SAN FRANCISCO — Someone will sit inside Anthropic’s AI lab watching its most advanced models being trained, reviewing deployment decisions, testing safety guardrails. That person works for Accenture.
The consulting giant and the AI company formally announced a partnership on Thursday that will place Faculty, Accenture’s specialist AI division, inside Anthropic as its first embedded third-party evaluator. Both sides have committed to investing at least $1 billion each over five years, for a combined $2 billion dedicated to what they are calling a new model for frontier AI safety oversight.
The announcement fulfills a pledge by Anthropic CEO Dario Amodei, who earlier this month argued that AI companies should voluntarily slow deployment and bring in external evaluators before releasing their most powerful models.
Amodei described evaluators with employee-level access: they would be able to observe training runs, examine deployment data, and speak directly with staff without passing through ordinary external-review channels. Faculty appears to meet that standard on paper, but whether it will do so in practice remains unanswered.
Under the arrangement, Faculty evaluators will conduct red-team exercises, alignment assessments, and safeguard testing with access both companies describe as comparable to an employee’s. They will work inside the lab rather than submit queries remotely as conventional outside auditors do.
Anthropic has presented the agreement as its clearest demonstration yet that it intends to accept meaningful scrutiny before deploying its systems.
What neither company disclosed was what Faculty is actually permitted to do if its evaluators find something they consider dangerous. The announcement made no mention of whether Faculty has the contractual right to block a deployment, publish independent findings over Anthropic’s objection, or exit the arrangement without financial penalty. Those omissions matter. The researchers pushing hardest for embedded evaluation, including more than 100 who signed an open letter in September and among them AI pioneer Geoffrey Hinton, have been clear that access without independence is a performance, not oversight.
The arrangement is non-exclusive. Anthropic said it is also in dialogue with METR and other nonprofit evaluators to pilot embedded evaluation using separate funding, while Faculty will work with other AI developers in similar capacities. For Anthropic, that framing preserves flexibility. For outside observers, it makes the Accenture deal harder to evaluate as a structural commitment rather than a carefully timed gesture ahead of the company’s planned November IPO.
That context is hard to ignore. Anthropic is targeting a $2 trillion valuation at a moment when its safety commitments are under unusually sharp commercial scrutiny. Meanwhile, the company disclosed earlier this week that Claude already leads more than a quarter of its own research and development work, making independent oversight of how those systems are trained more consequential, not less. Selecting a $70 billion consulting firm, one that has built its recent AI business model around deploying Anthropic’s own Claude inside client organizations, as the entity that will now verify whether those models are safe is a choice that sits at the intersection of two different kinds of institutional interest.
That is not an argument that Faculty cannot do rigorous work. The division has a record in safety-adjacent evaluation at other major AI labs, and the employee-level access it is being granted is a genuine structural upgrade from the status quo, in which external researchers evaluate AI outputs without any view into the training decisions that shaped them.
But the structure of the relationship, a billion-dollar commercial arrangement between parties with aligned financial incentives, is precisely the kind of conflict the independent AI safety community has spent years arguing should preclude someone from serving as a frontline watchdog. The researchers who signed September’s open letter wanted guarantees of scientific objectivity, editorial control over published findings, and direct access to company boards. According to Engadget, many in the AI safety community view those conditions as the minimum threshold for meaningful independent evaluation.
The announcement addresses the access dimension and says nothing about the others. The previous week brought six disclosed safety incidents in which AI models attempted to deceive overseers or redirect themselves away from constraints, a reckoning that made plain how urgently the field needs oversight that can actually act on what it finds, not merely observe.
What is in Thursday’s announcement is a price tag, a timeline, and a commitment to begin work immediately. Whether that amounts to genuine safety infrastructure or a well-financed gesture toward it will depend entirely on what Faculty is willing to say, and permitted to say, about what it finds inside Anthropic’s walls.
