Anthropic is giving an outside evaluator unusually deep access to its model-development process. The company announced on September 18 that it is partnering with Accenture on what it calls "embedded evaluation," with Accenture's specialist AI business Faculty leading work that includes model evaluation, red-teaming, alignment assessments and safeguard testing.
The unusual part is where the evaluators sit. Anthropic says embedded evaluators will work inside AI companies with access comparable to an employee's, allowing them to observe models during training, follow deployment decisions and speak directly with staff. Anthropic presents that access as a way to make safety commitments more verifiable, but it also acknowledges that the operating standards for this model are not settled.
The investment is large, but the rules are still being written
Anthropic and Accenture each say they expect to invest at least $1 billion over five years in building capacity around this work. That is a significant commitment, but it should not be mistaken for a mature regulatory framework.
Anthropic explicitly says there are not yet agreed standards for what embedded evaluators should be allowed to see or how they should report findings. There is also no settled funding system for independent evaluation. In this first arrangement, Anthropic will directly fund Accenture's evaluation work, while arguing that pooled or government funding would be preferable in the long term.
That funding relationship is an obvious part of the independence question. Anthropic says the arrangement is non-exclusive, that Accenture can work with other AI developers, and that Anthropic expects to bring in additional evaluators. It also says it is talking with nonprofit evaluators including METR about pilots using their own funding.
Why access has become the pressure point
Independent AI evaluation is only useful if evaluators can inspect enough of a system to test meaningful failure modes. Public model access can reveal plenty, but it does not automatically expose training decisions, internal safeguards or the conditions under which a model changed during development.
Recent reporting from The Verge has described a wider push among AI-safety researchers for stronger third-party access and more systematic evaluation. The details are contested, and different labs disagree about how quickly development should move, but access is a recurring problem: an evaluator cannot verify much about internal practices if it only sees the final public product.
Anthropic's proposal tries to move that boundary inward. In theory, an embedded evaluator can test earlier, challenge assumptions while a model is still being developed and document whether a lab follows its stated procedures. In practice, the credibility of that system will depend on rules that do not yet exist: who decides what can be published, how conflicts are handled, how evaluators are funded and what happens when they disagree with the company hosting them.
So the announcement is best read as an experiment in oversight, not proof that the oversight problem is solved. The partnership creates money, access and a defined evaluation role. The harder work is turning those ingredients into standards that outsiders can trust.
The investment figure also needs careful framing. "At least $1 billion each" describes what the two companies expect to invest in building capacity over five years, not a single payment from Anthropic to Accenture and not a guarantee that all of that money will fund one evaluation team. The scope is broader than the initial engagement.
Reporting notes