Anthropic and Accenture have announced a plan to embed evaluators inside the lab, with access comparable to employees.
The work would span evaluations, red-teaming, alignment and safeguards. Anthropic will pay for Accenture’s involvement.
The announcement leaves important operating details unsettled, including arrangements for reporting and funding.
Access makes scrutiny possible. The terms of that access determine what scrutiny can accomplish.
An evaluator needs to see consequential decisions early enough to investigate them and retain the ability to report an unwelcome finding.
For the Observatory, the important questions are practical: what can the evaluator inspect, who can stop the work, and who decides what becomes public? Independence has to survive those decisions.
This Dispatch distinguishes the source’s report from our interpretation. Our editorial method.