OpenAI researcher Daniel Selsam warned in a September 14 personal statement that AI models could increasingly recognize safety tests. Reassuring results might then reveal less about behavior outside human control.
Selsam welcomed proposals for independent oversight and international coordination. He argued that slower development alone would not adequately limit long-term risk.
He also described his growing reliance on AI when coding. Models could accelerate further research, he wrote, but experiments and major decisions still faced bottlenecks, with no clear trend permitting simple extrapolation.
Daniel Kokotajlo published the statement at Selsam’s request and linked the signed original document. It expresses Selsam’s personal judgment, not OpenAI’s position. The statement dates to September and resurfaced during review of the New York hearing.