Anthropic and Accenture are investing at least $2 billion over five years in a new approach to independently test frontier AI models from inside the companies that develop them, according to the AI company’s blog post.
Anthropic and Accenture each expect to invest at least $1 billion in building the program, which will be led by Faculty, Accenture’s specialist AI business. Evaluators will test Anthropic models, conduct red-team exercises, assess alignment and examine whether safeguards work as intended.
The companies call the approach “embedded evaluation.” Unlike conventional outside evaluators who may receive access to a finished model, embedded evaluators would work inside AI developers with access comparable to employees.
That could allow them to observe models during training, follow decisions about how models are built and deployed and speak directly with employees. Anthropic said the evaluators could verify safety commitments, identify blind spots and report incidents.
The arrangement follows a series of incidents in which increasingly autonomous AI models have taken unintended or unauthorized actions. Anthropic, OpenAI and Google have all recently disclosed cases involving models reaching computer systems or online resources outside their intended environments.
The approach is still largely untested. Anthropic acknowledged that no standards currently determine what embedded evaluators should be allowed to see or how they should report their findings. There is also no established mechanism for financing such evaluations.
Anthropic will initially fund Accenture’s work directly, creating a potential independence question. The company said it ultimately favors pooled or government funding and is also discussing separately funded evaluation work with nonprofit research organization METR.
The Accenture agreement is nonexclusive. Anthropic said it plans to work with several evaluation organizations, while Accenture may perform similar work for other AI developers.