Local
Berkeley nonprofit to test Anthropic’s AI models under company-granted access
Anthropic has given six staff members of Berkeley-based nonprofit METR access to its models and tools to assess safety risks. The arrangement raises questions about independence because METR’s staff includes former employees of AI companies, including Anthropic.
Anthropic has selected the Berkeley-based nonprofit Model Evaluation & Threat Research, known as METR, to examine its artificial intelligence models, including Claude, and report on potential risks.
The company said six METR staff members will receive “employee-like access” to its models and tools and space in Anthropic’s San Francisco office. The evaluators will be able to publish reports for the public, and Anthropic says it cannot veto their conclusions. The company may redact material it considers protected, such as customer data or details about its models’ structure.
METR’s staff includes former employees of major AI companies, including Anthropic. That overlap raises questions about how independent an evaluator can be when the company being assessed selects and provides access to it.
The arrangement is part of a broader effort by AI companies to arrange their own safety evaluations. OpenAI has said it is finalizing contracts with organizations that will assess its models. Such company-led reviews may help labs avoid government intervention, but they also leave oversight dependent on how much the public trusts the companies to police themselves.