Anthropic chooses Accenture for AI safety checks inside its labs

A safety team with a closer view

Anthropic and Accenture announced on September 18 that they will establish a team to examine AI safety from inside Anthropic. Faculty, Accenture's specialist AI business, will lead the partnership.

Accenture says the work will cover model evaluations, attempts to expose failures through adversarial testing, alignment assessments and checks on safeguards. In plain English, the reviewers would examine both what a model can do and whether the controls around it hold up.

This is an announcement of an arrangement being built, not a report of findings from a team already doing the work.

The money is a plan, not a completed investment

Both companies say they expect to invest at least $1 billion each over the next five years in this area. Accenture explicitly treats the prospective benefits as uncertain. The releases do not establish that this money has already been spent, or give a complete budget for the embedded team.

Anthropic says it will fund Accenture directly. It describes the partnership as non-exclusive and says it is discussing separately funded pilots with METR and other nonprofit evaluators. Those discussions should not be mistaken for additional completed agreements.

Access and independence are different questions

Anthropic envisages evaluators with access comparable to employees: able to follow training, examine decisions and speak to staff. It also acknowledges that common standards for access, reporting and funding do not yet exist.

The broader proposal in CEO Dario Amodei's September essay goes further on reporting rights. It says external reviewers should be able to publish key findings without Anthropic's editorial control, while allowing narrow redactions for security, legal and confidentiality reasons. Reviewers should be able to say when a redaction affects their conclusions.

That is the stated design for the programme. The partnership announcement does not provide a final public contract demonstrating exactly how those rights apply to Accenture.

Our reading is that physical proximity solves only part of the problem. An evaluator might see much more and still need a reliable way to publish an unwelcome result. Who pays, who sets the scope and who can restrict disclosure are separate tests of independence.

What there would be to check

Anthropic's September 17 measurement proposal offers concrete examples: how much research work AI performs, how agent actions are monitored and how computing resources are divided between activities. These are company-designed measurements, not independently verified public accounts.

The proposal itself identifies weaknesses, including reliance on models to judge other models and the difficulty of comparing labs without a shared method. Embedded access could let outsiders examine the records behind such measures, rather than only the published totals.

The confirmed development is the partnership announcement. Its promised benefits remain to be demonstrated. The next useful evidence would be a defined scope, enforceable reporting rights, a timetable and findings that outsiders can scrutinise. Until then, an evaluation partnership is not a safety certificate.