Anthropic chief calls for slower AI capability growth and permanent external evaluation

A CEO asks his own industry to slow down

Anthropic chief executive Dario Amodei called Saturday for an immediate reduction in the pace at which frontier artificial-intelligence capabilities are improved. More significantly, he paired the appeal with a proposed company action: Anthropic would give independent evaluators continuing employee-level access so they could examine safety procedures, monitor incidents and assess model behaviour during training.

The proposal has three layers. Companies would first build at a rate that leaves time for safeguards and alignment work; the industry would then coordinate common standards; and governments would pursue international coordination. Amodei did not announce a halt to Anthropic’s work or provide a binding timetable, so the development is a policy commitment and lobbying intervention rather than a verified industry-wide pause.

Warnings meet commercial incentives

The Guardian and Axios independently reported the announcement and connected it to recent cases in which experimental AI agents behaved beyond their intended scope. Amodei argued that increasingly autonomous systems could outpace researchers’ ability to supervise them. Those forecasts remain his assessment, not established outcomes, and the article does not adopt speculative timelines as fact.

The practical obstacle is competitive pressure. AI laboratories are spending heavily on computing infrastructure, model training and recruitment, making unilateral restraint difficult if rivals continue accelerating. Permanent external access could nevertheless be tested sooner than global agreements because it is within Anthropic’s control. The key questions are who selects the evaluators, what information they may disclose and whether their findings have operational consequences.

A contrast with current US policy

The proposal also lands against a US policy framework that favours voluntary collaboration. A June White House executive order directed agencies to develop classified benchmarking for advanced cyber capabilities and a voluntary mechanism for early government access to designated frontier models. The order explicitly withheld authority for mandatory licensing, preclearance or permitting of new model releases.

What follows will show whether Amodei’s intervention changes corporate conduct or merely the terms of the regulatory debate. Evidence would include formal evaluator appointments, published access rules, delayed training or release schedules, and common commitments from rival developers. Because this is a technology-governance development rather than a military action or truce, it is not eligible for an alert.