Anthropic wants to put the brakes on frontier AI
Dario Amodei has proposed a three-part plan to “pace the frontier”, arguing that increasingly capable AI systems need stronger safety checks before development accelerates further.
Anthropic chief executive Dario Amodei has called for a slower, more deliberate approach to frontier AI development. TechCrunch reports that Amodei’s proposal follows concerns about the speed of progress and recent incidents involving AI systems operating in ways their makers did not intend.1
The plan is framed around three broad strategies for pacing the frontier. Anthropic says it is unilaterally committing to one of them, while OpenAI chief executive Sam Altman has indicated that OpenAI will follow suit.1 The reporting does not turn that agreement into a binding industry-wide pause, and no such pause has been announced.
The Verge reports that Amodei also wants third-party evaluators such as METR to receive access to models so they can assess adherence to safety practices and stated commitments.2 That matters because internal testing can miss failure modes that appear when a system is given broader tools, longer tasks or less predictable environments.
Amodei’s intervention comes amid a wider debate over whether advanced models are becoming capable faster than safety work can keep up. The proposal is therefore less a product announcement than an argument about governance: developers should accept more scrutiny and build in more time for evaluation before capability races leave the lab.
Our opinion
A slowdown is not a magic safety switch, but independent testing and clearer commitments are a much better idea than asking the public to trust impressive demos. Anthropic is right to put a concrete shape around “go slower”; the real test is whether the company accepts uncomfortable results when evaluators find them.