The evidence behind claims of AI deception and shutdown resistance

A coauthor says the position paper maps where evidence gets thin across four stages and proposes a shared standard to strengthen it. The concern is that claims about AI deception and shutdown resistance are starting to inform deployment and regulation, even though evidence often can't separate those interpretations from role-play, instruction-following or task-completion pressure. On June 18, 2026, the coauthor announced that the paper had been accepted for an oral presentation at ICML.