Open Letter: Minimum Conditions for Embedding Evaluators
4 hours ago
- The letter supports frontier AI companies embedding third-party evaluators to independently assess AI risks, including system evaluation and incident response.
- Evaluators must be scientifically objective, transparent, and independent, with protections against interference from companies.
- Key conditions include: independence from company ownership/governance, no payment contingent on findings, and multiple evaluators with diverse expertise.
- Transparency is required for methods, findings, access, and terms, with limits on NDAs and a time-limited redaction process for critical interests.
- Evaluators need protection from retaliation, including legal and funding safeguards, and must have access equivalent to senior internal staff.
- These conditions should be standardized and codified, with AEF-1 as an early example, but more work is needed.
- Embedded evaluations complement, not replace, broader oversight like public transparency and independent researcher access.
- The letter lists over 100 signatories in personal capacity, including notable figures like Geoffrey Hinton and Stuart Russell.