Could we switch off a dangerous AI?