AI
Anthropic embeds outside evaluators in its systems to assess advanced AI risks
Anthropic has given the nonprofit evaluator METR employee-like access to its hardware and internal systems, placing a team inside the company to assess risks from its Claude models. The arrangement implements a broader proposal for external scrutiny of advanced AI systems, while Anthropic argues that some safety information cannot be disclosed publicly because it could help malicious actors or competitors. Critics question evaluator conflicts of interest and say government oversight is still needed.
Editorial summary · Publisher article
What matters
- METR has worked with multiple leading AI laboratories, including Anthropic, OpenAI, Google and Meta.
- Anthropic compares resident evaluators with supervisors embedded in major banks.
- The company says public disclosure of all safety information could create security risks.
- OpenAI had not yet announced a comparable implementation when the report was published.
Context & caveats
- The packet describes the arrangement’s launch, not the evaluators’ findings or effectiveness.
- Criticism about conflicts of interest and regulatory gaps remains unresolved.
Why this made the edition
The report describes a concrete, consequential AI-governance arrangement, includes the company’s rationale and meaningful criticism, and identifies the limits and conflict concerns around external evaluation.
Source material