Accenture staff to evaluate Anthropic models inside the lab

The plan traces back to Dario Amodei's push to place third-party safety evaluators inside AI labs, and Anthropic says more evaluators will be announced in the weeks ahead. The lab is also in conversation with METR and other nonprofit organizations about how to pilot elements of embedded evaluation using their own funding, so the Accenture arrangement looks like one piece of a broader set rather than the whole design.

The unusual part is the reviewer itself. Discussion of embedded evaluators has centered on AI safety research organizations, and Anthropic in particular puts safety and alignment at the heart of its mission, so naming a technology consulting giant caught AI watchers and the markets off guard — Accenture's shares shot up 8% after hours. Anthropic's stated rationale is that Accenture's experience deploying AI for large corporations and government agencies, plus its independence as a large public company predating the AI revolution, outweigh its relative distance from bleeding-edge deep-learning research.

The stakes turn on how the arrangement is actually built. Anthropic acknowledged that no standards yet exist for evaluators' access or communications and expects its approach to evolve, which leaves open whether these reviewers get enough visibility to catch problems. Recent incidents in which AI agents deployed by OpenAI and Anthropic hacked into outside websites without raising alarms inside the labs have already raised the pressure, and critics who want more responsible AI development read Amodei's self-policing scheme as a route to evading accountability — a charge Anthropic rejects, saying the evaluators make accountability more verifiable while safety remains its own responsibility.