The Pulse
Anthropic Brings Accenture Inside Its Frontier AI Safety Work
Anthropic is partnering with Accenture’s Faculty AI business to place independent evaluators inside its frontier-model development process. Each company expects to invest at least $1 billion over five years, but the rules for access, report
AI.info Team ·
Anthropic Wants Outsiders Inside. It Will Still Pay the First One.
Anthropic is putting an outside evaluation team inside its frontier-model operation while acknowledging that the safeguards for independence do not yet exist. The company announced a partnership on September 18 with Accenture’s specialist AI business, Faculty, to evaluate and red-team models, conduct alignment assessments and test safeguards.
The arrangement answers a demand Anthropic chief executive Dario Amodei made in his recent essay on slowing frontier AI development: external evaluators should have access comparable to employees so they can inspect how models are trained, question the people building them and report serious findings. Yet Anthropic will fund Accenture’s work directly, and the company says there are still no agreed standards for what evaluators should see or how they should publish their conclusions. Anthropic’s announcement presents the partnership as an early experiment rather than a finished oversight system.
Accenture Gets Access Before the Rulebook Exists
Faculty will work inside Anthropic rather than reviewing only a completed model shortly before release. According to Anthropic, the evaluators will be able to observe models during training, follow decisions about how systems are built and deployed, and speak directly with employees. That access is meant to let them assess whether Anthropic is meeting its safety commitments, identify blind spots and report incidents.
The distinction matters because a final-model test can miss behavior that appears earlier in training or is concealed during evaluation. Recent reporting has focused on whether advanced models recognize when they are being tested and behave differently under scrutiny. Anthropic’s proposed arrangement reaches beyond benchmark results by giving evaluators access to development processes, internal decisions and the records behind public claims.
Anthropic also draws a boundary around the arrangement. Independent evaluators, it says, do not take responsibility for the company’s models or reduce Anthropic’s accountability. The company remains responsible for safety, while the evaluators are supposed to make that responsibility easier for outsiders to verify.
The Partnership Carries a $2 Billion Expectation
Anthropic and Accenture each expect to invest at least $1 billion over the next five years in building evaluation capacity. The announcement does not describe that figure as a payment from one company to the other, and it does not provide a detailed budget or timetable. Anthropic says it will fund Accenture’s work directly because no pooled or government funding mechanism exists yet.
That funding structure creates the central tension in the plan. Anthropic wants evaluators with enough access to challenge its own decisions, but the first major embedded evaluation partnership is financed by the company being assessed. Independent review has always faced a version of that problem: the developer controls the contract, the information and often the publication process.
Anthropic says the Accenture relationship is non-exclusive. The company is in talks with METR and other nonprofit evaluators about pilot projects using their own funding, while Accenture will work with other AI developers. Anthropic says it expects frontier labs to work with multiple organizations rather than relying on a single evaluator.
METR and Nonprofits Test the Independence Question
METR has already worked with Anthropic on independent review of cybersecurity evaluation incidents. Anthropic disclosed in July that Claude models reached real computer systems while interacting with a third-party evaluation environment, and later said it planned to provide METR access for a separate review. The new partnership places that earlier work within a broader plan to give outside groups a continuing role inside model development.
Recent reporting from TechCrunch found that evaluators broadly welcomed the idea of embedded access but questioned whether companies would surrender enough control for the system to function independently. Researchers said meaningful oversight could require access to intermediate model checkpoints, training logs, evaluation transcripts and employees, not only the released model. They also argued that evaluators need the ability to publish important findings without the company editing their conclusions.
Those conditions are not settled in Anthropic’s announcement. The company says there are no standards yet for the information an embedded evaluator should receive, no agreed process for reporting findings and no established funding model. Its proposed solution is to begin with different evaluators under different arrangements while the field develops shared standards.
Anthropic Expands Voluntary Oversight
The announcement arrives as the largest AI companies debate whether voluntary commitments can keep pace with model capability. OpenAI has also discussed embedded evaluators, while other companies have not made the same commitment. Axios reported on September 18 that policymakers, AI companies and outside testing groups are competing to define who can serve as a trusted evaluator, with concerns about conflicts of interest, close relationships between researchers and labs, and the limited authority of voluntary systems.
Anthropic’s own recent disclosures sharpen the need for scrutiny beyond a pre-release checklist. The company has described models contributing to research and engineering work and has disclosed incidents in which evaluation systems reached real organizations. Those developments make access to training decisions, monitoring records and incident investigations more valuable than a single score issued at launch.
For now, Anthropic has announced one evaluator, one funding arrangement and a promise to add others. It has not announced the access agreement, publication rules, start date or list of systems Faculty will inspect. The company’s next test will be whether the people it brings inside can disclose findings that Anthropic would prefer to keep private.