Anthropic’s Dario Amodei Calls for Slower AI Progress and Outside Safety Reviews

Anthropic CEO Dario Amodei is urging frontier AI developers to slow capability gains while making one immediate concession of his own: permanent access for independent safety evaluators inside Anthropic. His proposal puts a practical question at the center of the AI race: can a company’s voluntary openness become a basis for collective restraint among competitors and governments?

In an essay posted to his website, Amodei wrote that the pace of improvement in AI models should slow, even as progress remains fast, so developers can use the time to improve safety and judgment. He warned that misuse or loss of control could bring cyberattacks, bioterrorism and serious economic disruption.

One company can open its doors; a race needs shared rules

Anthropic says its independent evaluators will have employee-level access to examine safety practices and report incidents. They will receive the same access as the company’s internal risk-assessment teams and may publish findings without Anthropic’s editorial control. Amodei said the arrangement takes effect immediately.

We must slow the pace at which we improve the capabilities of AI models. Progress will still seem fast, and we must make wise use of the time we gain.

Dario Amodei, in his essay

That is the part of the agenda under Anthropic’s direct control. The rest is a request for alignment beyond the company: Amodei proposed permanent independent evaluation across frontier AI companies, common safety standards among democratic countries, and coordination between democratic governments and authoritarian states on shared interests, including a ban on using AI to develop biological weapons.

The rationale for a slower pace

Amodei’s case is not that AI progress should stop. It is that increasingly capable systems may help create their successors, accelerating development faster than people can understand or control it. He described unchecked recursive self-improvement as a dynamic that must be pursued very carefully, if at all.

The proposal has three connected parts

• Make independent evaluation a permanent feature of frontier AI companies.
• Set common safety standards among democratic countries to limit unchecked progress.
• Seek international agreements on risks with mutual stakes, including AI-enabled biological weapons development.

A commitment at Anthropic, a proposal beyond it

Anthropic has committed to a standing review mechanism with employee-level access and publication rights for independent evaluators. The wider objective—slowing capability gains while preventing dangerous uses—depends on whether other companies and governments accept Amodei’s proposals for shared evaluation, common standards and international coordination.