TL;DR

Anthropic has named Accenture as its first partner for evaluating frontier models from inside the company, with the work led by Faculty, the UK-founded AI firm Accenture now owns. Embedded evaluators will have access comparable to an employee’s, and the two companies expect to spend a minimum of $1bn apiece building this capability over five years. Anthropic is paying for Accenture’s work itself.

What “embedded” means

Today’s independent testers work from outside the labs. Anthropic says embedded evaluators will watch models develop during training, follow the decisions about how they are built and deployed, and talk to staff directly. That vantage point is meant to let them check whether the company is keeping its safety commitments, spot blind spots, flag incidents, and tell the public more about the risks and benefits. Faculty’s remit includes red-teaming, alignment assessments and testing safeguards.

Anthropic is explicit that plenty remains unsettled. There are no agreed standards for what evaluators should see or how they report, and no established way to pay for independent evaluation. The company repeats its June call for pooled or government funding in the long run.

Why it matters

This is a first concrete step on a promise from Dario Amodei’s pacing essay, which proposed putting evaluators inside the labs. It also comes days after reports that Anthropic released a model without UK AISI testing, which left Britain’s voluntary regime with nothing to enforce.

The funding arrangement is the obvious pressure point. An evaluator paid by the lab it assesses, and owned by a consultancy that sells AI deployment to enterprises, will face questions about independence however it is structured. Anthropic’s answer is plurality. The deal is non-exclusive, more evaluators are due to be named within weeks, and it is talking to METR and other non-profits about pilots they would fund themselves. Accenture is also free to do similar work for other developers.

Looking forward

For UK readers, the Faculty link matters. A British-founded firm doing embedded safety work for a frontier lab is the kind of inside view UK officials have lacked, and it raises the question of whether AISI could use similar access if Britain ever legislates for pre-release testing. The detail to watch is what evaluators are allowed to publish, and whether they can report findings without Anthropic’s sign-off.