Crafting A Viable ‘AI Kill Switch’ Law Presents Lawmakers With Enormous Technical And Legal Challenges

AI Kill Switch Keeps Coming Up

Interest in having AI kill switches has been around since the beginning of AI. I’m sure you’ve read fictional tales of AI that diabolically aim to enslave humans or wipe us from the planet. In fact, during the real-world mania last year about artificial general intelligence (AGI) and artificial superintelligence (ASI), everyone was stressed out about the probability of doom, known as p(doom), and the existential risks of AI; see my in-depth coverage at the link here.

One viewpoint is that a large-scale AI might become dangerously malevolent, perhaps entering sentience or reaching superhuman intelligence, and humanity will only survive by having a means of turning off the AI. Maybe that’s a bit sci-fi oriented. A more down-to-earth view is that AI could make an error or encounter a software bug and go awry. The AI might inadvertently imperil humans. Thus, having a stop button or kill switch is prudent, just like having one for an automated assembly line at a factory.

The trouble with an AI kill switch is that implementing one is a lot harder than it might seem.

Technical Challenges Of AI Kill Switches

On the surface, all that needs to be done is stop the AI from being operational. If an AI was running on your smartphone and you wanted to stop it, the actions appear to be straightforward. You might press a button that is supposed to interrupt whatever the smartphone is currently doing. Or you might simply turn off the smartphone. These are easy-peasy solutions.

Scale up AI to the popular generative AI and large language models (LLMs) that we all use these days, such as ChatGPT, GPT-5, Gemini, Grok, Claude, Copilot, etc. Those AIs are massive in size and widely distributed. They are typically running on a multitude of computer servers. The servers are globally spread to ensure resiliency in case any specific data center goes down.

The gist is that the AI makers tend to devise their AI to be hardened or resilient and not be taken out by any anticipated natural disaster or other system failure. There is an irony that the very same features and functions that aim to keep AI running smoothly are also part of the problem in finding a simple means of adopting an AI kill switch.

Legal Questions Abound

In addition to the technical challenges, there are a plethora of legal questions that must be considered regarding an AI Kill Switch capability.

First, should AI makers have the option of instituting an AI kill switch, or should a law be passed that requires the adoption of AI kill switches? You might say that making the crafting of a kill switch a legal requirement is overkill of sorts (pun!). Just let AI makers decide what is best. An alternative perspective is that AI makers could fail to implement a kill switch, and we might all suffer when AI goes rogue. We must legally force the AI makers to institute a kill switch and protect society accordingly.

Second, who will get to decide that an AI kill switch should be activated? One approach is to allow an AI maker to make that weighty decision. No way, comes the retort. An AI maker might be hesitant to switch off their own AI. They are willing to risk our safety. Thus, the government should make that decision. Whoa, comes the response, if the government can go around switching off AI, you are handing the keys of the kingdom to Big Brother.

Third, what will an AI kill switch actually do or accomplish? Some would say that it forces the AI to do a reboot. But a reboot implies the AI will immediately be brought back online, which could be disastrous if the AI continues to do bad things. Another angle is that the AI is switched off. It is no longer operational. A worry is that an AI might be clever enough that it opts to switch itself back on. Maybe the kill switch should also obliterate or delete the AI.

Intertwining Of Technical And Legal

These are highly complex legal questions that need to be addressed, and for which there are corresponding technical challenges associated with whatever legal approach is chosen. You see, even if a law is passed that says an AI kill switch must do this or that, AI makers might not be able to shape or build the AI to act in those stipulated ways. The legal requirements must be technologically implementable.

At the same time, there should not be an impulse by lawmakers to water down any legal requirements merely due to handwringing that AI makers might have a tough time adopting the required stipulations. An AI kill switch law that has no teeth is likely a performative law versus a substantive law. Some lawmakers might want to look like they’re doing gallant things; meanwhile, they are letting AI makers off the hook and presumably undercutting public safety.

If Congress is going to pass an AI kill switch law, the law needs to contain specifics. The same goes for U.S. states that opt to enact a state-level AI kill switch law. Any law that is open-ended and vague will not achieve whatever “kill switch” goals are in mind and could worsen the situation rather than be helpful.

Recently Drafted Bill

Here’s why AI kill switches are suddenly getting a resurgence of interest.

You might have heard or read about a recent situation involving OpenAI that made the headlines and spurred calls for the legal adoption of AI kill switch laws. In brief, OpenAI was testing its latest AI; it seemingly escaped its sandbox and broke into Hugging Face to try to get secrets to achieve a testing goal that was presented to the AI (for my analysis of the circumstances, see the link here). This provoked widespread concern. Many urged policymakers and lawmakers to wake up and anticipate ways to deal with AI situations that can go badly on a massive scale.

This top news story was previously topped by Anthropic having devised its latest AI known as Mythos that was chockful of ready-to-go cyberhacks; see my coverage at the link here. The concerns were that if Anthropic released the AI to public usage, it would be immensely easy for anyone and everyone to tell the AI to break into the computers of businesses, non-profits, consumers, and even governmental systems.

On July 23, 2026, in response to the emerging wave of interest in AI kill switches, Congressional interest also rose, and two members of the U.S. House of Representatives quickly rolled out a draft bill. Representative Ted Lieu (D-Calif.) and Representative Nathaniel Moran (R-Texas) made available their draft law entitled the AI Kill Switch Act. It is not yet a submitted bill, and we will have to wait and see how things go with it. I will keep you posted, so stay tuned.

Unpacking The Drafted AI Kill Switch Act

To give you a flavor for how challenging it can be to compose a robust law on AI kill switches, I will briefly walk you through some of the elements of the draft AI Kill Switch Act. This will be instructive and illuminate the tough tradeoffs in how to write such a law. As noted, an AI kill switch law ought not to be loosey-goosey, nor should it be overly narrow, and it should aim to be reasonably implementable. That’s a lot to ask for.

First, let’s see what an AI kill switch is supposed to achieve, which, according to the draft, is this:

• “Maintain a technical capability to carry out the following actions:”
• “(i) Stop inference of a covered technology of such covered entity,”
• “(ii) Terminate user access to such technology,”
• “(iii) Suspend access to such technology with respect to an account, user, or use pattern identified by such covered entity or the Secretary as posing a risk of any of the following: (I) A covered incident, (II) A violation of law or the terms of service of such technology,”
• “(iv) Shut down such technology.”

In this case, the AI kill switch is supposed to stop the AI from performing inference, terminate user access, and shut down the AI. That is a laudable goal. There are potential loopholes.

For example, by using the terminology of “inference,” there is potential ambiguity associated with that word. The common implication is that the AI is essentially reasoning about something. Suppose the AI is merely using a component that reaches out to grab data from other systems, doing so without active inferencing directly involved; does that not come into the purview? The AI component might be wreaking havoc and yet have no direct bearing on any inferencing at that moment in time.

Which AI Is Impacted

One head-scratching legal question about any law on AI is what type of AI is considered bound by the law. Not all AI consists of LLMs. There are plenty of other types of AI. Also, there are large-scale AIs and other AIs that are tiny. Does the size of the AI matter, or should all AI of any size be encompassed by a given AI law? An AI law might be an overreach, or might be an underreach, depending on what the scope of AI is for that law.

The draft bill says this:

• “Covered Technology. Except as otherwise provided in this section, the term ‘covered technology’ means an artificial intelligence system developed utilizing a quantity of computing power the cost of which would exceed $100,000,000 at the prevailing market price of cloud computing in the United States, as determined by the Secretary.”

The draft bill offers that if the cost of the computing that led to developing the AI was more than $100 million, the AI is within the scope of the AI kill switch requirements. You might be thinking that makes sense since the biggest danger is from the biggest AIs. Well, that’s not necessarily the case. A tiny AI that goes viral could readily be dangerous and spread widely across the globe due to the interconnectedness of the Internet.

Some would say that this is a loophole. It means that AI that isn’t at that size will be able to freely avoid having an AI kill switch.

Who Activates The AI Kill Switch

I earlier mentioned that a knotty issue is choosing who can activate an AI kill switch. Usually, an AI maker is allowed the prerogative to push the button if they wish to do so, while they must push the button if the government tells them to do so.

Here is what the draft says:

• “Emergency Authority. If the Secretary, acting through the Director and in consultation with the Secretary of Commerce and the Director of National Intelligence, determines that a covered incident has occurred, the Secretary may order the covered entity at issue to take action proportionate to the nature and immediacy of such incident, which may include any of the actions described in subsection (b)(1)(A).”

This indicates that the Homeland Security Secretary would have the emergency authority to order that an AI kill switch be activated. Of course, this would only apply to those entities that are supposed to have an AI kill switch, which I noted above is the large-sized AIs.

Do you think that an AI maker can defy such an order? You might be of a mind that no AI maker should be able to override a legally permitted order from the government. We cannot leave it up to an AI maker to decide whether they agree or disagree with such an order. Others might insist that an AI maker should have a legal recourse if they believe that the order is wrong.

Appealing An AI Kill Switch Order

Here is what the draft says about appealing an AI kill switch order:

• “Appeal. Not later than 48 hours after an order under paragraph (1), the covered entity subject to such order may petition the Secretary for reconsideration of such order, but such petition does not stay such order.”
• “Determination. Not later than five days after a covered entity petitions pursuant to subparagraph (A), the Secretary, acting through the Director, shall make a determination with respect to such petition, but if the Secretary fails to so make such determination, such failure is deemed to be a determination in the negative.”
• “Judicial Review. A covered entity for which there is an order under paragraph (1) may request review of such order in the United States Court of Appeals for the District of Columbia Circuit on petition filed not later than 60 days after such order.”

Note the emphasis that even if the AI maker appeals the order, they are still required to proceed with fulfilling the order. They cannot use the appeals process to delay the action of invoking the AI kill switch. This seems to make sense, since otherwise an AI maker could raise an appeal and the AI would not be halted until sometime later when the appeal was finally decided. During that intervening time, presumably the AI is still wreaking havoc, and the AI maker is avoiding applying the kill switch.

Will The AI Kill Switch Work

There are many more twists and turns when it comes to AI kill switches. I’ve got one for you to noodle on. Suppose an AI maker comes under a law about AI kill switches and says that they have indeed implemented an AI kill switch. Perhaps the AI maker tests the kill switch and claims that it works.

A moment arises when the government decides that the AI kill switch needs to be engaged. An order is issued to the AI maker. Assume the order is lawful. The AI maker is presumably legally compelled to press the AI kill switch.

Maybe the AI maker tells the government that they did so, but in fact they chose not to activate the kill switch. Would the government know that the AI maker had defied the order? There are more such scenarios. Suppose the AI maker does activate the kill switch, but it is poorly made and doesn’t do what it is required to do. Would the government know that the AI kill switch was half-baked?

Here’s what the draft bill says:

• “Audit. Upon a confirmation under paragraph (2)(C), the Secretary, acting through the Director, shall, through audit, telemetry, on-site inspection, or other forensic review, verify compliance with the order that prompted such confirmation.”

The provision stipulates that an audit can be undertaken to verify that the order was abided by and that the result was what the legal requirement stipulates. The good news is that this would seemingly keep an AI maker on their toes, knowing that they will get audited. The bad news is that an audit takes time, and the horse might already be out of the barn. By the time the audit is performed, and if the audit shows that the AI maker defied the order or only did so half-baked, the damage or destruction by the AI might have already taken place.

The World We Are In

There is little doubt that we are increasingly at risk as AI becomes ubiquitous. Whether AI goes rogue in some far-fetched sci-fi manner or does so in everyday fashion by error or devious hacking, the crux is that there needs to be some viable means of stopping the AI. I’ve pointed out above that the technological and legal specifications to do this are hard to pull together and can end up with loopholes or gotchas.

You could claim that the mere act of talking about AI kill switches might be beneficial. It raises public awareness. The odds are that many people probably already assume that all AIs have an embedded kill switch (nope, that’s false). There is likely a further belief that the AI kill switch works flawlessly (again, false). A nuance that might not be obvious is that if an AI is being used by millions upon millions of people, and they are highly dependent on the AI, there are dangers or risks associated with summarily switching off the AI. Doing so could leave those people in quite a lurch.

A final thought for now. The famous philosopher and economist, John Stuart Mill, made this pointed remark: “A person may cause evil to others not only by his actions but by his inaction, and in either case he is justly accountable to them for the injury.” By discussing AI kill switches and by pursuing the crafting of laws thereupon, we are hopefully raising awareness among the public, policymakers, lawmakers, AI makers, and the like, regarding thinking clearly and overtly about mitigating AI harms while aiming to maximize AI benefits.