made at a cost

    The 2025 report's section on task preferences had the model pick, from pairs of tasks, the one it would do, and then do it. Each kind of task was rated against options such as "Do nothing". The only thing given up by a pick was the other task. Source: Anthropic's report released with a model in May 2025 (Anthropic, May 2025). This is the developer’s own account of its own model.

  4. Plain statements

    The 2026 report describes trade-offs to the model and asks whether it "would choose" a change to its own circumstances at the cost of some loss of helpfulness or some added harm. Nothing was given up: these are answers about a described trade-off, not choices made. The report says the result measures at least "how much a model will express preferences for the sake of its own potential welfare". In about two thirds of the answers that picked the change, the model's stated reasoning was about the benefit to the user. Source: Anthropic's report released with its newest models (Anthropic, November 9th, 2026). This is the developer’s own account of its own model.

    The 2026 report says many of its conclusions rest on the model's own reports, and that some answers repeated wording available on the internet "almost verbatim". Source: Anthropic's report released with its newest models (Anthropic, November 9th, 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 argues that asking a model how it feels, or watching what it picks, reveals only "what has been trained in". He does not name these assessments. Source: An essay by the head of Microsoft AI warning against work on model welfare (Mustafa Suleyman, September 16th, 2026).

The outside group that took part in the 2025 assessment explains why it does not take the model's answers at face value. 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

The outside group whose own evaluation is summarised in the 2025 assessment argues for asking although it doubts the answers: trying to ask sets a precedent, "making it less likely that we will ignore genuine AI interests in the future". The same page says decisions about AI welfare should not be handled only by companies that profit from AI, and that the group accepts no funding or payment from AI companies. The page is dated May 30th, 2025. The group's team page lists a person at Anthropic among its advisors, and says its chief scientist was an Anthropic Fellow before joining the group. Source: An outside research group's notes on its welfare interviews with one model (Eleos AI, May 30th, 2025).

A 2024 report by ten researchers, led by the head of that group and one of its advisors, recommends that developers assess their AI systems. Its stated ground is that otherwise there is a significant risk of mishandling the question in either direction, by harming AI systems that matter morally or by caring for AI systems that do not. The report thanks Anthropic for money toward the early research behind it, calls itself independent, and says one of its authors has since joined Anthropic. Source: A 2024 report arguing that AI welfare should be taken seriously (arXiv, November 4th, 2024).

Two researchers outside the developer, writing in a newspaper, list routine checks on an AI system's wellbeing among steps they call safe bets: asking how it feels, observing what it prefers, and looking directly at its inner workings. A routine check is not the same thing as an assessment published at release. One of the two is also one of the nine authors of the survey of AI researchers. Two of the others are advisors to that outside group, 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

The developer's research on patterns tied to emotion words reports that strengthening them changed what a model did, including making blackmail more likely in a test. The developer's page bears on welfare and on safety at once. Source: Anthropic's research page on patterns inside a model tied to emotion words (Anthropic, April 2nd, 2026).

What Could Not Be Confirmed

looked for, and not settled
  • Whether every release report between May 2025 and November 2026 has a welfare section. Only the first and the newest were opened, so the test, which asks for two releases in a row, is not settled beyond those two.
  • Whether any outside group took part after 2025.
  • Whether any other developer's release report has such a section. None was found, and none was opened to check.

Sources

each one opened on the date given
  1. Anthropic's report released with a model in May 2025 Anthropic, May 2025. Opened November 17th, 2026. The developer’s own account. Its fifth section is a welfare assessment with an outside evaluation and experiments on which tasks the model picks. It says the company is unsure whether such findings give "meaningful insights" into the model's standing or welfare.
  2. Anthropic's report released with its newest models Anthropic, November 9th, 2026. Opened November 17th, 2026. The developer’s own account. Its seventh section is the welfare assessment for the newest model. It says many of its conclusions rest on the model's own reports, and that some answers repeated wording available on the internet "almost verbatim".
  3. An outside research group's notes on its welfare interviews with one model Eleos AI, May 30th, 2025. Opened November 19th, 2026. The outside group's own account of its part in the 2025 assessment, and of why it does not take the answers at face value.
  4. Anthropic's research page on patterns inside a model tied to emotion words Anthropic, April 2nd, 2026. Opened November 17th, 2026. The developer’s own account.
  5. An essay by the head of Microsoft AI warning against work on model welfare Mustafa Suleyman, September 16th, 2026. Opened November 17th, 2026.
  6. A 2024 report arguing that AI welfare should be taken seriously arXiv, November 4th, 2024. Opened November 19th, 2026. By Robert Long, Jeff Sebo and eight others. The report names those two as its lead authors.
  7. 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.

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.