September 29, 2026 — Anthropic opened a new What do you want from AI? study on Anthropic Interviewer. The hook is not another survey slider. After a ~15-minute chat, you can choose to publish the complete transcript plus country, with no name or email attached.
That is a real change from December 2025, when ~81,000 Claude users sat through Interviewer and the public only got high-level results and a handful of quotes. This round is also open on Claude Code, not only consumer Claude.
The window is short: September 29 through October 6, 2026. If you have a two-week-old Free, Pro, or Max account, you can sit it today. The rest of this post is the question people actually ask after they see the headline: should you make it public?
TL;DR — should a builder take this, and go public?
| Question | Direct answer |
|---|---|
| Who is eligible? | Free / Pro / Max on Claude or Claude Code; account ≥ two weeks old |
| How long? | About 15 minutes, Interviewer asks |
| Must I publish? | No. Analysis still happens unless you hit End interview early (that deletes the partial) |
| What is public? | Full transcript + country. No Claude name/email. Unedited, with a later “responsible to publish” filter |
| Can I undo public? | Not fully. Anthropic can pull its copy. Saved copies elsewhere stay |
| Re-ID risk? | Anthropic says yes, possible. Researchers already showed it on Interviewer text |
| Sample? | Claude users, not the public. Public subset is whoever opts in |
| Why bother? | Anthropic says a public corpus is harder for labs and policymakers to ignore than a private survey |
Primary: anthropic.com/research/your-thoughts-on-ai.
What Anthropic is actually asking
The company lists three “big questions,” then refuses to spoil the script:
- Most meaningful positive and negative experiences with AI.
- Anything about work, school, healthcare, or government you want AI to help change.
- What you want from companies developing AI, including Anthropic.
That last item is the one worth taking if you ship on Claude. December’s 81,000-user round shaped the Anthropic Institute agenda and was shown at the World Economic Forum. The July 2026 “hard questions” campaign already used Interviewer as a pillar. This study is the follow-up after models, products, and user sentiment moved.
Anthropic’s own framing is the usual 2026 pair: more useful, more dangerous; science and medicine on one side, a single security incident costing more than a year ago on the other. For the capability side of that sentence, see this week’s GLM-5.3 cyber eval. Interviewer is the user-voice layer, not a red-team paper.
Who can sit it (and who cannot)
Anthropic is explicit:
- Dates: September 29 – October 6, 2026.
- Plans: Free, Pro, and Max.
- Surfaces: Claude and Claude Code.
- Account age: at least two weeks.
- Length: ~15 minutes.
- Interviewer: an AI, not a human moderator.
If your Claude Code seat is on a team workspace but the login is younger than two weeks, you wait. If you only have API keys and no Claude/Claude Code product login, this form is not for you.
Ending early is the nuclear option: End interview permanently deletes the partial. It will not be analyzed or published, including by Anthropic. That is the path if you start talking and realize you are about to name a customer.
What people are asking
Should I make it public?
Default no, unless you wrote the transcript as if a journalist, an employer, and a model trainer will all read it. Anthropic’s FAQ is unusually blunt for a growth-loop research product.
Benefits they list:
- Researchers and policymakers can read full answers in participants’ languages, not a lab’s summary.
- Outsiders can check whether Anthropic’s later findings match what people said.
- A public record is “hard for AI companies … to ignore” if many participants demand the same change.
Risks they list:
- Employers, relatives, governments, and other companies can read it.
- Age, job, city, and school were not asked because they re-identify. Do not volunteer them.
- Scammers can misuse anything you paste.
- Permanent in practice: Anthropic can take down its copy; it cannot unspread a screenshot.
Anthropic also says it will not edit interviews before publishing, then adds a post-study filter for confidential content and policy-violating chats. Plan for verbatim. Do not assume a human will strip your Slack channel name.
How is this not the Palisade 22?
Different corpus. Palisade’s frominside.ai series is named lab staff, filmed, safety-skewed, many still private. Interviewer is Claude users, anonymous-ish, optional public dump, Claude-only sample.
Use Palisade when you want insider quotes about RSI and evals. Use Interviewer when you want product and lived-experience language at volume. Do not average them into “what people think about AI.”
Is a Claude-only sample useful?
Anthropic says no, it is not representative. Published interviews are a further subset of people who opt in. Questions shape answers. Those limitations are first-party. Quote them if you cite the study later.
The December round still moved an Institute agenda. A biased sample can be useful if you name the bias. It is not a substitute for a population survey.
How to talk if you go public (builder edition)
You do not need a novel. You need specific, non-identifying product sentences Anthropic cannot summarize away.
Safe and useful
- “Claude Code weekly limits made me split a repo across two days.”
- “I want labs to publish when default effort or routing changes mid-session.”
- “Cowork / computer-use should have a visible approval log I can export.”
- “I want the same evals on the subscription path that API customers get.”
Do not put in a public transcript
- Employer, customer, or student names.
- Internal model IDs, unreleased eval scores, or screenshots of private UIs (PixelLeak is what happens when agents paste those into git; humans can do the same in Interviewer).
- Health, immigration, or legal facts about you or a third party.
- Enough project detail that a coworker would recognize the sprint.
Anthropic will show examples of sensitive information during the chat and let you reread before the public toggle. Use that screen. The re-ID paper they allude to is the reason: small details compose.
If you want the interview analyzed but not on the open web, skip public release. Anthropic still uses it for Societal Impacts, model and product decisions, and aggregated write-ups.
How this sits next to Anthropic’s other “listening” products
| Artifact | What you see | Who speaks |
|---|---|---|
| Dec 2025 Interviewer | Stats + quotes | Claude users, private |
| Jul 2026 hard-questions film | Ad + Public Record | Mixed; brand |
| Sep 2026 Interviewer | Optional full transcripts | Claude / Claude Code users |
| Palisade frominside.ai | Named video | Lab employees |
| GLM-5.3 / Mythos papers | Numbers | Frontier Red Team |
The July campaign promised Anthropic would track actions it takes in response to hard questions. This study is the first time the raw Interviews could sit next to that promise. If you care whether PBC “listening” is a dashboard or a Super Bowl spot, a public transcript is the receipt.
For the policy essay layer, see Amodei on pacing the frontier and Coxon’s resignation. Interviewer will not replace those. It will give journalists a user-voice corpus to pair with them.
What Anthropic will do with a private interview
Even if you never toggle public:
- Societal Impacts analysis.
- Public findings, not your full text.
- Input to models and products.
- Optional early delete if you abort.
That is still a Claude-user research panel, with the usual “this chat may inform the company” bargain. If that is unacceptable, do not start.
A 15-minute runbook
- Confirm the account is ≥ 14 days and you are on Claude or Claude Code (Free/Pro/Max).
- Open the participate page from a machine that is not showing a confidential IDE.
- Answer in outcomes, not org charts.
- On the review screen, search your own text for city, school,
@company, ticket IDs. - Default private. Go public only if the transcript would still be fine on a conference slide with your face off it.
- Treat public as permanent. Calendar a reminder for whenever Anthropic posts the corpus, then read whether your sentence survived the “responsible to publish” filter.
Honest limitations (theirs and ours)
- We did not sit the September script; Anthropic hides the exact questions on purpose.
- Eligibility and FAQ text are as of the September 29, 2026 research post. Product gates can change inside the week.
- “Researchers have shown” re-ID is Anthropic’s claim about prior Interviewer data. This post does not reproduce that attack.
- Public interviews appear after analysis, not as a live feed.
- Claude-user + opt-in public is a double filter. Do not cite it as “the public wants X.”
Related on explainx.ai
- Claude “hard questions” ad and the 81,000 Interviewer round
- Palisade’s 22 lab interviews
- GLM-5.3 open-weight cyber eval
- Amodei: pace the frontier
- Jacob Coxon resigns
- Anthropic threat intelligence, September 2026
- Claude values across models and languages
- PixelLeak: agent screenshots in git
- Primary: What do you want from AI?
Eligibility, FAQ, and study dates are taken from Anthropic’s September 29, 2026 Societal Impacts post. Confirm the participate page before you start; the window closes October 6, 2026.