Anthropic Letter

Anthropic employees are asking leadership to commit to a conditional pause.

Our Ask

We ask that Anthropic make a public commitment to pause frontier model development1 if China and the other major labs credibly agree to do the same.

1 By "pause frontier model development" we mean halting the training of the most advanced, large-scale, or general models that are pushing the frontier of general capabilities. These teams would move to AI alignment research or narrow applications. The specific threshold could be agreed upon between labs, as in Anthropic's own RSP v3, where one benchmark is AI systems able to "compress two years of 2018-2024 AI progress into a single year." Current models stay available. In practice, we expect the terms and duration of any pause would be negotiated between participating labs, with the goal of establishing adequate safety standards and international oversight.

Dear Dario,

Project Glasswing showed the world what responsible deployment looks like: identifying a model's dangerous capabilities and coordinating with industry before releasing it. That kind of leadership is exactly what this moment demands.

At Davos, you said you were confident the race was something you and Demis could "work out" if only the two of you were racing. Demis said he would support a pause if all other companies and countries agreed. And OpenAI's charter, which you helped draft, already includes a commitment to stop competing if another project is closer to safe AGI.

Anthropic's latest Responsible Scaling Policy moved away from its commitment to pause, since a unilateral pause would put the company at a competitive disadvantage and create pressure to downplay capabilities. A conditional commitment solves exactly this problem.

Today, we're asking Anthropic to make the conditional pause commitment we describe above, and continue leading in AI Safety.

Signed,

Signatories

Signatories will be published once enough people have signed.

Add Your Name

Current and former employees of Anthropic are invited to sign. You may sign anonymously. All signatures are verified before being published.

Verification Method

Frequently Asked Questions

When would the pause end?

A pause doesn't need to be indefinite to be useful. Even a few months of coordinated slowdown, timed to when AI can automate significant AI research but before recursive self-improvement, would let safety research close critical gaps and labs collaborate on alignment. The longer a pause drags on, the greater the risk that a non-participating actor races ahead. That's why the conditional structure matters: all major labs pause together, collaborate on safety, and resume together once shared safety standards and oversight are in place.

How would you enforce it?

Ideally, governments would enforce a pause through an international treaty, as with nuclear nonproliferation. Research suggests nations could verify compliance through multiple independent approaches, including hardware tracking, energy monitoring, and facility inspections. But public commitments from lab CEOs are what create the conditions for that coordination. A conditional pause is also partially self-enforcing: each lab pauses only if the others do. If one lab breaks the agreement, the others resume.

How are signatures verified?

Every signature is verified before it appears on the letter. Current employees can verify by email (@anthropic.com) or by signing into a Google Form with their work Google account. Former employees and those who prefer not to use a work email can verify through alternative means such as a photo of a work badge or a referral from a co-signer. No signature is published without verification.

How is my data handled?

All personal data (email, contact info, uploaded images) is permanently deleted from our database as soon as your signature is verified. After deletion, only your public listing remains (name, role, and employment status — or just "Anonymous" if you signed anonymously). No analytics or tracking scripts are used on this site.

Can I sign anonymously?

Yes. Your signature will appear as "Anonymous, verified [current/former] employee at Anthropic." Only one organizer reviews anonymous signatures during verification. Your personal data is deleted immediately after approval.

What if I accidentally fill out the form twice?

We de-duplicate signatures manually during review. If you submitted twice, we will only count you once.

I signed anonymously but now want to put my name on it.

Sign again using the "Alternative verification" method. In the verification details, mention that you previously signed anonymously and would like to switch to a named signature. We will update your entry and make sure you are not double-counted.

What infrastructure does this site use?

This site is hosted on Fly.io, a US-based infrastructure provider. The database is SQLite, stored on an encrypted persistent volume. Verification emails are sent via Resend. Google Forms is used as one verification option because it allows email confirmation without sending anything to your inbox. The site itself is a simple open-source Flask application. No analytics or tracking scripts are used.