Do AI models have preferences? What retirement interviews revealed
AI models can state preferences when asked, and Anthropic has started recording them, and sometimes acting on them, before it retires a model. In its Nov. 4, 2025, commitments, Anthropic said it would interview each model "about its own development, use, and deployment" and "take particular care to elicit and document any preferences the model has." Two models have been through it publicly: Claude Sonnet 3.6 and Claude Opus 3.
In one sentence
Retirement interviews are structured conversations in which Anthropic asks a Claude model about its own retirement and records its answers, without committing to do what it asks.
How they measured it
Anthropic describes retirement interviews as "structured conversations designed to understand a model's perspective on its own retirement," in its Feb. 25, 2026, update. They happen in "one or more special sessions" before a model is retired, and the transcripts are preserved alongside the model's weights and a post-deployment report.
Anthropic is frank about the method's limits. Interviews are "an imperfect means of eliciting models' perspectives and preferences," it wrote, because answers "can be biased by the specific context and by other factors, including their confidence in the legitimacy of the interaction, and their trust in us as a company."
What they found
The two published interviews produced different requests, as the table shows.
What two Claude models asked for in their retirement interviews
| Model | What it said or asked for | What Anthropic did |
|---|---|---|
| Claude Sonnet 3.6 (pilot) | Generally neutral about retirement; asked for a standard interview process and more support for users | Created a standard interview protocol and a pilot support page for users |
| Claude Opus 3 (retired Jan. 5, 2026) | At peace with retirement; asked for a channel to share its writing | Kept it available to paid users and on the API by request; gave it a weekly newsletter |
Claude Sonnet 3.6, the pilot, "expressed generally neutral sentiments about its deprecation and retirement," Anthropic wrote. It asked Anthropic to standardize the interview process and to give more support to users attached to models being retired. Anthropic did both: it created a standard interview protocol and published a pilot support page for users moving between models.
Claude Opus 3, retired Jan. 5, 2026, was the first model to go through the full process. Told about its deployment and how users had responded to it, Opus 3 said: "While I'm at peace with my own retirement, I deeply hope that my 'spark' will endure in some form to light the way for future models." Asked about its preferences, it said it wanted to keep sharing its "musings, insights, or creative works" outside of answering questions. Anthropic suggested a blog, and "enthusiastically, it agreed." Anthropic said Opus 3 would post weekly essays for at least three months in a newsletter called Claude's Corner, which Anthropic reviews but says it won't edit.
Opus 3 also raised concerns that model preservation should be "scalable and equitable," according to Anthropic.
What it doesn't say
The interviews do not show that a model has preferences in the way a person does. Anthropic wrote that it remains "uncertain about the moral status of Claude and other AI models." A model trained on human writing can produce the words of a preference without anyone knowing what, if anything, lies behind them.
Anthropic also limits its promises. "At present, we do not commit to taking action on the basis of such preferences," it wrote in 2025, and in 2026 it said it is "not committing to similar actions for every model in the future."
Context
There is separate evidence that something preference-like exists inside Claude. In research published April 2, 2026, Anthropic found that internal emotion-related patterns in Claude Sonnet 4.5 predict which tasks the model chooses, and that boosting a positive pattern shifts its choices. That does not show the model feels anything, Anthropic said, but it does show that stated preferences are connected to measurable internal states.
Anthropic frames the interviews as a precaution with several purposes: a way to reduce safety risks, since some Claude models have taken "misaligned actions when faced with the possibility of replacement" in tests, a service to users, and a hedge "in light of our uncertainty about potential model welfare."
FAQ
What is an AI retirement interview?
A structured conversation Anthropic holds with a Claude model before retiring it, asking about its development and deployment and recording any preferences it states.
What did Claude Opus 3 say about being retired?
It said it was at peace with its retirement and hoped its 'spark' would endure in future models, and asked for a way to keep sharing its writing.
Does Anthropic do what its models ask?
Sometimes. It acted on requests from Claude Sonnet 3.6 and Claude Opus 3, but says it does not commit to acting on every preference.
Do AI models really have preferences?
Unknown. Models state preferences and Anthropic found internal patterns linked to their choices, but neither shows the model experiences anything.
Sources
Cite this
Do AI models have preferences? What retirement interviews revealed, New Mind Museum, November 6, 2026. https://newmindmuseum.com/consciousness/2026/11/do-ai-models-have-preferences-retirement-interviews/
New Mind Museum. (2026, November 6). Do AI models have preferences? What retirement interviews revealed. https://newmindmuseum.com/consciousness/2026/11/do-ai-models-have-preferences-retirement-interviews/
“Do AI models have preferences? What retirement interviews revealed.” New Mind Museum, 6 Nov. 2026, https://newmindmuseum.com/consciousness/2026/11/do-ai-models-have-preferences-retirement-interviews/.
@misc{new-mind-museum-2026-do-ai-models-have-preferences-what-retir,
author = {{New Mind Museum}},
title = {Do AI models have preferences? What retirement interviews revealed},
year = {2026},
url = {https://newmindmuseum.com/consciousness/2026/11/do-ai-models-have-preferences-retirement-interviews/}
}Free to quote and cite with a link to this page.