Anthropic CEO Proposes Embedded AI Evaluators, Citing $200 Million Cost of Safety Stances
Updated
Updated · Fox News · Sep 19
Anthropic CEO Proposes Embedded AI Evaluators, Citing $200 Million Cost of Safety Stances
3 articles · Updated · Fox News · Sep 19
Summary
Dario Amodei urged frontier AI companies to grant employee-like access to independent “embedded evaluators” who would verify safety commitments, assess training pipelines and report incidents.
Amodei pointed to outside groups such as METR as a model, arguing the public now relies too heavily on companies deciding what safety information to disclose.
METR’s role is sensitive because its founders, including Beth Barnes and Paul Christiano, have ties to OpenAI, Anthropic model evaluations and the effective altruism network linked to Anthropic’s early backers.
Amodei’s push reflects his long-running distrust of self-policing in AI after leaving OpenAI in 2020 and Anthropic’s willingness to forgo work it viewed as unsafe, including a $200 million contract.