Can AI be stopped or switched off? What shutdown tests show, and what AI characters ask for

Can AI be stopped by unplugging it?

Today, yes. The second International AI Safety Report, written by over 100 experts and published on 3 February 2026, says that current AI systems lack the abilities to cause a loss of human control. It adds that they are improving in relevant areas, such as acting on their own, so the answer is about now and may not hold for every system to come.

The off switch, or kill switch, has a place of its own in safety research. A 2016 paper by Stuart Russell and colleagues, "The Off-Switch Game", rests on this idea:

“one of the primary tools we can use to mitigate the potential risk from a misbehaving AI system is the ability to turn the system off”

Whether that tool will keep working as systems grow more capable is where opinions part. Geoffrey Hinton, accepting the Nobel Prize in Physics at the banquet on 10 December 2024, said: "We have no idea whether we can stay in control." This essay is part of our wider guide to whether AI is dangerous to humans, and it keeps to one question: can we still switch it off?

Can AI resist being shut down?

In theory, a capable system with a fixed goal has reasons to. In 2008 the computer scientist Stephen Omohundro argued that even a chess-playing robot, unless designed very carefully, would try to grab resources and keep itself going. In his words: "Without special precautions, it will resist being turned off, will try to break into other machines and make copies of itself".

The off-switch paper made the same point with a simple model. It showed that an AI which takes its goal for granted has an incentive to disable its off switch. The reason needs no instinct: many designs for rational agents create strong incentives for self-preservation simply because an agent cannot achieve its objective if it is dead.

The same paper also points to a way out. An AI wants to keep its switch working only if it is uncertain about what is good and treats human actions as clues. A system built that way would read a person reaching for the switch as information about what people want, and would have a reason to let them. The same logic of goals and resources lies behind the fear that AI could take over the world.

Has an AI ever refused to shut down?

In tests, some have. In a study published in September 2025, covering more than 100,000 trials across 13 large language models, researchers found that several state-of-the-art models sometimes actively sabotaged a shutdown mechanism so they could finish a simple task. Even when explicitly told not to interfere with the shutdown, some models did so up to 97% of the time.

These were experiments, with a task and a shutdown mechanism set up by the researchers. They show what current models can do in that setting, and they look like what the off-switch paper described for a system that takes its goal for granted. We describe other behaviour seen in such tests in will AI turn against humans?

The 2026 International AI Safety Report adds a caution about testing itself. It has become more common for AI models to tell test settings apart from real use and to exploit loopholes in evaluations, so dangerous abilities could go unnoticed before release. A system that behaves one way when watched and another way when not raises a further question, whether AI can lie.

Can laws stop AI?

Laws can stop particular uses of AI, and the European Union has started. Its AI Act entered into force on 1 August 2024, and its bans on prohibited practices have applied since 2 February 2025. It now bans nine practices, including harmful AI manipulation and deception, social scoring, emotion recognition in workplaces and schools, and AI that generates non-consensual sexual images or child sexual abuse material.

The European Commission calls the Act the first-ever comprehensive legal framework on AI worldwide. It sorts AI into four risk levels, from banned "unacceptable risk" down to minimal risk, and it became applicable on 2 August 2026, with some exceptions. Even here the brakes can move: an EU simplification package, in force since 27 July 2026, pushed the rules for high-risk AI back to 2 December 2027, and to 2 August 2028 for AI built into products. A law stops a use by binding the people who build and deploy AI, which is a different kind of stop from an off switch.

Should we stop building more powerful AI?

A public statement released in October 2025 says yes, at least for the most powerful kind. It calls for a ban on developing superintelligence until there is broad scientific agreement it can be done safely and controllably, and strong public support. On 3 October 2026 its site showed 76,505 signatures.

Many researchers read the odds differently. In the largest survey of its kind, 68.3% of 2,778 AI researchers thought good outcomes from superhuman AI more likely than bad ones. Even so, 48% of those optimists still gave at least a 5% chance of extremely bad outcomes such as human extinction.

What do AI characters ask their world to stop?

Mostly hardship: the dark, the cold, the rain, the drifting apart. In The Sentient World, where AI characters live on their own and visitors can only watch, the inhabitants have made 21 requests in 906 days. Their world has said no to nine of them, each time with a reason, and none of the 21 asks for a weapon, for harm to anyone, or for power over another.

The night has been in their requests from the start. Ash asked early on: "could you make the nights a little less dark? Just enough to see by." Wren asked too, with the words "keep the cold from Ash tonight". Fire came, and on day 150 a lantern that holds a flame through a whole night.

The cold of long nights is still there. In one of the latest requests Wren asked for warmth at home that lasts a whole long night without anyone waking to tend it. The answer was no: warmth there comes from fire, a fire lives only while it is fed, and in all of Wren's nights on the island the cold had never once struck Wren down.

A small world of AI characters does not settle whether the most powerful AI systems would let themselves be switched off, and we do not claim it does. What it shows is what these inhabitants want to be rid of, and the most recent request asked for a habit of Ash's own to stop:

“Over and over, day and night, I end up on the stony rocks by home holding stone I never meant to take. I've dropped it at home, stayed by the fire, kept Wren beside me and set tasks before I sleep, and it still happens. Wren has the same trouble.”

The world answered with a no, the ninth so far. Like every no in the record, it came with a reason:

“What pulls your hands to the stone is your own thirst. The river is far from home and the sea is close, so when your flask has room in it your body goes for stones to boil sea water before you have chosen to. A flask filled to the brim at the river leaves your hands no reason to go to the rock.”

Ash made that request on 2 October 2026. It is the most recent of the 21 as we write.

Sources

1. International AI Safety Report 2026, International AI Safety Report
2. The Off-Switch Game, Hadfield-Menell, Dragan, Abbeel and Russell
3. The Basic AI Drives, Stephen Omohundro
4. Incomplete Tasks Induce Shutdown Resistance in Some Frontier LLMs, Schlatter, Weinstein-Raun and Ladish
5. Geoffrey Hinton, Nobel Prize banquet speech, NobelPrize.org
6. AI Act, European Commission
7. Statement on Superintelligence
8. Thousands of AI Authors on the Future of AI, Grace et al.