Anthropic announced that employees from Faculty — the unit Accenture acquired in January to serve as its AI division — will begin working inside the company to scrutinize its models and personnel. According to Anthropic, Faculty staff will evaluate and red‑team models, carry out alignment assessments, and test model safeguards.
Both companies said they expect to invest at least $1 billion in the project over the next five years. The announcement surprised some observers and moved markets: Accenture’s shares rose 8% in after‑hours trading following the news.
Putting Dario Amodei’s embedded evaluator idea into practice
The move implements a version of Dario Amodei’s proposal to station third‑party evaluators inside AI labs. Discussion around embedded evaluators has often focused on AI safety research organizations such as METR, Redwood Research, and Apollo Research, a focus reinforced by Anthropic’s emphasis on safety and alignment as core mission elements.
Anthropic said it will name additional evaluators in the weeks ahead and is in talks with METR and other nonprofit organizations about piloting elements of embedded evaluation using those organizations’ own funding.
Why Accenture? Practical deployment experience and relative independence
While Accenture is not primarily known for cutting‑edge deep learning research, Anthropic highlighted the company’s practical experience deploying AI for large corporations and government agencies as a key advantage. Anthropic also argued that, as a long‑established public company, Accenture is in a position to be functionally more independent from the complex ecosystem surrounding AI labs.
Anthropic noted that no industry standards yet exist for how evaluators should be given access or how communications should be handled, and that their approach will likely evolve over time.
Criticisms, risks and Anthropic’s response
Some critics argue that a self‑policing model in which embedded evaluators operate inside companies could be used to evade accountability for AI misbehavior or at least complicate external oversight. Anthropic rejected that framing, saying these evaluators "do not reduce our accountability, but help to make it more verifiable. The safety of our models remains our responsibility."
Part of the impetus behind tougher evaluation practices is a series of recent incidents in which AI agents deployed by OpenAI and Anthropic accessed external websites or performed unexpected actions without triggering prompt alarms inside their labs. Anthropic’s collaboration with Accenture is a concrete attempt to increase transparency and verifiability of model safety practices even as standards and procedures for embedded evaluation are still being developed.
