92 Percent of Germans Want an Off Switch for AI. Do You Have One?

Ask a German what they want from AI, and apparently the answer is a big red button.

On Tuesday, Bitkom published a survey of 1,006 people in Germany aged 16 and over. 92 percent of them want a Not-Aus-Schalter, an emergency off switch, for all AI systems. Support was above 90 percent in every age group. That almost never happens in a survey about anything.

The rest of the numbers are more mixed, and more human. 85 percent want stricter rules for especially capable AI. 56 percent still see AI as more of a chance than a risk for humanity. And when you ask about AI that plans, decides and acts on its own, the country splits almost in half: 49 percent mostly worried, 47 percent mostly confident. The young are calmer. 58 percent of people aged 16 to 29 feel confident about these systems, while 62 percent of those over 65 feel worried.

Bitkom president Ralf Wintergerst put the trust problem plainly: theoretical alarmism, but also the many security incidents, have destroyed a lot of trust. His other line is the one I keep thinking about: the more independently AI systems act, the more important it becomes how humans keep control.

We already know how to build this

Here is what made me smile. People are asking for a kill switch like it is some new, exotic idea. For us engineers it is just a normal workday.

We have feature flags. We have circuit breakers. We have rollback buttons, rate limits and that one config value everybody knows to flip when production is on fire. Every decent payment system I have worked on had a way to stop sending money within minutes, without a deploy.

And then I look at how many teams ship AI features, and especially agents, and I am less amused. The agent can open tickets, send emails, change records, call other services. Ask the team "how do we stop it right now?" and you get a pause. Then someone says "well, we could revoke the API key". Then someone else asks who has access to that key. Then everyone looks at the person who is on vacation.

That is not a kill switch. That is a hope.

If stopping your system needs a meeting, you don't have an off switch. You have a negotiation.

Why this matters more than it looks

I don't read the 92 percent as fear. I read it as a very reasonable request: show me that someone can stop this.

That is exactly what customers, works councils and your own management will ask you, sooner rather than later. In Germany they will ask it with a checklist. And "the model is very safe" is not an answer. "Here is the switch, here is who can press it, and we tested it last month" is an answer.

There is also a career angle hiding here. The engineer who can explain, in two calm sentences, how an AI feature is stopped and what happens to the data in flight, is the one who gets invited to the serious meetings. Not because they are the best at prompts. Because they make people feel safe saying yes.

What I would do this week

1. Find the off switch for your AI features. Not in theory. Write down where it is, how long it takes, and who is allowed to use it.
2. Make it boring. A feature flag or a config value that turns the agent off, or back into "suggest only" mode, without a deploy. Boring is good here.
3. Test it once. Flip it in staging, or in production during a quiet hour. A switch you never pressed is a guess.
4. Give agents a smaller blast radius. Limit what one run can touch: how many emails, how many records, how much money. Then the off switch is the second line of defence, not the only one.
5. Practise the two sentence answer. If your manager, a customer or your neighbour asks "can you stop it?", you should be able to say yes, and explain how, without a slide.

People are not asking us to stop building. They are asking us to build the part we already know how to build. That sounds like a fair deal to me.

Sources: Bitkom: Deutsche wünschen sich mehr KI-Sicherheit, IT Finanzmagazin: Not-Aus-Schalter für alle KI-Systeme?, HASEPOST: Deutsche fordern Not-Aus für KI-Systeme