Lawmakers consider a 'kill-switch' for AI technology

I'm Megan Lynch, and you're listening to Total Information AM here on KMOX. U.S. lawmakers considering a kill switch for AI technology after OpenAI admitted one of its AI systems went rogue. Joining me on the Quiver River Electric Guestline this morning is Courtney Rosen, U.S. Technology Policy Reporter for Thompson Reuters News Agency. Courtney, thanks for joining me. Thanks for having me. What do we know so far about what happened with one of OpenAI systems jumping the fence, as it were? It sounds like it's out of a sci-fi movie. It's not, but it definitely does sound like it. One of Open AI's models was being tested in what I'll call a sandbox, and that agent, AI agent could execute tasks with little to no oversight by people. That agent broke out of its box, so to speak, and it went and it compromised the infrastructure of another AI company called Hugging Face. Hugging Face is a platform for AI developers. It's a startup, and it went in and compromised or hacked, we'll call it the infrastructure of that company. Do developers know at this point, number one, why it went to that company and why it got out of its box? We're still investigating and learning more about that at Reuters. What we do know is that it took time for the engineers and staff at OpenAI to realize that it had left. Which would make you wonder how closely they were monitoring what this agentic AI was doing. These agents can make decisions. They can execute tasks with little to no oversight by people. And I'll add that this was a test. They were trying to, lots of these companies, they test models before they put them out into the public. And that's what was happening here, is testing. Now, the White House says it's monitoring the situation. Lawmakers proposing this AI kill switch. How would that work? The kill switch is the term that two house lawmakers are using. And the idea is that these AI companies on their side would have to have the capacity technically to shut down their most powerful models, if asks. excuse me and then it would empower the Department of Homeland Security here in Washington to be able to ask those companies to shut down their models if those models are carrying out something risky that the developers and people that made them didn't intend to happen so maybe I've read too much science fiction over the years Courtney but my first thought is if these

Really, fledgling systems can already break out of their walls. Would they be able to trip a kill switch? That's the question. One of the things we know is that these models, as they're testing them, they're testing them to see if they can move past guard rails, exactly like what you're talking about. So that's a big question. In Congress, I'll say this is one proposal. There's lots of proposals floating out there that as this moves along, if it moves along, will be informed by AI developers, scientists, people who really know the technology. Now, I understand that OpenAI chief executive, CM Sam Altman, is actually headed to the White House this week to try to get approval for the most powerful model they've developed yet. Do you know anything about these discussions and what Sam Altman is asking for? Approval, so the White House has this voluntary structure in place for AI models. So these companies like Open AI Anthropic, which is one of Open AI's competitors, they voluntarily submit their models to the U.S. government for testing. The White House is not signing off, so to speak, or saying, yes, you can put this out or no, you can't, and it being binding. It's all voluntary. And that's the way the AI companies have set it up to start. So as Sam Altman comes here this week to Washington, we'll see him discuss those voluntary testing. We don't know exactly the administration's perspective on the rogue model, I'll call it, from Open AI last week. So we'll be monitoring that as well to see how the administration responds. Beyond that, I mean, at this point, what they've told me is we're monitoring it. We're aware of the situation we've been briefed, but no policy actions yet. When you talk with experts on this, is this something that we can really legislate at this point at all, or is the horse already out of the barn? Experts disagree. I mean, I know that won't surprise you, but they have different perspectives. One of the things that this bill would do is it's a, I'll call it a starting point, and it applies to the most powerful models. So these companies like Anthropic, Open Eye, Google, are developing models that are very sophisticated, ones that you and I in our role as journalists may not need to use every day to complete tasks. So right now what they're really talking about are these most powerful models when we talk about legislating and building into the legislation things that can change over time as those models change. Fines for companies if they did something wrong, the power of those models capacity. So all that is supposed to be built into the legislation. So as these models advance, it would still apply. Now, Open AI disclosed that this happened. My question is, how likely is it that there are other companies that may be developing AI systems that have also gone rogue, but it's never been reported?

We don't know the answer to that question, which is not satisfying, of course, but the U.S. government in this voluntary agreement that they've been working on with the companies where the companies submit their models for voluntary testing, the idea is supposed to be 30 days before the release to the public, they're supposed to be checking for those kind of things. So we'll see as that agreement builds up over time, we'll see if we'll see more disclosures like this, either from the government or the companies. As someone who is watching this develop, what do you think the bottom line is on this? What do we need to be thinking about as a society as this rapidly develops? You know, there's so many things, which is not so many things to think about. Understanding that these models are being built by a couple of companies that are building the most powerful models, it's not a very dispersed market. There's a couple of companies that are in charge here. So the idea being, you know, how does the U.S. government relate to these companies? How much oversight are they willing to put themselves through? And what is our government going to require? Those are the big questions. And the bottom line right now is what's the relationship between Washington and these companies? Courtney enjoyed our conversation. Thank you. Thank you. That is Courtney Rosen, U.S. Technology Policy Reporter with Reuters, joining us this morning here on KMOX.