Anthropic pays Accenture to embed evaluators in its model development

Anthropic said it would embed a team from Accenture, working through Accenture’s Faculty division, inside its own organisation to evaluate and red-team models, run alignment assessments and test safeguards. Anthropic and Accenture each said they expected to invest at least $1 billion over five years in the arrangement.

The distinguishing feature was access. Anthropic said embedded evaluators would get “access comparable to an employee’s” — able to observe models as they are trained, follow the internal decisions that govern how they are built and deployed, and speak directly with staff, rather than testing a model only after release through an API. The company framed this as the first concrete implementation of a proposal it had made days earlier: chief executive Dario Amodei’s essay arguing frontier labs should slow capability gains until independent oversight could keep pace had called for exactly this kind of embedded, continuous evaluation.

Anthropic was explicit that it saw the funding structure as a stopgap. It said long-term funding for this kind of oversight should ultimately come from pooled industry or government sources rather than from individual companies paying their own evaluators — a structure that does not yet exist — and that in the meantime it would fund Accenture’s work directly. The company also said it was in discussions with nonprofit evaluator METR and other independent groups to pilot parts of the embedded-evaluation model using their own, separately sourced funding.

The partnership is non-exclusive on both sides: Anthropic said it planned to bring on additional evaluators in the coming weeks, and Accenture said it would offer similar embedded-evaluation services to other AI developers. Critics of company-funded safety evaluation have long pointed to the conflict of interest in a lab paying the people who assess it; Anthropic’s own framing of the deal as an interim measure pending pooled or public funding was, in effect, a concession to that criticism rather than a rebuttal of it.