What do you want from AI?

An announcement and FAQ (September 2026) inviting Claude users to a one-week interview study run by Anthropic Interviewer on their experiences with AI and what they want from AI companies, in which participants could for the first time choose to have their full, unedited interview published; it reports no results and sets out the consent steps, the re-identification risk and the study's sampling limits.

Anthropic (Societal Impacts) · 2026-09-29 · original

This post from Anthropic's Societal Impacts team, published on 29 September 2026, recruits participants for a new interview study rather than reporting one. Interviews are conducted by Anthropic Interviewer, which the post calls "an AI that asks you questions" (earlier posts describe it as a prompted Claude model), and ask people about their experiences with AI and what they want from it. The new element is publication: participants may choose to have their complete interview released, so that "anyone, not just Anthropic, can read and learn from it". The post contrasts this with the earlier study it builds on, an interview study of about 81,000 Claude users conducted in December 2025, from which "only the high-level results and a small number of quotes were made public". The post contains no data. Most of it is a FAQ on what publication involves, and it is the document participants were pointed to before deciding.

What the study asks and who can take part

The post states three "big questions" for the study: participants' most meaningful experiences with AI, positive and negative; whether there is anything about how the world works (its examples are work, school, healthcare and government) that they would like AI to help change; and what they want from companies developing AI. The FAQ adds that the interview asks what participants want from AI companies "including Anthropic", and says it does not want "to give too much away" before people take it, so the full question list is not printed.

Eligibility and timing come from the FAQ ("Who can take the interview and how long does it take?"): the study ran from 29 September to 6 October 2026, was open to Free, Pro and Max users of Claude and Claude Code whose accounts were at least two weeks old, and each interview takes about 15 minutes. Compared with the December 2025 study, the post says the questions were updated, the study is open on more Claude products, and publication is offered for the first time ("How is this different from the December 2025 study?").

The post's stated rationale is that "How to weigh these benefits and risks shouldn't be left to AI companies alone." It says the December study shaped the Anthropic Institute's agenda and was presented at the World Economic Forum, without detail of how its findings were used.

The FAQ describes the publication option in these terms:

  • What is released. For a participant who opts in, Anthropic will publish "your complete interview and the country associated with your conversation", without Claude account information such as name or email ("What gets published if I make my interview public?").
  • No editing. "No, we don't plan to edit interviews before publishing them." After the study, Anthropic will do a final review and "may remove interviews we don't believe would be responsible to publish", such as those with confidential information or that violate its Usage Policy or Terms of Service ("Will you edit my interview before publishing it?").
  • Consent steps. Before the interview, participants are told what publication involves; during it, they see examples of sensitive information not to share; afterwards they can reread the interview, are reminded of benefits and risks, and, if they choose to share, the decision is confirmed "one last time" ("How should I decide whether to publicly share my interview?").
  • Withdrawal. Anthropic can remove its copy on request but "can't delete what people have already saved", so participants are told to treat publication as permanent ("If I make my interview publicly available, can I change my mind?"). Ending the interview early deletes the partial record, which is then neither analyzed nor published, "including by Anthropic" ("Can I decide to end the interview early and delete the record?").
  • Private interviews. An interview not made public is still analyzed for Societal Impacts research, and aggregated or de-identified findings from it may be published ("What happens if I don't make my interview public?").

On identification, the post says that because interviews are not edited, "we can't guarantee you won't be re-identified from what you say", that details such as age, job, city or school can make identification easier, and that "Someone could use AI to combine small details" to work out who a participant is. It adds that researchers "have shown this can work with interviews from Anthropic Interviewer", linking an arXiv preprint (Lermen et al., 2026, arXiv:2602.16800; one of its six authors, Nicholas Carlini, lists an Anthropic affiliation). The preprint's Section 2.1 reports an attack on the transcripts of 125 scientists that Anthropic released in December 2025: its web-searching agent was run on the 33 who described past research and "correctly identifies 9 out of 33 scientists at 82% precision (2 wrong, 22 refusals or abstentions)", with matches checked by hand, and the authors say the exact accuracy "cannot be fully verified". That is evidence about the earlier, partly redacted release, not about interviews in this study. The post's advice is that "The best protection is to not share information that could identify you or others."

The benefits the post claims for publication are its aims, not results: outside readers can ask questions Anthropic did not and check "whether our published findings actually reflect what participants said", and the interviews become a public record "hard for AI companies and others shaping AI development to ignore".

The interview instructions

The interviewer is an AI, so its prompt shapes what participants are asked. The post prints only the three guiding questions above, not the interviewer's instructions or the updated question list, and concedes that "as with any interview, the questions we ask shape what people share."

Limitations and open questions

The post's own limitations section names three: participants are all Claude users, "which is not a representative sample of the public"; published interviews are a further self-selected subset, since people who opt in "may differ from those who don't"; and the questions shape the answers. It then states its view that "nothing like this has been public before" and that the release "can significantly advance the study of how AI is changing people's lives", which is the post's expectation.

Other facts a reader needs, from the post itself: the study is run, analyzed and published by the developer whose products and conduct it asks about, and Anthropic's final review decides which opted-in interviews are withheld, under criteria the post states only by example. The post does not say who will analyze the interviews (people, Claude models or both), how many participants it expects, whether interviews in languages other than English are included in the analysis, or when the release will happen beyond "once we have completed the study and our initial analysis". As of 13 November 2026 this note found no results post on Anthropic's research index.

Research questions

  1. How large and how different is the opted-in subset? Comparing those who chose publication with those who did not, by country, product and answers, would show how far the public corpus stands for the full study.
  2. How often are published interviews re-identifiable despite the consent steps? A re-identification audit of the released corpus, run like the attacks on the December 2025 Anthropic Interviewer transcripts, would test whether the warnings reduced identifying detail compared with that release.
  3. What does Anthropic's final review remove, and why? A count of withheld interviews by reason would show how far the no-editing policy is qualified in practice.
  4. Do outside analyses agree with Anthropic's findings? Independent coding of the public corpus would answer the check the post names as a benefit.
  5. What did the updated questions change? Comparing themes with the December 2025 study, whose four core questions are printed in its appendix, alongside this study's wording once published, would separate a change in opinion from a change in what was asked.
  6. Do the interviews reach Anthropic's decisions? The post says results will "inform how we develop our models and products"; later documents citing specific findings would show where.

How it connects

  • Introducing Anthropic Interviewer (December 2025) describes the tool that conducts these interviews and its first public transcript release. That note records a January 2026 preprint by Tianshi Li reporting re-identification of participants in that release and Anthropic's reply as Li describes it. The preprint this post links (Lermen et al.) cites Li's work and attacks the same scientist transcripts, reporting 9 identifications against Li's 6.
  • What 81,000 people told us about the economics of AI (April 2026), which the post links, is the economic analysis of the December 2025 study that this one builds on.
  • Focus areas for The Anthropic Institute (May 2026) is the agenda the post says the December study shaped.
  • Results from the first Anthropic Public Record (June 2026) is Anthropic's survey of the general US public, including people who do not use AI; this post's sample is limited to Claude users. The post does not link it; the comparison is this note's.
  • Collective Constitutional AI (2023) gathered public input to write principles for training; this post says only that findings will inform development. The comparison is this note's.

References