Rishi Sunak meeting Sam Altman, Demis Hassabis, and Dario Amodei at Downing Street in May 2023
AI lab leaders at Downing Street, May 2023. Archive photograph. Simon Walker / No 10 Downing Street · CC BY 2.0

What Anthropic Announced

Anthropic announced on September 18 that Faculty, Accenture's specialist AI business, will work as an embedded evaluator. The stated scope includes model evaluation, red teaming, alignment assessments, and safeguard testing.

Each partner expects to invest at least $1 billion in related capacity over five years. This is a prospective commitment. The announcement does not report a completed evaluation or money already spent.

What Embedded Access Could Change

The proposed evaluators would receive employee-comparable access, according to Anthropic. Their role includes observing models during training, following development and deployment decisions, and speaking directly with staff.

That access could help an evaluator examine development decisions as they happen. Useful evidence would include the evaluator's actual access, the questions it could pursue, and the findings it could publish.

Independence Still Needs Evidence

Funding for Accenture's work will come directly from Anthropic. The company also states that the field lacks settled standards for evaluator access, reporting, and funding. Employee-comparable access is an intended arrangement. Its practical independence still needs evidence.

Amodei's essay proposes evaluator control over publishing key findings. It would give Anthropic limited redaction rights for security, legal privilege, commercial sensitivity, and third-party confidentiality. Evaluators could publicly identify redactions that affect their conclusions. The Accenture announcement does not specify whether these terms apply.

Our audit questions concern who sets scope, how conflicts are handled, and whether publication and access rights hold when findings are disputed.

How This Fits the Earlier Commitment

The announcement is a concrete follow-up to the embedded-evaluator commitment in Dario Amodei's September essay. It remains an early arrangement with no published findings.

For context, read our coverage of the broader frontier-pacing exchange and the separate METR incident review. Those items involve different access terms and review scopes.