Dario Amodei calls for slower AI progress and shared safety checks

Anthropic chief executive Dario Amodei is calling for a slower pace of AI model development, with outside evaluators given deep access and shared safety standards. His proposed checks make the debate more concrete than another vague promise to “be responsible”.

Watch Desk analysis

What happened

In an essay titled “We Must Pace the Frontier”, Amodei argues that the industry should slow the rate at which AI models improve to reduce the risk of serious harm. International Business Times Australia reports that he proposed three steps: external evaluators with “employee-like access”, common safety standards and international coordination on the pace of progress.

The article sets that call against a competitive industry where major labs continue shipping products on different schedules. It does not establish that the companies have agreed to adopt Amodei’s proposals.

Why it matters

“Independent evaluation” can mean anything from a brief outside test to sustained access to a system and the information needed to examine it. Amodei’s employee-like-access proposal points towards the latter, while common standards could make it easier to compare how labs assess their models.

International coordination is the hardest part: a shared approach would need to survive commercial pressure and disagreement over who sets the rules. The essay makes a case for slowing capability gains, but the report offers no evidence that the industry has acted on that request.

Our read

This is a substantial proposal, not proof of a new safety pact. Evaluators need enough access to do meaningful work, and standards need to be specific enough to test. Otherwise “independent oversight” risks becoming a reassuring label on a process nobody outside the lab can properly inspect.

For now, the useful thing to watch is whether Amodei’s three proposals turn into named commitments, with clear access rules and a way to check whether they are followed.

What to watch

• Whether Anthropic or other labs publish concrete terms for external evaluators’ access.
• Whether any common safety standards are proposed, and who would assess compliance.
• Whether international discussions produce commitments on the pace of AI development.

Discussion spark: Would employee-like access for independent AI evaluators be enough to build trust, or should the labs also have to publish common test results?

Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.