Anthropic has handed the job of independently checking its models to an outside firm, and the arrangement comes with a price tag measured in the billions. The company said it has appointed Accenture as an independent evaluator of its model safety, with the work led by Faculty, the British AI company Accenture bought in January for about $1 billion.
The two companies have each committed to spend at least $1 billion over five years, a combined $2 billion that will fund teams embedded inside Anthropic’s research process with roughly the access of full-time employees. Their brief is to run red-team attacks, test the models’ alignment, and inspect the guardrails meant to keep the systems behaving, then to turn that work into a third-party verification that customers in regulated industries can point to.
Faculty’s chief executive, Marc Warner, described the philosophy behind the deal in a sentence: AI should be safe by design rather than safe by accident. The line captures the shift Anthropic is trying to make, from safety work done in-house and judged by the company itself to safety work performed and vouched for by someone on the outside.
The timing is not accidental. The announcement came a day after Dario Amodei, Anthropic’s chief executive, called publicly for giving third-party evaluators greater access to the systems they are meant to assess. Until now, most independent safety reviews have relied on limited access that critics say makes it hard to find real problems, since the evaluators cannot poke at the models the way a determined attacker or a careless deployment would.
Anthropic has built its brand on the idea that safety is not an afterthought, and the Accenture deal is an attempt to put money behind that claim. The company has spent years publishing research on how to align powerful models with human intent and on how to test for the harms those models might cause, and it has positioned itself as the cautious counterweight to rivals moving faster.
The commercial logic runs alongside the safety logic. Anthropic sells its models to banks, insurers, health systems, and government agencies, customers that face their own regulators and cannot simply take a vendor’s word that a model is safe. An independent seal of approval, backed by a firm of Accenture’s size, gives those buyers something to show an auditor and gives Anthropic a way to differentiate itself in a market where the leading models are increasingly hard to tell apart.
Faculty brings a particular history to the role. The London firm began life more than a decade ago advising governments and intelligence agencies on data science, and it spent years working on safety-sensitive problems before Accenture acquired it. That government pedigree is part of what made it attractive for a job that sits at the intersection of technical competence and public trust.
Whether the arrangement meaningfully changes anything will depend on the details, many of which have not been disclosed. The commitment to embed evaluators inside the research process suggests more than the paper reviews that have typified past efforts, but embedding raises its own questions about how independent an evaluator can remain when it sits at the same table.
Amodei’s larger argument has been that safety testing should become a standard part of how the industry operates, not a gesture a company performs before launch. The Accenture deal is one concrete step in that direction, and its size is meant to signal that the work is being treated as a serious, recurring cost rather than a one-time exercise.
For Accenture, the partnership is a bet that AI safety will become a durable line of business as the technology spreads into regulated industries. The consulting giant has been assembling an AI practice in recent years, and the Faculty acquisition gave it a team with deeper technical credentials than a traditional advisory shop would carry. Winning the Anthropic engagement places that team at the front of a field many expect to grow quickly.
Anthropic has organized its safety work around a set of internal policies that call for more intensive testing as models grow more capable, and the Accenture engagement plugs into that framework as an external check. The company has long argued that the stronger a model becomes, the more it needs scrutiny from people who do not work for the people who built it. The five-year commitment is meant to make that scrutiny a standing feature of the process rather than a response to a crisis.
Neither company said how the results of the testing would be published or how disputes over findings would be resolved, and those are the points where such arrangements tend to break down. What the deal establishes for now is a commitment, unusual in its scale, to let outsiders look hard at models before they reach the public, and to keep looking after they do.
