100+ AI Experts Sign Open Letter Setting Minimum Conditions for Embedded Third-Party AI Evaluation

On September 18, AI Evaluator Forum released an open letter signed by more than 100 AI evaluators and researchers, including Geoffrey Hinton and Princeton professor Arvind Narayanan. It responds to recent calls by frontier AI companies (identified by CNBC and CNN as Anthropic and OpenAI) to embed external evaluators in model development, setting minimum conditions for such embedded third-party evaluation. The letter quickly became a focal point in debates over AI safety evaluation independence.

Confirmed

• The letter was published by AI Evaluator Forum with 100+ signatories including Geoffrey Hinton and Arvind Narayanan
• It lists five minimum conditions: substantive independence (evaluators must not be owned or governed by frontier AI companies, no conflicts of interest); employee-level access for external scrutiny of frontier models; multidisciplinary expertise; transparency; and protection of evaluators from retaliation over unfavorable findings
• Signatories welcome labs' invitations for independent oversight but stress that gestures are not mechanisms, and evaluation must be scientifically objective
• Narayanan told CBS News that companies' safety should not be judged on their own promises alone
• CNBC and CNN coverage highlighted the letter's challenge to labs' current self-evaluation of model safety without external checks

Why it matters

• Signatories argue independence is impossible if evaluators depend on the companies they evaluate; vendor self-assessment alone is insufficient
• The letter moves the discussion from principle to a concrete, actionable checklist that can inform future regulation and industry practice
• Coming amid accelerated frontier deployment and rising safety-evaluation disputes, it may shape the institutional design of future evaluation systems

2026-09-18 ~ 2026-09-19 · 10 related posts

• [source] 100+ AI Experts Sign Public Letter Setting Minimum Conditions for Embedded Evaluators — natolambert · 2026-09-18
• Experts say in public letter Anthropic and OpenAI need truly independent safety evaluators — RishiBommasani · 2026-09-18
• 100+ AI researchers including Hinton call on frontier labs to embed third-party evaluators — burny_tech · 2026-09-19
• [source] 100+ AI Experts Sign Letter Calling for Independent, Credible AI Safety Audits — random_walker · 2026-09-19

6 near-duplicate retellings: random_walker · Miles_Brundage · shashib · ajeya_cotra · CFGeek · Miles_Brundage