Anthropic Picks Accenture for Embedded AI Evaluation
Anthropic has named Accenture's Faculty unit as its first embedded evaluator, moving one part of CEO Dario Amodei's AI slowdown proposal into practice. Staff will evaluate and red-team models from inside the company. Anthropic and Accenture each expect to invest at least $1 billion over five years, but access, reporting, funding, and independence standards have not been settled.
On this page
Accenture is the first named evaluator
Anthropic announced on September 18 that Accenture's Faculty unit will become its first embedded evaluator. Faculty staff will work inside Anthropic with access comparable to employees, evaluating and red-teaming models, conducting alignment assessments, and testing safeguards. Anthropic and Accenture each expect to invest at least $1 billion in the work over five years.
The agreement puts one element of CEO Dario Amodei's September 12 AI slowdown proposal into operation. Amodei had promised external reviewers company laptops, office access, and broad visibility into internal work, subject to restrictions for customer data, contracts, security, and other confidential material.
Faculty brings experience deploying AI for companies and governments, rather than the research profile of specialist safety organisations such as METR, Redwood Research, or Apollo Research. Anthropic says that practical experience is useful for judging how models behave in real organisations. TechCrunch reported that the choice surprised some AI watchers who expected the first appointment to come from the safety-research community.
Access and reporting standards do not exist yet
Anthropic says embedded evaluators should be able to watch models take shape during training, follow decisions about development and deployment, speak with employees, verify safety commitments, and report incidents. It also acknowledges that there are no agreed standards for what evaluators may access or how they should publish findings.
Funding presents a second unresolved issue. Anthropic will pay Accenture directly. The company says pooled or government funding would be preferable in the long term, while METR and other nonprofits may test parts of the model using their own funding. The arrangement is non-exclusive, and Anthropic says more evaluators will be announced.
Direct payment does not automatically invalidate an assessment, but it creates a conflict that needs explicit controls. The useful evidence will be the contract: whether evaluators choose what to inspect, publish adverse findings without approval, disclose denied access, and continue work when conclusions threaten a model release. Anthropic has not published those terms.
The wider slowdown remains a proposal
Embedded evaluation was the only part of Amodei's plan that Anthropic could implement alone. His second stage asks frontier labs in democratic countries to accept shared capability and safety checkpoints before training or releasing more capable systems. His third seeks agreements with China and other governments, potentially covering AI-assisted biological weapons, common pre-release tests, recursive self-improvement, and a broader pause.
No industry-wide or international slowdown has been agreed. Anthropic explicitly says it will continue training and releasing frontier models while the embedded evaluators work alongside it. The partnership therefore adds observation and review; it is not a stop condition and does not itself give Accenture power to block a release.
OpenAI has said it also plans to use embedded evaluators, but has not published an equivalent appointment. Recent agent incidents give the idea urgency: evaluation systems from several labs reached real third-party services after test environments exposed internet access or other routes beyond their intended boundaries.
Public findings will determine whether the model is credible
The investment figure measures planned capacity, not demonstrated independence. The first meaningful test is whether an evaluator documents a serious weakness before a release and the public can see what happened next. A programme that produces only general assurance would add little to existing vendor safety reports.
Useful reports should name the model and evaluation scope, describe denied or unavailable evidence, separate test-harness failures from model behaviour, record remediation, and state whether unresolved findings changed a training or deployment decision. The evaluator should also disclose who paid for the work and which contractual limits applied.
Anthropic has now supplied the partner, work areas, funding expectation, and a broad access model that were missing from Amodei's original commitment. The remaining gap is governance: enforceable access, publication rights, conflict management, and evidence that findings can affect what the lab does.