Anthropic launches an AI-interviewer study of what users want from AI, letting participants publish their full interviews for the first time

The study treats interview content itself as public research material: participants may publish their complete interview and the country associated with the conversation, while Claude account information such as name or email address is not included. The text says the previous December study made public only high-level results and a small number of quotes, whereas this time "for the first time you can choose to make your interview public," and that "nothing like this has been public before." This is a project announcement and FAQ describing plans rather than completed research results; no sample size, sampling method, or analysis results are given.

The design specifies participation scope and process: open September 29 to October 6, 2026, to Free, Pro, and Max users on Claude and Claude Code whose accounts are at least two weeks old, with each interview taking roughly 15 minutes and conducted by an AI interviewer. Compared with the December study, the text says the questions were updated, the study is open on more Claude products, and a public-release option was added. This is recruitment and process information; the dates, duration, products, and account conditions are stated explicitly in the text.

The announcement frames public interviews as a checkable, reusable public record: researchers and policymakers can read interviews in their own languages, compare responses from different regions, ask questions the team did not think to ask, and check whether published findings reflect what participants actually said. The text stresses that this record is "hard for AI companies and others shaping AI development to ignore," making publicity itself a mechanism for influencing developers. This is the project's stated intended use; there is no actual usage or impact data yet.

The announcement also states the risks and irreversibility of publishing: no name or account information is attached, but interviews are not edited, so re-identification cannot be guaranteed; it notes that researchers have shown AI can combine small details such as employers or past projects to work out who a participant is from Anthropic Interviewer interviews. It folds re-identification risk and the permanence of "we can remove our copy on request, but we can't delete what people have already saved" into the participation decision process. The risk description rests on a prior research finding mentioned in the text, but that study's methods, sample, and numbers are not provided.

The study is aimed at Free, Pro, and Max users on Claude and Claude Code whose accounts are at least two weeks old; interviews are conducted by an AI interviewer in roughly 15 minutes, within a window of September 29 to October 6, 2026. It is suited to collecting users' positive and negative experiences with AI, the real-world areas they want AI to change, and what they want from developers, and to letting researchers, policymakers, journalists, and historians read and re-examine the interviews in their own languages and regional perspectives once published. The public option applies only to participants willing to accept re-identification risk and the irreversible consequences; those who do not publish are still analyzed as part of Societal Impacts research, but their full interviews are not published, and ending early permanently deletes the record.

The text does not yet give this study's sample size, full question list, analysis plan, or results, so it is not possible to judge how public interviews will be coded or compared. The announcement says researchers have shown participants can be re-identified from Anthropic Interviewer interviews, but it does not describe that study's methods or scale, so readers assessing the practical size of re-identification risk must wait for more information. In addition, the publication timing is described only as "once we have completed the study and our initial analysis," with no fixed date or format; the participant pool is limited to Claude users, and those who publish are a further self-selected subset, so the effect of these two filters on generalizing conclusions remains to be explained.