Anthropic and OpenAI need truly independent safety evaluators, experts say in public letter
A coalition of over 100 AI experts are urging independence and transparency from Anthropic, OpenAI and other foundation model labs to conduct evaluations.
TL;DR
- Over 100 AI experts and evaluators are demanding necessary resources and protections for independent AI safety testing.
- The group wants to ensure AI companies provide scientific objectivity, transparency, independence, and robust protections for third-party evaluators.
- This initiative aims to hold foundation model providers accountable for their pledges to support third-party AI safety testing.
- Key proposals include independent ownership, avoiding commercial conflicts, and accepting payment only for work performed, not for specific findings.
- Evaluators should have access equivalent to internal employees, including candid communication and sensitive internal data, with limited exceptions.
- Companies should facilitate transparency, limit NDAs, and allow prompt communication with boards and public release of findings.
- Evaluators need protection from retaliation, including from retaliatory litigation.
- The efforts complement, not replace, internal evaluation efforts and broader external oversight by companies.