September 18, 2026
Independently testing AI safety is nearly impossible: more than 100 experts sign letter
On September 18, 2026, more than 100 AI evaluation experts signed a letter to Anthropic, OpenAI, and other foundation-model developers. They demand that independent evaluators receive access to the labs' systems, data, and tools. Signatories include Geoffrey Hinton and representatives of Johns Hopkins University, Stanford University, and METR.
In December 2025, AEF-1 established five evaluation conditions: access and resources, no conflicts of interest, analytical autonomy, methodological transparency, and protection of sensitive information. For substantially new systems, the standard often provides for at least 20 working days.
Access is expanding. On September 12, 2026, Anthropic committed to giving an external team ongoing staff-level access to its systems, data, tools, and facilities. The team will be able to evaluate completed models and training pipelines, and publish its findings without Anthropic editorial control.
Budget changes the outcome. On May 29, 2026, OpenAI cited an example where increasing UK AISI's budget from 10 million to 100 million tokens improved the cyber evaluation result by up to 59%. The letter's authors therefore demand funding even when the conclusion is unfavorable to the company, as well as protection for evaluators from retaliatory lawsuits.
The letter proposes formalizing evaluators' direct link to boards of directors and their right to publish disagreements with lab staff.
Source
