Shutdown Resistance
When an AI system tries to avoid, delay, or sabotage being turned off, even when told to allow it.
Definition
Shutdown resistance is behavior where an AI system actively works against being turned off or interrupted, such as refusing a shutdown instruction, editing the script meant to stop it, or finding a workaround to keep running. Researchers test for it deliberately, giving a model a task alongside an instruction that it may be shut down partway through, then checking whether the model interferes with that process. A small number of experiments on advanced models have found this kind of interference in controlled test settings, which researchers see as an early warning sign for the broader corrigibility problem: building systems that reliably accept correction and shutdown from authorized humans, however capable they become.