What do you want from AI? news update: What do you want from AI?

Anthropic launched a public qualitative research study using Anthropic Interviewer to record and evaluate user perspectives on artificial intelligence, as detailed in Anthropic Research's report. The interactive interview initiative runs from September 29 to October 6, 2026, gathering firsthand accounts directly from active model users across consumer and coding environments. For the first time, respondents can elect to have their completed interview transcripts published online for independent review by outside researchers, historians, and policymakers.

This report examines how the automated study operates, who qualifies to take part, the data privacy controls offered to participants, and how the collected material feeds into ongoing safety research. It is written for AI developers, policy analysts, enterprise engineering teams, and Claude subscribers seeking to understand how user feedback directs model development priorities.

01Automated interviews with Anthropic Interviewer

Anthropic is deploying Anthropic Interviewer, an automated conversational system that conducts interactive questioning sessions lasting roughly fifteen minutes. The tool presents participants with open-ended prompts concerning their most significant positive and negative moments with AI systems. The questioning also explores institutional shifts across schools, workplaces, healthcare systems, and government functions, while asking users what behaviors they expect from developers building frontier models.

The system allows participants to review their entire dialogue before deciding whether to contribute their answers to the open research repository. Unlike standard static questionnaires, Anthropic Interviewer probes context based on responses while advising users against submitting sensitive identification markers. Participants can exit the process at any point prior to final submission.

1. Start interview Access the conversational interface within supported user accounts.
2. Respond to prompts Answer questions regarding personal AI encounters and institutional expectations.
3. Review sensitive data warnings Check dialogue records against built-in privacy guidance.
4. Inspect full transcript Reread complete logged exchanges before selecting publishing preferences.
5. Confirm publication status Make a final choice to release the transcript publicly or restrict it to internal researchers.

02Expansion from the previous study

The current program builds upon an earlier study conducted in December 2025, which gathered AI expectations and concerns from 81,000 respondents. Feedback from that initial effort directed research initiatives at the Anthropic Institute, guided internal Societal Impacts and Economics research, and was presented to international policymakers at the World Economic Forum. However, the December 2025 project restricted public disclosure to high-level statistical summaries and a limited selection of approved quotes.

The 2026 update broadens both reach and transparency across model ecosystems. The research team updated core inquiry questions to capture shifting user attitudes and opened testing access across expanded product tiers. Most notably, the new protocol provides an option for participants to share their unabridged interview texts openly with the broader public.

03Public record distribution and re-identification risks

Transcripts marked for public release will appear online accompanied only by the country associated with the participant's conversation. Anthropic does not append Claude account details such as registered account names or email addresses to the files. Outside investigators, journalists, and regional policymakers will receive unrestricted access to read records in original languages, audit corporate conclusions against source testimony, and run external analytical queries.

Public release carries permanence risks because external visitors can download and store files locally. Anthropic can delete its hosted copy upon request but cannot retrieve or purge cached copies retained by third parties.

Although direct account handles are omitted, unedited interviews might allow outside parties to piece together identities using occupational histories, past projects, or locations. Researchers have previously demonstrated that re-identification using AI synthesis remains possible with Anthropic Interviewer transcripts.

Note: Published transcripts are open to permanent copying by third parties, meaning participants cannot undo third-party storage if they change their minds later.

04User eligibility and early exit protocols

Eligibility for the study is limited to Free, Pro, and Max subscribers on Claude and Claude Code whose accounts are at least two weeks old. Users who choose not to release their conversations publicly can still complete the session for internal review; Anthropic includes those submissions in Societal Impacts research while withholding raw text from public repositories. Aggregate findings and de-identified figures from internal sessions may still appear in downstream publications.

If a participant objects to internal research use while an interview is underway, they can click the End interview button. Triggering this command ends the session instantly and permanently purges the partial exchange from internal databases. Once deleted, the discarded content will not undergo model analysis or appear in any research papers published by Anthropic.

Anthropic does not edit transcripts before publication, though it conducts a screening review to withhold records that violate its Usage Policy, breach Terms of Service, or expose confidential information.

05Sample limits and research applications

Collected interviews will guide Anthropic as it trains upcoming foundation models and refines product features. The study evaluates the balance between frontier model breakthroughs across medicine and science and the growing hazards of misuse, where a single security breach can cause significantly greater damage than it did twelve months prior. Published datasets will give outside watchdogs permanent documentation to hold model creators accountable to stated user requirements.

The research team noted structural constraints inherent in the study population. The participant group comprises active Claude users rather than a representative cross-section of the general population.

Furthermore, participants who approve public release represent a specific sub-tier whose perspectives may diverge from users selecting private processing. Anthropic plans to publish the raw interview archive and its accompanying analytical report online following post-study evaluations.