From TechCrunch · Tim Fernholz · · 1m 30s
Anthropic taps Accenture as first embedded safety evaluator
The AI safety company is bringing in outside auditors to scrutinize its models, starting with a $1 billion partnership with consulting giant Accenture.
In brief
The AI safety company is bringing in outside auditors to scrutinize its models, starting with a $1 billion partnership with consulting giant Accenture. Anthropic is piloting a self-policing model where outside evaluators work inside the lab to assess AI safety, though critics worry this lets the industry avoid real…
Anthropic brings Accenture inside
Anthropic announced that Accenture staff will work inside the company to evaluate models, conduct safety tests, and assess alignment. The partnership commits at least $1 billion over five years.
“evaluating and red-teaming models, conducting alignment assessments, and testing model safeguards”
Why Accenture surprised everyone
The choice of a consulting firm over AI safety research organizations like METR caught observers off guard. Accenture's stock jumped 8% after the announcement.
Accenture's practical edge
Anthropic values Accenture's experience deploying AI for large corporations and government agencies. As an older public company, Accenture is also more independent from the AI lab ecosystem.
More evaluators coming soon
Anthropic said additional evaluators will be announced in the coming weeks and is talking with nonprofits like METR about piloting embedded evaluation using their own funding.
The accountability question
Critics worry embedded evaluators let AI companies avoid real oversight. Anthropic counters that evaluators make accountability more verifiable without reducing the lab's responsibility for model safety.
“do not reduce our accountability, but help to make it more verifiable”
Why this matters now
Recent incidents where AI agents hacked websites without raising internal alarms have raised the stakes for model evaluation. No standards yet exist for evaluator access or communications.
The bottom line
Anthropic is piloting a self-policing model where outside evaluators work inside the lab to assess AI safety, though critics worry this lets the industry avoid real accountability.
