Lawmakers introduce 'AI Kill Switch Act': How would it work?

I want to start with the AI kill switch  act. Okay. My understanding reading from. the act is it would require developers of the  most powerful AI systems. This is what you. write to maintain the technical capability to  throttle, suspend or shut them down. Now I'm, you know, about a world where we're  hearing about hugging face and swarms. of agents and different civilizations that  don't even know their agents. And you know, makes me think, how in the world could there  be a kill switch? How exactly would it work?

Uh thank you Erin for your question. Let me first  say that today on September 11th we will never. forget the victims of the horrific terrorist  attacks and we will never forget the bravery. of our first responders. Also today is Rashashana  and so to the Jewish community. Happy Rashashana. Now in answer to your question, society has had  a simple solution to machines that can do great. good and cause great harm. It's a kill switch. So,  heavy machinery has kill switches. Highspeed rail,

jet skis have kill switches, and even your Apple  iPhone, you can erase it remotely if it's stolen. And so, we're saying the AI industry should not  be exempt. That's why it has bipartisan support. People on both sides of the aisle realize this  is a necessary thing. It should be the first. principle of AI safety that humans must be able  to turn off AI systems and AI agents. and we. let the AI companies figure out how to do that.  They have the smartest people working at these.

frontier labs. Okay. So I I I'm still trying to  understand only because you know some of the most. terrifying things we hear whether it's open AI or  claude or chat GPT whatever is when they say oh. uh we didn't even know what the agents were doing  right and it would seem like well the agents will. find out there's a um AI will find out there's  a kill switch and they'll go turn it off. I mean. the British government had rejected a similar kill  switch idea. They said the UK cannot simply turn. AI off. Um and and there was a open AI researcher  who spoke to the BBC and said that if you switch.

off access to an AI model in emergency, it  will do little to protect you. Um what how. do you think through all of those things? Oh  yes. So you would have to be able to turn it. off safely and there are cases where you can't  turn off immediately if it's doing something. that is critical for infrastructure or let's say a  hospital or something else. But the goal is humans.

have to be able turn off AI systems or agents that  go rogue. And if the response is we can't do that, then look, we might as well just pack up our bags  and go to some island in the South Pacific and. wait for the apocalypse because we have to got to  believe that AI models and agents can be turned. off by humans. You know, I'm just thinking when  you made that comment about an island of the South. Pacific, some of the very people who are involved  in these AI companies who are running them, they. went and built bunkers in places like New Zealand  years ago. And I just wonder if maybe it's because.

they saw where this was going long before any of  of us did. To that end, Jeffrey Irving, a former. Open AI employee who Time magazine once named  one of the most influential people in AI, just. posted something um which hit my feed, so you can  tell what I click on. It says, "I think we have. approximately 50% chance of all dying as a result  of super intelligence, mostly due to actions in. the next few to 10 years." You're on the AI  subcommittee. Do you believe that's hyperbole?

Uh, I do. Now, one reason Congressman Nathaniel  Moran and I introduced a bipartisan AI kill switch. bill is we also want to put this requirement into  the AI frontier lab company so they actually work. on kill switch technical proposals and we want  to make sure that they're not so focused on. just building the most amazing model they can  without thinking about how we might be able. to turn it off. And I am uh very confident  that with the smartest people in the world,

they can start injecting in abilities to turn off  these AI agents and models. Congressman Moran, when I hear the words kill switch, I picture a  giant red stop button. But AI lives on computers, on the internet, and in the cloud. So what  does pulling the plug actually look like? Well, what it looks like under our bill is  the ability to throttle it back, to pause it, to slow it down, or in uh emergency situations,  actually stop it. And the way you can really think.

about this is look, if we want to go down the  highway of innovation quickly, just like a car, you need to have a brake installed. Sometimes  you're going to have to use that brake a little. bit and tap it on. Sometimes you might need  to pull the car over. That's exactly what uh. Congressman Lou's bill and my my bill does is  say to companies, these frontier companies, the large ones, you must have the ability  to throttle it back, pause it, slow it down, cut off user access, or stop it altogether  in an emergency. This bill, I understand,

only covers what are called closedweight models,  systems like what Open AI or Enthropic have, but it doesn't cover AI models that could be  running on my personal home server, for example, or open weight models. Uh, how do you introduce  an AI kill switch bill that doesn't cover all. forms of AI that potentially that potentially  could go rogue and become very dangerous? Well, thank you, Wol, for that question. So, the  closed models are the most advanced models right.

now, and I'm honored to partner on this bill  with Congressoman Nathaniel Moran. And our bill. is based on a very simple principle. Humans must  be in control of AI, not the other way around. And. at a minimum, that means humans must be able to  shut down AI models and agents. Now, what you're. talking about is open- source models or models  that exist on your phone. right now they are. not as advanced and so they don't pose the same  catastrophic risks as the closed models do and.

that's why you saw three AI CEOs Elon Musk Durio  Amday and Sam Alman saying we need to take a step. back and slow down development as put in guard  rails and you just noted Congressman Louu all the. warnings from AI leader leaders but Congressman  Moran President Trump responded to all of this. saying it's really a conspiracy do you think it's  a conspiracy. Well, I think we all recognize that.

there are likely uh some serious problems out  there, but there are solutions out there. And. I'm again, I think Representative Lou hit the  nail on the head. This is a this is a bipartisan. effort to address some of the some of the really  uh difficult issues that might might be out there. But I think really what President Trump was trying  to convey is whatever we do, we can't stifle. innovation. We cannot let the Chinese get ahead of  us in this in this race. But in order to move fast. down the innovation uh highway, we have to have  the ability for humans to always be in control and. to be able to throttle it back, pause it, cut off  that user access or stop it all together in the.

kill switch bill is one that actually would  do that. Would you like President Trump um. Congressman to be more forthcoming then if that's  really what he's trying to say that there needs. to be more human control over AI? Because so far  he hasn't said that. He just has said he wants to. win the race against China. that's what's most  important here and that these warnings from AI. leaders are a conspiracy. Do you think he needs  to be sounding the alarm more about the potential. threats here? I think he's identified the priority  threat and that is uh losing uh this technology.

innovation battle to China. Uh whatever the risks  are with AI, the risks get a whole lot worse if. China gets ahead of us. When we talk about that  race to recursive self-improvement, uh we have got. to always stay ahead of China. Right now, we have  the leverage. We have the ability to to influence. uh China's uh AI stack. We hold the technology  advantage. They're dependent upon us for their. development. And so, right now is the time to  really put the screws to China to help them.

understand they need to uh they need to be careful  as they proceed with AI development as well. So, I think the threat of China having a greater  AI ability is certainly the priority of threat. Congressman Lou, what's your reaction to what  we're hearing from President Trump repeatedly. over the past few days about AI uh and the  potential threat? He's calling it a hoax. Well, let me say I agree with Congressman Moran  about the threat and the reason we're getting.

bipartisan support for our bill, including Senator  John Kennedy who's going to introduce a version. of a kill switch on the Senate side is because  our kill switch by definition does not slow down. innovation. It only applies to models that have  already been built. So, here's an analogy. We. don't tell car manufacturers to not make cars. The  federal government literally buys a car after the. car has been made. We crash test it and if there's  a problem, we tell the manufacturer to fix it and.

if it's already been deployed on the roads, we  can recall that car, tell the manufacturer fix it, and then they re reput the car back out on the  road. That's an analogy. So, our bill does nothing. to stop innovation. So, just to follow up on that,  Congressman Lou, your bill says the government can. only force a shutdown if AI threatens catastrophic  harm. what exactly counts as catastrophic? So the way it's defined in the bill is if it kills  people over causes significant economic damage.

And again the analogy to the auto industry I  think is uh very similar. Uh we don't order. a recall on a car if for for example uh the seat  cushion doesn't look right. We order a recall on. a car if the brakes don't work. And so our bill  is very similar. If there's a significant problem. that could pose a catastrophic risk, we ask the  AI companies to slow it down or shut it down and.

fix it or if they don't do it, then we give the  federal government a process to do so. But oh, go ahead. I just want to follow up. Who would be  in charge of alerting the uh government of that. catastrophic harm, threat, or risk as you call it?  Would it be on the AI companies to raise the flag? Well, yeah. We we want to we want to make sure  that those companies have the responsibility to. uh notify DHS when there is an incident when  there is a an issue that needs to be addressed. That transparency is very important. Uh there's a  lot of us that have uh incident reporting uh bills.

that are out there and this one also includes  that requirement in there. We want to make sure. that they are reporting that. Once they report  it then there is uh there is then the decision. from DHS. Okay, do we need to take some action  or have they taken the action that's necessary? And these companies, we've seen a lot of companies  already act very responsibly over the summer when. they did not have the uh the legal mandate to  disclose. They've been disclosing that. They're. trying to address this from the private sector  commercial uh perspective. We want to see more of.

that corporate responsibility. To the degree that  they are responsible corporate citizens, I think. uh government should be very careful about uh  regulations that would stifle future innovation. Just really quickly, what is the appetite on  Capitol Hill right now to pass this bill? Do. you sense that there is the kind of urgency that  is needed to get this bill through or do do you. just see a partisan divide right now? Congressman  Louu? Uh so we're very pleased that this bill.

is bipartisan and again Republican Senator John  Kennedy is introducing a version of a kill switch. bill on the Senate side and just in the last  few days you saw uh Republican Congresswoman. Representative Luna come out in support of a  kill switch concept and on the Democratic side. Pete Buddha Judge said the same thing. So we're  building uh coalition on both sides of the aisle.