Anthropic said on Sept. 18 that it will work with Accenture to place external evaluators inside the company to independently assess frontier AI models. In its official announcement, Anthropic said each side expects to invest at least $1 billion over the next five years to build that capability.
Embedded reviewers will get access comparable to employees
The clearest detail in the announcement is the scope of access. Anthropic said the embedded evaluators will work inside the company and have access "comparable to that of an employee."
The company said their responsibilities will include:
- evaluating models and conducting red-team testing
- running alignment evaluations
- testing model safeguards
- observing models during training
- tracking company decisions on model development and deployment
- speaking directly with employees
- assessing company operations and auditing safety commitments
- identifying blind spots and reporting incidents
Anthropic said the work will be carried out by Faculty, Accenture’s specialist AI business.
Anthropic links the move to Amodei’s earlier proposal
The announcement ties the partnership to CEO Amodei’s essay, We Must Pace the Frontier, which argued for placing evaluators inside Anthropic.
That means an earlier public commitment now has a named implementation partner and a stated funding figure. The $1 billion figure refers to the expected investment from each side over a five-year period.
ABMedia said this is the first move on that track to come with a concrete number, after earlier reporting that Anthropic and OpenAI had made an unusual commitment to slow the race.
Company says outside evaluation does not shift responsibility
The announcement also includes a line that Anthropic highlighted directly: "Independent embedded evaluators do not reduce our accountability; they make it more verifiable. The safety of our models remains our responsibility."
The statement addresses an obvious question raised by the arrangement: whether outsourcing part of the evaluation process also shifts responsibility outward. Anthropic’s answer in the announcement is no, and it states that responsibility for model safety remains with the company.
Launch timing and follow-up process remain unclear
The announcement does not give a start date. Anthropic also said the arrangement is not exclusive. The company said it will announce other evaluators in the coming weeks and is in discussions with nonprofit evaluation groups including METR.
No senior executive was quoted by name in the announcement. Anthropic also did not explain how issues found by evaluators will be handled, whether reports will be released publicly, or how "auditing safety commitments" will work in practice.