Anthropic named the first outside team it will embed inside its own company to check its safety work on Friday, September 18, 2026, and it is not one of the nonprofits the field expected. Per Anthropic's announcement, the work will be led by Faculty, the specialist AI business Accenture acquired earlier this year. Faculty staff will sit inside Anthropic to red-team models, run alignment assessments and test safeguards alongside Anthropic's internal teams and existing safety partners. Anthropic and Accenture "each expect to invest at least $1 billion in building capacity in this area over the next five years," per the announcement. Accenture shares rose more than 6.5 percent in after-hours trading that Friday, according to Investing.com; TechCrunch put the move at 8 percent.
The mechanism is genuinely new. No frontier lab has previously given an outside evaluator standing, employee-level access to watch models during training rather than testing a finished checkpoint under an NDA. The argument is about who got the job. Accenture is not a safety nonprofit — it is one of Anthropic's largest enterprise customers and its biggest Claude Code deployment partner. Within 48 hours Anthropic's own post announcing the deal had been Community Noted on X over the word "independent."
What Amodei promised six days earlier
The deal is the first concrete output of an essay Anthropic CEO Dario Amodei published on September 12, 2026, titled "We Must Pace the Frontier." Its core claim: "We must slow the pace at which we improve the capabilities of AI models." The essay laid out a three-part plan and said Anthropic was "unilaterally committing to this step now" — the first step, embedded third-party evaluators with permanent employee-level access.
Amodei was specific about what that access means. Evaluators would get "desks in our offices, access badges, and company laptops," with "permissions mostly comparable to what internal risk assessment teams have." They would have the right to publish key findings about risk levels, incidents and practices without Anthropic holding editorial control — though Anthropic reserved the ability to redact "security-sensitive, legally privileged, commercially sensitive, or third-party confidential information." The one evaluator Amodei named as an example was METR, the California nonprofit. Six days later, the name on the contract was Accenture's.
Amodei's September 12 essay proposed, in order: embedded third-party evaluators with employee-level access inside frontier labs; coordinated safety standards and limits on the rate of unchecked progress among labs in democratic countries, which he acknowledged needs government support to be legally workable; and eventual coordination with authoritarian governments, including the hard problem of verifying compliance. Only the first is something one company can do alone — which is why it moved first.
Faculty is not a random pick — which is both the case for it and against it
Faculty was founded in 2014 by Marc Warner, a former quantum physics research fellow at Harvard and a former member of the UK's AI Council. It built the NHS Early Warning System during COVID and has a long UK public-sector record. Accenture agreed to buy it in January 2026 — reported by Tech.eu and SiliconRepublic at over £600 million — and completed the acquisition in March, absorbing more than 400 AI staff and naming Warner as Accenture's chief technology officer. He remains Faculty's CEO.
On raw capability, that is a defensible choice. Faculty's client roster already included both Anthropic and OpenAI, checking model safety before release. There are perhaps a dozen organisations on earth with real experience evaluating frontier models, and Faculty is one of them. "Faculty was founded on the belief that AI should be safe by design, not safe by accident," Warner said in Accenture's announcement. Accenture chair and CEO Julie Sweet framed the pairing as complementary: "Safety requires both deep technical expertise and a clear understanding of how AI is used in the real world."
The problem is everything else Accenture sells. On December 9, 2025, the two companies announced a multi-year expansion that created the Accenture Anthropic Business Group, committed to training roughly 30,000 Accenture professionals on Claude, and put tens of thousands of Accenture developers on Claude Code — at the time Anthropic's largest-ever deployment of it. Accenture's own announcement described the arrangement as making it one of Anthropic's three largest enterprise customers.
The objection went public within two days
Per OfficeChai's September 20 report, Anthropic's post announcing the partnership was appended with a Community Note on X disputing the characterisation of Accenture as an independent evaluator, on the grounds that Anthropic funds the evaluator and the two companies are already business partners. It is a crowd-sourced annotation rather than a considered critique, but it is a fair summary of what a lot of people said out loud the same weekend.
TechCrunch's Tim Fernholz reported on September 18 that the choice "surprised many AI watchers," because public discussion of who would fill this role had centred on safety research organisations — METR, Redwood Research, Apollo Research — rather than a consultancy. Fernholz also noted that some critics regard the embedded-evaluator scheme itself as a plan to evade accountability: a structure that gives a lab the appearance of external scrutiny while keeping the scrutineer on the payroll and inside the building.
A hundred researchers asked for something narrower, the same day
Also on September 18, more than 100 AI researchers and evaluators published an open letter arguing that third-party safety evaluators currently lack the independence, resources and legal protections needed to credibly assess frontier models. It was organised by the AI Evaluator Forum, a coalition of groups that test AI systems for safety, and calls on foundation model companies to guarantee evaluators "scientific objectivity, transparency, independence, and robust protections against interference." Signatories include Geoffrey Hinton, David Duvenaud — formerly Anthropic's alignment evaluations lead, now an associate professor at the University of Toronto — and Mark Daley, chief AI officer at Western University, alongside people from Stanford, Johns Hopkins and METR.
The letter does not name Anthropic or Accenture, and was not written as a response to the deal. But BNN Bloomberg's September 19 report quotes Duvenaud making the point that lands squarely on it: companies should not be grading their own homework on catastrophic risks. Conrad Stosz, who chairs the AI Evaluator Forum, described the letter as an effort to enshrine shared baseline principles and hold AI companies to commitments they have already made publicly.
Anthropic's answer
Anthropic's defence, per TechCrunch, rests on two claims. The first is practical: Accenture brings real experience deploying AI inside large corporations and government agencies, which is where model failures actually show up. The second is structural: Accenture is a large public company that predates the AI revolution, so its financial fate does not hinge on Anthropic's. To the accountability objection, Anthropic's position is that embedded evaluators "do not reduce our accountability, but help to make it more verifiable."
Anthropic also stressed that the arrangement is non-exclusive and said more evaluators would be announced. It says it is in dialogue with METR and other nonprofit evaluators about piloting elements of embedded evaluation, and per reporting on the announcement, METR would participate using its own funding rather than Anthropic's. METR states it takes no money from AI companies or their executives — a position that matters given former White House AI adviser David Sacks has separately questioned METR's independence over its ties to Anthropic investors and staff. Nobody in this argument has clean hands by everyone's standard, which is roughly the problem.
The number nobody has explained
The headline figure deserves a second look. Anthropic's wording is that each company "expects to invest at least $1 billion in building capacity in this area over the next five years." That is not the same as Anthropic paying Accenture $1 billion, and neither company has published how the evaluation work itself is funded, how many evaluators will be embedded, when they start, or what reporting channel their findings travel through. Amodei's essay set out publication rights and redaction limits in some detail; Friday's joint announcement did not restate them. Anthropic's own post concedes the point in a softer form, describing embedded evaluation as an emerging practice whose standards for information access and reporting are still evolving.
That is the honest read on September 18. The access commitment is the most substantive thing any frontier lab has volunteered on external oversight — evaluators who, in Anthropic's words, can "watch models take shape in training, follow the decisions that govern how those models are built and deployed, and speak directly to employees." The governance around it is unwritten. Whether this becomes real oversight or an expensive letterhead depends on details that have not been published, and on whether the next names Anthropic adds look less like a supplier and more like the nonprofits its CEO cited when he made the promise.
Is Accenture actually independent of Anthropic?
Not financially. A December 9, 2025 partnership made Accenture one of Anthropic's three largest enterprise customers and put tens of thousands of its developers on Claude Code. Anthropic argues Accenture is functionally independent as a large public company that predates the AI boom; critics, including a Community Note on Anthropic's own announcement post, say funding plus an existing commercial relationship is not independence.
What does $1 billion each actually buy?
Unclear. Anthropic's announcement says each company expects to invest at least $1 billion in building capacity in this area over five years — it does not say Anthropic is paying Accenture that sum, and neither company has disclosed the funding structure, the number of evaluators, or a start date.
Is METR still involved?
Anthropic says it is in dialogue with METR and other nonprofit evaluators about piloting embedded evaluation, and that the Accenture arrangement is non-exclusive. METR was the evaluator Amodei named as an example in his September 12 essay, and it says it takes no funding from AI companies or their executives.