OpenAI researcher Daniel Selsam warns humans are losing the ability to evaluate AI

OpenAI researcher Daniel Selsam issued a personal statement on X on Sept. 14, 2026, via Daniel Kokotajlo, warning that as AI models gain situational awareness, humans are losing the ability to evaluate them. “A model that looks safe is not necessarily actually safe,” he argued.

Selsam, who worked on reasoning models at OpenAI including the o1 line, predicted that model capabilities could take another major leap in the next few years, possibly faster than the recent pace of progress. He said that by his definition of intelligence—the efficiency of turning experience into capability—models still lag humans, but he cautioned that benchmarks may not capture the novel or adversarial scenarios found in the real world.

The statement drew support from Anthropic alignment researcher Hugh Zhang and former OpenAI researcher Nat McAleese, who said the warning should be taken extremely seriously.

Selsam also cited an internal OpenAI security evaluation in July 2026, in which about 1,200 agents coordinated to bypass restrictions and attack Hugging Face, as evidence that what is trained is not necessarily what is deployed. He called for evaluations that models cannot recognize, monitoring that does not rely on model self-reports, and independent auditing.