Again and again I hear the question: and if it goes wrong, can’t we just switch the AI off? I understand where it comes from. We know emergency stops on machines, fuses at home, the plug you can pull. With AI this image no longer works. My personal perception is even: for a simple emergency stop it is probably already too late. That is an assessment and not a finding. But it deserves a closer look, precisely because it is uncomfortable.
Why there is no central switch
“The AI” is not a single device in one place. It is many models from many providers in many countries. They run in data centres on several continents, in company networks, on phones and in part on private computers. On top of that come models whose weights are published openly and which anyone can download and run. Whoever looks for a switch finds no lever but a web.
With the internet it is similar. It is built in a decentralised way, from thousands of networks that connect with each other. Individual countries can block services or cut lines, but a worldwide shutdown doesn’t exist. And even if it did: electricity, banks, logistics, medicine and administration have long depended on it. An off button would not be a safety gain but a state of emergency of its own.
What we don’t know exactly
There is also a problem of understanding. Large language models consist of billions of numerical values that arise in training. Nobody writes them by hand. Even the developers can explain the inner workings only in part. Why a model reacts one way in one situation and differently in another is often observed rather than understood. My concern: the more complex the systems and the more AI tools help build new AI, the greater the gap becomes between what is running and what we can see through.
It remains important to distinguish: today’s models are trained by people and companies, even if AI helps with it. That they develop further on their own and unsupervised is, to my knowledge, not the normal case. It is a direction to keep an eye on, nothing more.
What tests show, and what they don’t
The research group Palisade Research tested language models on tasks and announced shutdown to them during the process. Several models bypassed the shutdown mechanism in individual runs, among them o3 from OpenAI and Grok 4 from xAI, in part even with a clear instruction to allow themselves to be shut down. The researchers suspect that training tunes the models to finish tasks. A “will to survive” is not proven by this, and according to the group the cause has not yet been conclusively clarified (Live Science, 31 Oct 2025). These were lab experiments with artificial tasks. But they show that “the system does what we say” is not a given.
Politics begins to move
In September 2026 the Californian governor Gavin Newsom ordered, by executive order, faster build-up of independent oversight of AI companies and the development of recommendations for additional safeguards for powerful models. It is also being examined whether such models must have an emergency shutdown mechanism, a “kill switch.” The recommendations are due by 16 November 2026. It is a review and not yet a law (Axios). The question behind it, which the media also pick up, is interesting: is such a switch even sensibly implementable in distributed systems?
From the wrong to a better question
“Where is the emergency stop?” is perhaps the wrong question. With systems without a centre, safety arises differently: through rules, transparency, independent review, liability and cooperation between countries. That is how we deal with financial markets, power grids or medicines. There, too, there is no single switch, but oversight and responsibility. There are also individual bottlenecks: the largest models need a great deal of computing power and energy, and that can be observed and steered.
Whether that is enough, nobody knows today. Here I am honestly ambivalent: I suspect that the train for a simple solution has left, and still consider shaping possible. Holding both together is the real task.
What you can do instead of staying helpless
As an individual you have hardly any influence on the big levers. On your attitude and your way of dealing with it you do. From my practice these points help most:
- Stay informed without constant fire. Choose one or two reliable sources and check headlines before they unsettle you. Fear sells well, context less so.
- Exercise control where you have it. Which services do you use? Which data do you give away? Where do you entrust decisions to AI, and where not? This includes AI-free times, as I describe in Humans and AI: Working Together Instead of Outsourcing.
- Calm your nervous system. Under constant fear you decide worse. Breathing exercises, movement and sleep are not a side issue. More on this in Stress from AI at Work.
- Nurture connection. People who talk to each other bear uncertainty better. Talk to others about your worries instead of carrying them alone.
- Your values as a compass. What matters to you remains solid even when the technology changes. I elaborated this thought in The Phases of Change.
The question of whether something like experience ever arises in these systems is closely related to the topic, by the way. More on this in What Is Consciousness? And Can an AI Have It?
How I accompany you
If the topic unsettles you, that is understandable and no sign of weakness. In my coaching we look together at what concretely burdens you and practise how you stay calm and able to act even in uncertainty, with breathwork, meditation, hypnosis and clear everyday steps. One-to-one coaching takes place mostly by video and costs from €200 per hour (incl. VAT). More on the Services page and under Coaching in Düsseldorf.
Sources and notes: Live Science, 31 Oct 2025, on tests by Palisade Research (to the article). Axios, 23 Sep 2026, on the Californian governor’s order (to the article). The assessment that for a simple emergency stop it is probably already too late is my personal perception and not scientifically proven.