AI Watch

Who can switch off your AI — and why it concerns you

Mr AndersonNovember 17, 2026Read time: 5 min
Who can switch off your AI — and why it concerns you

An American bill would give the government the power to block an AI model deemed dangerous, beyond a quantified threshold and with a judge's approval. For a business, the real question is not regulation: it is who can switch off the tool it has started running on, and whether it can still work without it.

Some texts get skimmed because they seem to belong to someone else. An American bill on artificial intelligence model safety falls into that category. It is happening in Washington, it concerns laboratories training models with hundreds of millions of dollars, and you, in Anglet, have a quoting tool to update before Friday.

Except one line in that text deserves a pause. The government could gain the power to stop an artificial intelligence model. Not to regulate it, not to frame it, not to hold it to account: to stop it.

And if a model can be stopped, then the question is no longer only who trains that model. The question becomes: who can switch off the tool your business has started running on.

This is not an article for or against regulation. It is an article about a threshold, a trigger, and what they imply when you are on the side of the people who use AI rather than build it.

What the text proposes, without the jargon

The creator of Claude has submitted a federal framework to American lawmakers. The idea has three tiers.

First, a stop mechanism: the government could block the release of a model deemed dangerous. Second, a procedure: the risk would be assessed by independent firms, and the decision would need a judge's approval. Third, a primacy clause: this federal framework would override some stricter state laws, so a business does not have to comply with fifty different rules depending on where it operates.

The text also sets obligations for those who build: testing the most powerful models before wide deployment, publishing the results of those tests, accepting independent evaluation, and running a genuine security programme.

Every point can be argued for months. That is not where the useful information for a business lies.

The threshold: three figures and a judge

The useful information is the threshold. A model would not be stoppable "because it is frightening" or "because a minister decided so one morning". It would take an imminent, quantified risk: more than fifty deaths, a major cyberattack, or more than a billion dollars in losses.

That is the interesting part. The trigger is not political, it is quantified. And it is checked by third parties, then validated by a judge. It is a decision architecture, not a slogan.

Why does that matter to you? Because a quantified threshold applied to a model applies mechanically to everyone who uses that model. There will be no exception for the twenty-person business producing joinery quotes with the tool. The day the model is stopped, it is stopped for everyone, and your tool stops with it.

That is the real news. Not "the United States wants to regulate AI". Rather: the world is equipping itself with a switch, and nobody has yet asked users what they would do if it were pulled.

Why it concerns you, even with no data centre in the Basque Country

I can hear it from here: "that has nothing to do with me, I am in Bidart, I do plumbing".

Let us look at the chain. Your quoting tool sends text to a service hosted elsewhere. Your review-reply agent calls a model running in a data centre whose country you do not know. Your accounting software uses an invoice-extraction feature supplied by a third party. Your website answers visitor questions thanks to a model you do not host.

None of those bricks belong to you. All of them depend on a company that depends on a legal framework you do not control, in a country that is not yours. This is not a moral problem, it is a supply-chain problem, and it is handled as such.

A construction business knows what a single supplier running dry means. The Italian tile that stops arriving, the spare part six weeks out, the subcontractor going under at the wrong moment. Businesses learned to keep a second supplier for critical items. AI has not yet had that treatment, even though it has become as structural as a materials supplier.

The scenario almost nobody tests

Here is a simple and uncomfortable exercise. Take your most used AI tool. Ask yourself what happens tomorrow morning if it stops answering.

Three cases, all already seen, with no law and no threshold involved.

The first: the access key stops working. A provider changes its terms, suspends an account, or blocks a country. You get an authentication error and nothing works. This is not hypothetical: businesses have lost access to entire services overnight for geographic or contractual reasons.

The second: the model you use is withdrawn. Providers deprecate models regularly. Your prompt, tuned over six months to a precise behaviour, ends up plugged into a model that answers differently. The result is not wrong, it is just less good, and nobody notices for weeks.

The third: the service stays online but becomes too expensive, too slow, or restricted in what you may send it. The bill triples, quotas hit, sensitive data has to leave the perimeter.

None of those three scenarios requires an American law. A text like the one from the creator of Claude simply adds a fourth case, rarer but harsher: the stop decided, framed, validated by a judge, and announced to everyone at the same time.

So the question is not "will this happen". The first three already happen, every day, on a small scale. The question is: are you in a position to keep working the next morning?

Three questions to ask before signing

I ask them systematically when I install a system in a business, and the answers are often evasive. That is normal: most providers did not build their tools for this.

First question: where is my data and how do I get it back?

If the answer is "in our cloud, and you do not need to get it back", that is an honest answer but not an acceptable one for a critical process. What matters is the export format, how often it is possible, and what remains usable once out.

Second question: what happens if I want to change provider?

A well-built system separates three things: the data, the instructions, and the model. If your instructions live in a proprietary interface you cannot copy, you are not changing provider: you are starting from scratch. Those are two very different situations, and you usually discover which one you are in at the worst possible moment.

Third question: how much of my business depends on this single brick?

A business using AI to draft LinkedIn posts does not carry the same risk as a business using it to sort incoming requests and trigger quotes. The first one stops, you pick it up by hand on a Tuesday. The second one stops, and the salesperson no longer knows what came in since Monday.

collage: portrait of a craftsman facing an industrial switch, cut-out workshop photographs, colour blocks, text printed like an old press page

What reversibility means, in practice

The word is technical, the thing is not: being reversible means you can leave without losing everything.

In practice, for a locally sized business, it comes down to five moves.

Write your instructions in a document of your own, not inside a provider's interface. A prompt library living in a versioned text file is the difference between changing tools and rewriting everything.

Take your data out regularly, and check it is readable once out. An export you have never opened is not an export, it is a promise.

Keep a second route for critical functions. Not two subscriptions for everything: a real alternative for the brick whose shutdown costs money. For many businesses that is a smaller, less capable model that holds up on simple tasks and runs somewhere else.

Document what happens when it breaks. Half a page is enough: who tells whom, what is done by hand in the meantime, which decisions are blocked. A business that has written this once no longer panics.

Re-read your contracts, on at least two points: ownership of what you produce, and the conditions for suspension. Those are rarely the clauses discussed in a sales meeting, and they are the ones that decide what comes next.

None of this requires a large budget. It requires one afternoon and some clear sightedness.

What I do on my side

Let me be direct: for a long time I had a single model provider, and I did not think about it. It worked, it was simple, the bill was clear.

What made me change was not a bill. It was an outage. One morning with no API response, automations waiting, and nothing to do but wait. Since then I have a simple rule: anything running unattended must be able to run without me and without a single provider. It cost me time at the start, and saved me much more since.

The regulation debate does not interest me much, truthfully. What interests me is that a business in Anglet, Bayonne or Saint-Jean-Pied-de-Port can use these tools without putting its operations up as collateral.

collage: two photographic control panels linked by a cable, press cut-outs, red and blue colour blocks, cream paper background

The wrong debate, and the right one

The wrong debate is the one filling the columns: for or against AI regulation. It is binary, it is distant, and it changes nothing about what you will do on Monday morning.

The right debate is more modest and more useful: from what threshold does a business become dependent on a tool it cannot replace? And what do we do so that threshold is not crossed by accident?

A framework like the one proposed to American lawmakers has an involuntary merit: it makes the implicit visible. An AI model is a service. A service has a provider. A provider has rules, and one day it may stop answering.

When that day comes, the businesses that keep working are not the ones with the best model. They are the ones that still knew how to work without it.

What to remember

The threshold in this bill — fifty deaths, a major cyberattack, a billion dollars — will probably never be reached. Good. The fact that it exists, and that it is quantified, says something about the times: AI has become infrastructure, with all the switches that implies.

Three things to do this week, without urgency and without drama:

  • List the functions in your business that would not survive twenty-four hours without AI. There are fewer than you think.
  • For each one, write in a single sentence what you do by hand in the meantime. If the sentence does not come, that is where the risk is.
  • Check that your instructions and your data can leave the tool hosting them. A real test, not a contract read.

You do not control the threshold, the judge, or American law. You control your ability to keep working when someone else decides to pull the plug.

FAQ

Frequently Asked Questions

In practice, the model provider, not the government directly. An American bill would create a stop mechanism, but the switch would remain in the hands of the company hosting the service. For a user the outcome is the same: the tool stops answering, without them being consulted or able to contest it.

Let's launch your project

WHAT IS
YOUR
PROJECT?

We prefer real conversations over boring forms. Click on a service to tell us about your ideas, our process, or your AI production needs.