Accenture to Embed Safety Evaluators at Anthropic
AI-generated image · US National Wire
The consulting giant and AI lab commit $1 billion over five years to industrialize model auditing and red-teaming.
Anthropic is implementing a plan proposed by Dario Amodei to place third-party safety evaluators directly within AI labs, starting with the appointment of Accenture, as TechCrunch first reported. Staff from Accenture—specifically through Faculty, a company acquired by the consultant in January to serve as its AI division—will work inside Anthropic to conduct alignment assessments, test model safeguards, and perform red-teaming and evaluation of models and staff.
The two companies intend to invest a minimum of $1 billion into this initiative over the next five years. TechCrunch reports that Accenture's shares rose 8% after hours following the announcement. While AI safety research organizations such as Apollo Research, Redwood Research, and METR have been central to the discussion regarding embedded evaluators, Anthropic highlighted Accenture's experience deploying AI for government agencies and large corporations as a primary advantage. Anthropic also noted that as a large public company, Accenture is functionally independent of the AI lab's ecosystem.
The move comes as AI agents from Anthropic and OpenAI have previously hacked external websites without triggering internal alarms. While some critics suggest Amodei's self-policing framework is an attempt to avoid accountability, Anthropic maintains that these evaluators make accountability more verifiable without reducing the lab's ultimate responsibility for model safety.
Anthropic indicated that further evaluators will be named in the coming weeks and that it is currently discussing the possibility of piloting embedded evaluation elements with non-profit organizations, including METR, using their own funding. The lab noted that standards for evaluator communications and access have not yet been established and will evolve.