Anthropic has committed to embedded third-party evaluators, while OpenAI says it will follow the same strategy as the companies discuss AI safety and limits on frontier development.
Anthropic and OpenAI commit to embedded evaluators
Anthropic has committed to embedded third-party evaluators with employee-like access, and OpenAI says it will adopt the same arrangement. Anthropic’s commitment is part of a broader strategy to pace frontier AI development.
What the evaluator arrangement would involve
Dario Amodei proposed using embedded evaluators from third-party organizations such as METR as a first step. The evaluators would receive company badges, desks and laptops, allowing them to work with employee-like access and verify whether the companies are following their pacing and safety commitments. Amodei has also called for leading AI companies in democratic countries to coordinate on common safety standards and limits on unchecked AI progress.
The commitments come amid wider safety talks
OpenAI, Anthropic and Google DeepMind have been in talks about AI safety for several weeks. OpenAI supports a provision in the FRONTIER Act that would require top frontier labs to allow independent verification organizations into their companies. Amodei’s essay also proposed a narrow government waiver allowing safety coordination.
A broader dispute over how to pace AI
The stated goal of the “We Must Pace the Frontier” proposal is to slow the pace of AI progress, and it asks governments to require other frontier companies to match its approach. The proposed law would require AI models offered to the public to be released as open weights, while exempting internal and research models.
Antitrust questions remain
The talks could put the companies at risk of violating antitrust law if the coordination is found to suppress competition.