Anthropic commits to embedded external evaluators as Amodei calls for slower AI gains

Anthropic CEO Dario Amodei called for a slower pace of improvement in frontier AI capabilities and committed Anthropic to bringing in independent evaluators with access broadly comparable to its internal risk teams. That commitment is the immediate operational piece of a wider three-part proposal; the remaining steps depend on coordination among companies and governments.

The company said it intends “in the near future” to give an external review team desks, access badges, company laptops, and access to tools and permissions mostly comparable to those available to employees who assess risk. Legal, contractual, security and third-party privacy restrictions would still apply. Anthropic did not name the organization it plans to invite or provide a launch date.

The reviewers would assess models, training pipelines and company processes, verify compliance with safety practices and commitments, and report incidents. Amodei said their contract should let them publish key findings about risks, incidents, Anthropic’s practices and the access they received without editorial control from the company. Anthropic would retain narrow rights to redact sensitive material. Reviewers could disclose when a redaction materially changed their conclusions.

The proposal calls for continuing access inside Anthropic, a different arrangement from a discrete outside assessment such as a UK AI Security Institute test of Anthropic and OpenAI agents.

Amodei described the evaluator arrangement as a unilateral Anthropic commitment. He separated it from the proposal’s second step: coordination among frontier AI developers based in democratic countries to pace capability gains. He said that coordination could require government support or a narrow antitrust waiver.

The third step would seek global coordination among governments, including engagement with authoritarian governments. Neither coordinated step was presented as an agreement already reached. Amodei also said pacing would not mean stopping model training or technical progress.

Multiple news organizations reported that OpenAI CEO Sam Altman backed the pacing proposal and said OpenAI would also adopt independent evaluators with employee-like access. They also reported that Elon Musk said Amodei was right. Those reactions do not establish implementation at OpenAI or any Musk company, or an industry-wide agreement on the broader coordination proposals.