Dario Amodei, CEO of Anthropic, has announced a unilateral decision to slow the pace of AI model training and development. He stated that this measure allows third-party evaluators, such as METR, to access their systems and verify adherence to safety commitments. In a detailed essay, the executive outlined a three-step plan to pace the frontier, a phrase meaning companies must halt rapid advancement to let regulators assess risks and build necessary safeguards. This approach marks a shift from the previous industry standard of prioritising speed above all else, as Amodei argues that unchecked growth poses existential threats to democracy and security.
The move matters because it forces a pause in a sector where capabilities often outstrip human oversight. By granting external auditors direct model access, Anthropic attempts to create a transparent verification process that private testing alone cannot guarantee. This strategy also signals a potential industry-wide realignment, where safety protocols become a prerequisite for releasing new tools rather than an afterthought. The proposal challenges the current race to deploy increasingly powerful systems without clear governance structures in place.
- External auditors will receive direct access to Anthropic’s models for safety verification.
- The plan requires the wider industry to cooperate on establishing shared safety standards.
- Regulators need time to evaluate risks before new models reach public use.
