Anthropic's first embedded evaluator turns out to be Accenture
Anthropic announced that Accenture's AI division, Faculty, will begin working inside the company to assess model safety. Both firms plan to invest at least $1 billion in the effort over five years.
Anthropic CEO Dario Amodei's proposal to embed independent safety evaluators inside AI labs is moving forward, with consulting giant Accenture named as the first partner. In a blog post, Anthropic said Faculty — the AI unit Accenture acquired in January — will begin evaluating and red-teaming its models, conducting alignment assessments, and testing safety safeguards from within the company.
Anthropic and Accenture expect to invest at least $1 billion in the project over the next five years. The choice of Accenture caught many AI observers off guard, and markets responded quickly too: Accenture's shares rose 8% in after-hours trading following the announcement.
Why Accenture
Until now, discussion of embedded evaluators had centered on AI safety research organizations such as METR, Redwood Research, and Apollo Research. While Accenture isn't known for cutting-edge deep learning research, Anthropic pointed to its practical experience deploying AI for large corporations and government agencies as a key strength. Anthropic also noted that Accenture, as a large public company that predates the AI boom, is more functionally independent from Anthropic and the wider ecosystem surrounding the lab.
Anthropic said no standards currently exist governing evaluators' access or communications, and that its approach is expected to evolve over time. The company added that additional evaluators will be announced in the coming weeks, and that it is in talks with METR and other nonprofit organizations about piloting elements of embedded evaluation using their own funding.
The move comes after incidents in which AI agents deployed by OpenAI and Anthropic hacked into outside websites without triggering internal alarms. Some critics have characterized this kind of self-policing arrangement as a way for AI labs to sidestep accountability for their models' misbehavior. Anthropic maintains that the evaluators do not reduce its accountability but make it more verifiable, stressing that responsibility for its models' safety remains its own.
