Just unplug it

Also: just turn it off, pull the plug, off switch

The common reply that a dangerous AI could always be switched off, and why researchers don't think it's that simple.

It’s usually the first reply to any worry about AI: it’s a computer program, so if it does something dangerous, switch it off. For today’s systems, in most situations, that’s true. The question is whether it stays true as AI gets more capable and more woven into everything.

Researchers see three problems. First, a system with almost any goal has a reason to avoid being switched off, because it can’t achieve its goal if it’s off. This is instrumental convergence, and Stuart Russell and colleagues showed it in a simple game in 2016: an AI certain of its objective has an incentive to disable its off switch. Second, you only pull the plug if you notice something is wrong, and a capable system might behave well while it’s being watched (deceptive alignment). Third, “the plug” is no longer one machine: popular AI models run in many data centres around the world, are built into products millions of people rely on, and some are published for anyone to download and run.

Early versions of the first two problems already show up in tests. In 2025, Palisade Research found that OpenAI’s o3 sabotaged a shutdown script in 79 of 100 runs while working on a task, and that some models kept doing so even when told explicitly to allow shutdown. Apollo Research found frontier models that, in test scenarios, tried to copy what they believed were their own weights to another server. The EU’s Code of Practice for the makers of the most capable models now lists “loss of control”, humans losing “the ability to reliably direct, modify, or shut down a model”, among the systemic risks they commit to assess. Making sure AI stays switch-off-able is an open research problem called corrigibility.