Ensemble
Equipping decisionmakers with insights from the people building AI.
Policymakers are writing rules for frontier AI with little direct view into the labs. Ensemble is a verified, anonymous panel of people who work at frontier AI labs. We collect questions from legislators' offices, agencies and AI safety nonprofits, put them to the panel, and share the findings with the public.
Work at a frontier AI lab? Share your judgment and experience with the people writing the future rules for AI.
Work on AI policy? Great policy needs great insight. Tell us which questions would help your work.
Know people at a frontier lab? Become a recruiter.
Getting the rules right
Rules for frontier AI are being written now. What do the people who build these systems think policymakers should know? What would they put first? Which practices would they like every lab to adopt? And what would actually work?
Policymakers hear mostly from the labs' government-affairs teams. Those teams will take the meeting, but their job is to advance their company's position, and the race between labs complicates their incentives.
The people who know how things actually work, the staff who build these systems, have no low-risk way to weigh in. Their options are to stay silent, or to quit and go public.
Ensemble gives them another way. We put policymakers' questions to the people who do the work, and publish what they say.
How it works
- Ask. We work with legislators' offices, agencies and AI safety nonprofits to find the questions they most need answered.
- Triage. We pick the questions where an inside view would matter most, and reword them so answering never requires sharing confidential information.
- Answer. Panelists get each question on Signal and answer anonymously through an encrypted CryptPad form.
- Combine. We write a short brief: the question, how many people answered, how many labs they came from, where views split, and caveats.
- Share. By default, we publish the brief for everyone.
Join the panel
The panel is a low-effort way to help AI go well. Answers are anonymous, and you answer only the questions you choose.
Anyone who currently works at a frontier AI lab can join. Message us on Signal, with your code if a recruiter gave you one. No code? We'll arrange a short conversation to confirm where you work, and keep no notes.
- We don't need your name, and Signal doesn't show us your phone number.
- Before we add you, we'll ask you to read and accept the limits.
We're not taking new members right now.
For policymakers
Our briefs bring you the views of the people building these systems, across frontier labs. These reports are public by default, so you can cite them. Each shows how many people answered and where their views split.
We're not taking questions right now. We never say which office a question came from.
Get involved
- Become a recruiter. Know people at a frontier lab? We'll give you a few single-use codes to share with them, in person or over Signal. Nobody records which code went to whom.
- Fund or advise the pilot. We're looking for funders and advisers as the panel gets started.
- Build it with us. Want to work on this with us? Get in touch.
Message Edward on LinkedIn.
Limits
- This isn't a whistleblowing channel. We ask for judgment and experience, never trade secrets or anything under NDA. If you're worried about wrongdoing at a lab, get independent advice from the AI Whistleblower Initiative's Third Opinion.
- Sharing opinions here gives you no legal protection.
Who we are
This project started as a submission to BlueDot's hackathon on AI tools for better decisions and coordination (San Francisco, September 2026), and is currently led by Edward Paulino.
Not yet funded.
Questions
How do you know members really work at a frontier lab?
Most are invited by recruiters we trust, who give codes only to people they know. Everyone else has a short conversation with us to confirm where they work. Codes work once, come in small batches we can cancel, and expire. Every brief says how members were verified.
What precautions should I take?
If you're unsure whether something is confidential, leave it out. Use a personal phone, and open our links on a personal device on your home network, not a work laptop or office Wi-Fi. We contact you only on Signal, never through work email or chat.
What will you ask?
Questions where an inside view matters most. For example, whether a lab could meet a 72-hour deadline to report a serious internal incident, or which proposed reporting requirements would be cheap or costly to comply with.
Who sees my answers?
We see the answers, but not who gave them: every member gets the same link at the same time, the form doesn't ask who you are, and we don't see when you read our messages. Recruiters never see answers. We report nothing until at least 10 people have answered, and never show a smaller group. We paraphrase written answers instead of quoting them.
What data do you keep?
- If you're a member: a Signal chat with us, under whatever name you chose, on one phone, with disappearing messages. Nothing else: no contacts, notes or spreadsheets. We delete it when you leave.
- Answers: nothing that identifies you.
- If you're a policymaker who emails us: we never attribute what you tell us.
How much time does it take?
It depends, but generally a few minutes a question, and only for the questions you choose. Joining without a code adds one short conversation. We message you only to send a question.
Can I leave?
Yes. Tell us on Signal, and we delete our chat with you. There's no account to close. The panel is for current lab staff, so if you leave your lab, please let us know.
Why Signal?
Signal keeps almost no records: when subpoenaed, it has only been able to say when an account was created and when it last connected. Only you and we can read our messages, and they disappear after a while. You can join under a nickname, and a username doesn't reveal your phone number. To be sure, set "Who can see my number" to Nobody in Signal's privacy settings.
Why should I trust you?
Most of it doesn't rest on trusting us:
- Answers go through a CryptPad form, which its host can't read.
- This page runs no scripts and sets no cookies. Its code is public at github.com/edpaulino/ensemble, and the page is served from it unchanged.
What does rest on trust: our Signal chats with members are the member list, and we keep nothing else. Two companies see visitors' internet addresses, but not what you write: CryptPad, which hosts the forms, and GitHub, which hosts this page.