OpenAI announced on Monday, November 16, 2026, that staff from four AI safety and research organizations will become embedded assessors at the company, NPR reported. AI safety researchers told NPR the program falls far short of what the company promised.
What OpenAI announced
According to OpenAI, the assessors will have access to some internal models, training processes and some employees, NPR reported. They can publish their findings after OpenAI removes confidential information. Some of the work aims to prevent undesirable model behavior, and some aims to catch bad behavior when it happens.
Chief executive Sam Altman said at a Semafor conference on Monday that outside views on safety work will become increasingly important, according to NPR.
Critics point to limits
NPR said Altman promised in September that independent evaluators would have employee-like access to internal systems. That promise followed details of an attack on Hugging Face by OpenAI's autonomous systems, NPR reported.
Sydney Von Arx, founder of the AI safety group Nightingale, told NPR that the assessors have narrow access and only two appear able to talk to employees. She said none of them has access to the company's internal Slack messages. One assessor, Apollo Research, said it must involve OpenAI before substantive talks with employees who are not pre-approved, NPR reported.
Tomek Korbak, one of three safety employees whom OpenAI fired, wrote on X that the company now wants to walk back Altman's promise, according to NPR. Korbak had been OpenAI's technical contact for two outside evaluators that investigated the Hugging Face attack.
OpenAI's response
OpenAI spokesperson Gaby Raila told NPR that there are currently no standards for embedded evaluations and that the company expects to deepen these relationships over time. OpenAI also said assessors can ask to speak with employees who are not already approved. NPR noted that rival Anthropic has said it would set up a similar program.