|
Getting your Trinity Audio player ready...
|
Anthropic has selected consulting giant Accenture to become its first embedded external evaluator for AI safety, a partnership the company announced on September 18 that puts outside personnel directly inside Anthropic’s walls to scrutinize its models before they reach the public. Both companies expect to invest at least $1 billion each over the next five years, bringing the combined commitment to roughly $2 billion for what they are describing as a new model for how frontier AI safety oversight should work.
The arrangement runs through Faculty, the applied AI division Accenture acquired in January, whose staff will now operate with what Anthropic has described as access comparable to an employee’s. According to Anthropic’s own blog post announcing the deal, Faculty personnel will handle evaluating and red-teaming Anthropic’s models, a practice that involves deliberately trying to make AI systems misbehave in order to expose weaknesses before they become real-world problems, alongside conducting broader alignment assessments and testing the safeguards built into Anthropic’s products.
The move fulfills a specific pledge Anthropic CEO Dario Amodei made just days earlier. In a September 12 essay titled “We Must Pace the Frontier,” Amodei argued that AI companies need to deliberately slow the pace at which they improve model capabilities and proposed bringing in independent evaluators with genuine, employee-level access inside frontier labs, rather than relying on the more limited, arm’s-length audits that have typically characterized external AI safety review. Amodei described the kind of access he envisioned in fairly specific terms, evaluators who could observe training runs directly, examine deployment data, and speak with staff without having to route every question through a formal external-review channel. The Accenture partnership represents the first concrete step toward that vision actually taking shape.
The choice of Accenture specifically caught many AI industry watchers off guard. Prior discussion of embedded evaluators inside frontier labs had largely centered on dedicated AI safety research nonprofits such as METR, Redwood Research and Apollo Research, organizations built specifically around the kind of technical alignment work Anthropic says it wants Faculty to perform. Accenture, by contrast, is best known as a massive, publicly traded management consulting firm, and the announcement moved markets almost immediately, with Accenture’s shares jumping more than 8 percent in after-hours trading following the news, reversing a nearly 5 percent decline the stock had posted during the regular session amid broader investor concern that AI tools could erode demand for traditional consulting services.
Anthropic has been careful to frame the Accenture deal as non-exclusive rather than a singular, complete solution to the independence question raised by embedded evaluation. The company said it remains in active discussions with METR and other third-party evaluators to pilot similar arrangements using separate funding sources, and noted that Faculty is expected to offer comparable embedded evaluation services to other AI developers beyond Anthropic itself, a structure intended to preserve some measure of Faculty’s independence from any single client relationship.
That independence question sits at the center of the skepticism the deal has attracted. Anthropic will fund Accenture’s work directly, at least in the near term, a decision the company has attributed to the absence of any established, mature financing model for independent AI evaluation work industry-wide. Critics point out that an evaluator paid directly by the company it is evaluating faces an inherent structural tension, regardless of how much operational access it receives, since proposed industry standards for genuine evaluator independence typically call for evaluators free from AI-company ownership, commercial ties and outcome-contingent payment structures, conditions the current Anthropic-Accenture arrangement does not fully satisfy on its face. Whether $1 billion in committed spending translates into substantive red-teaming and alignment research, or gets absorbed into standard consulting fees and infrastructure costs Accenture would have incurred regardless, remains genuinely difficult for outside observers to verify from the details made public so far.
The timing of the announcement has also drawn attention. Anthropic is reportedly targeting a valuation approaching $2 trillion as it moves toward a planned initial public offering expected around November, a context that has led some observers to read the safety partnership partly as a deliberately timed signal to regulators, investors and the public ahead of that listing, even as the underlying safety concerns Amodei has raised publicly, including his stated view that advanced AI carries a real chance of causing serious harm within the next decade, appear to be genuinely held rather than purely performative.
The broader industry backdrop lends some weight to the seriousness of the move regardless of how the independence questions eventually resolve. Prominent AI researchers including Geoffrey Hinton and Yoshua Bengio, both widely credited with foundational work underpinning the current generation of AI systems, have spent roughly two years warning publicly that AI labs are advancing capabilities faster than their internal ability to test for safety can keep pace. Market reaction beyond Accenture’s stock has reflected a similar undercurrent of risk-awareness, with cybersecurity companies including CrowdStrike and Palo Alto Networks posting sharp gains recently as investors position portfolios around the assumption that heightened AI capability will bring heightened AI-related risk requiring more oversight infrastructure, not less.
Whether the Accenture partnership becomes a genuine template other AI labs follow, or remains a one-off arrangement specific to Anthropic’s particular moment and incentives, will likely depend on how much substantive findings from Faculty’s embedded work become public, and whether Anthropic’s parallel discussions with independent nonprofits like METR eventually produce a second, more clearly independent track of evaluation running alongside this one. Continuing coverage of how AI safety oversight structures are evolving across the industry is available on Business Tech. Additional detail on the partnership is available through Anthropic’s official announcement, and further reporting on the deal’s reception can be found through TechCrunch’s coverage of the arrangement.
