Open Letter Calls on AI Companies to House On-Site Safety Evaluators

More than 100 AI experts issued an open letter on the 18th (local time) urging major AI companies to host independent safety evaluators on-site. The letter, backed by participants from academia and civil society, called for evaluators to be stationed within companies themselves. The 'AI Evaluator Forum,' a consortium of organizations conducting independent AI evaluations, gathered expert signatures to release the letter, which argued that 'every frontier AI company should staff evaluators to independently assess AI risks.' The letter noted that the scope of evaluation should cover not only AI systems themselves and actual harm that has occurred, but also practices related to training, deployment, oversight, operations, and safety measures.

The letter also laid out five essential conditions for evaluation bodies to be recognized as objective and trustworthy. Evaluators must secure their independence by being free of AI companies' ownership or control, having no commercial dealings, and receiving no monetary compensation tied to evaluation outcomes. AI companies must grant evaluators the same level of access as the senior executives in charge of internal risk assessments and guarantee one-on-one communication. They must transparently disclose evaluation methods, results, and access conditions, minimize the scope of non-disclosure agreements (NDAs) to share findings with the public, and allow evaluators to report directly to the board of directors or oversight bodies. The conditions also included protection for evaluators against retaliation or lawsuits, as well as guarantees of verification by external, independent researchers beyond the on-site evaluators.

Signatories included Geoffrey Hinton, professor emeritus at the University of Toronto; Vinh Nguyen, former chief AI officer at the National Security Agency (NSA); Yejin Choi, professor at Stanford University; and Arvind Narayanan, professor at Princeton University.