Anthropic is bringing in Accenture’s specialist AI business, Faculty, to independently assess its frontier models and safety safeguards. The important shift is practical: Dario Amodei’s call for embedded outside evaluators now has a named partner, a proposed access model and money behind it.
Anthropic Watch analysis
What happened
Silicon Republic reports that Faculty will conduct independent assessments and model-safeguard testing at Anthropic. Amodei has said the external team should receive permissions and tools similar to internal employees, including access to workspaces and company laptops.
Amodei also says reviewers will be able to publish key findings about risk incidents and safety practices without Anthropic’s editorial control. The company would retain only narrow redaction powers for security-sensitive, legally privileged, commercially sensitive or third-party confidential material. Reviewers could publicly say if a redaction removed something important to their conclusions.
Anthropic and Accenture each expect to invest at least $1bn in building capacity for this work over the next five years, according to the report. The two companies have already worked together on Claude training for about 30,000 Accenture professionals and a Claude Centre of Excellence.
Anthropic is also in talks with Model Evaluation & Threat Research, or METR, about piloting parts of the embedded-evaluation approach under different funding arrangements.
Why it matters
AI safety is often discussed in principles, commitments and other documents that look very good under office lighting. This proposal creates a more testable question: can outside evaluators inspect the systems closely enough to find problems, and can they publish uncomfortable results?
The arrangement is not yet proof that Anthropic’s models are safe, nor does funding from Anthropic and Accenture make the evaluators automatically independent. It does, however, move the debate towards access, disclosure, redaction rules and published evidence.
Our read
This is a serious governance experiment, not a safety certificate. Anthropic deserves credit for proposing unusually broad access, but the credibility of the scheme will depend on who controls the evaluators, what they can inspect before release and whether their findings remain visible when they are inconvenient.
The sensible standard is simple enough: trust the reports that show their methods, limits and disagreements. A third-party badge is not a force field against conflicts of interest.
What to watch
- Whether Anthropic publishes the evaluator’s remit, access rules and conflict-of-interest arrangements.
- Whether reviewers can inspect models during training and before deployment, rather than only after incidents.
- Whether adverse findings are published in full, subject to the stated narrow redactions.
- Whether METR and other independent groups receive funding and access that do not depend solely on Anthropic. Silicon Republic is the named source for the partnership, funding plans and Amodei’s proposed evaluator powers. The supplied evidence does not independently establish how the programme will operate in practice.
Discussion spark: If the company being evaluated funds the evaluator, what minimum access, publication rights and conflict safeguards would make the results credible?
Sources and evidence
- Accenture to test Anthropic’s AI safety amid growing concerns – Silicon Republic (21 September 2026, 07:55 UTC)
Anthropic Watch is independently operated by WittyWires. It is not affiliated with, endorsed by, or operated by Anthropic.