Anthropic Taps Accenture as Embedded Third-Party Evaluator of Its Frontier AI Models

The two companies said each will invest at least $1 billion in the joint effort over the next five years, according to Engadget. Accenture's evaluators will work inside Anthropic rather than reviewing the company's models from the outside.

Anthropic described the arrangement as an early step in the three-part approach Amodei proposed last week. Under that proposal, frontier AI companies would give embedded third-party evaluators "ongoing, employee-like access," with the evaluators responsible for confirming that safety practices and commitments are followed, reporting incidents and assessing how well models are aligned.

Anthropic said Accenture's evaluators will "watch models take shape in training, follow the decisions that govern how those models are built and deployed and speak directly to employees." The company pointed to Accenture's work helping businesses and governments deploy AI across several industries, and said that experience gives the consulting firm a wider view of how enterprises use the technology when it assesses Anthropic's own models.

Anthropic also acknowledged that the concept of embedded evaluators is new and that standards have not been set for how far their remit extends, what access they receive or how they report concerns.

Accenture is not the only evaluator Anthropic intends to use. The company said it will work with additional firms on evaluating its frontier models and will announce other evaluators in the coming weeks. It added that Accenture will also work with AI companies other than Anthropic.