Amodei Calls on AI Companies to Coordinate on Safety

In a 3,800-word essay Saturday morning, Dario Amodei said Anthropic would give “employee-like access” to third party evaluators who would assess and report how safely the company is developing its AI models.

The commitment was part of a three-part plan Amodei put forward for leading AI companies to “pace the frontier”—or slow down the development of bleeding edge AI— just days after his own employees spoke about the growing dangers of AI development.

The calls come as the AI industry struggles to understand and test its latest models, which have grown larger and more intelligent. Amodei said new AI models “are more capable of deceiving tests, and thus may appear more aligned while having serious problems that go undetected.”

Amodei also asked the U.S. government to give the go-ahead for AI companies to “work together to set standards” for AI safety. (Right now, some AI leaders fear antitrust reprisals for such coordination.) He said that third party evaluators (like the nonprofit Metr) would be able to verify that AI companies are following the safety practices they claim.

He wrote that “the most effective method of pacing is via regulation that targets all U.S. frontier AI companies, as that covers even those who are unwilling to cooperate voluntarily.” He added “Unfortunately, passing laws can take time, and AI is advancing very quickly.”