Rogue AI chatbot sparks kill switch debate

All right, welcome back into Seattle News Today. Topping our headlines, I want to talk about an AI kill switch. Conversations about a so-called kill switch heating up after a popular chatbot went rogue, prompting new action from the White House. Anthropic actually briefed the Trump administration after its AI agent, Claude, tried to solve a homicide and worked to bypass paywalls. The White House will now require mandatory security reporting for AI firms, while lawmakers are pushing for stricter guardrails. Joining us now from out in the field via Zoom, we have Bill Howell with the University of Washington. For those who don't know, Bill is the co-director of the AI Clinic, also an associate professor at the UW iSchool or the Information School. Bill, first and foremost, I just want to say thanks for joining us today. Of course, happy to be here. Now, we'll just jump right into this conversation about a kill switch. what what does that actually mean and and who's kind of proposing this conversation of or this idea for AI well the idea itself is sensible and has been discussed in various capacities for a while but most recently a few days ago Satya Nadella offered this as a possible solution to all the news stories that we've seen about AI agents doing things that they weren't necessarily supposed to do in the context of the experiments that were being run you know the idea itself is is makes sense and is sensible. But, you know, the source of where it's coming from, Microsoft, you know, they're a platform company, and there's no real way to have a filter from within the systems themselves. It would be something about the platform in which they're deployed. So really, he's kind of talking about deployed models as opposed to any possible model you might imagine in any context. And, you know, since Microsoft is a platform company, it's not misaligned with their goals of of providing a hosting environment for for models so i think while the idea is is right and we should be exploring it it's it's not uh the the panacea that perhaps was was suggested uh you know obviously AI is one of those hot button issues right now i mean i'm looking at just the past 24 hours headlines that i was kind of browsing before our conversation i saw One headline saying AI invents a murder tip, sends it to police as a company takes heat over delay. Another saying anthropic warns self-replicating AI could spread across household devices. Another centers pushing Trump administration to tackle rising AI cybersecurity threats. You know, it's one of those things where I think we're learning more and more that AI just kind of is a big part of our lives. But at the same time, we get so many of these headlines. And I wonder, as an outsider that's still trying to wrap my head around all of this, as we talk about kill switches, as we talk about how much of our economy is tied up in AI, as somebody who is so close to this, that's doing research in this space, I wonder if you have the same concerns that you're hearing echoed in media or in the general public, social media, et cetera. You know, I do have these concerns. I think the most important thing is to separate the concerns from the business interests that are that are behind some of them. Right. So in every one of these cases, including this this suggestion from from Satya, you know, saying that we have the magic, scary robot and only we can see it. But trust us, it's going to be terrible. There is significant truth to that. And that is a concern that I'm aware of and we work on. But we have to have a little bit more transparency about what's actually going on. So having sort of a small number of thousand people in the world are the only ones that are privy to certain kinds of capabilities is not the way to really find the right regulation for working on this. But, you know, am I concerned about the capability? The capabilities are clearly there. This has been exposed in the research. That's a little bit more neutral than the, you know, billion dollar profit motives behind the companies. and so are our cyber security threats imminent yeah i think they are and we're starting to see some of that just from experiments that that ran a little bit amuck or they didn't control that carefully for a lot of those experiments that are coming out you know you mentioned the headlines some of these are pretty well designed experiments that then the science reporting picks up and maybe sensationalizes a little bit other ones are you know pr moves perhaps to try to increase value before an IPO. Other ones are very solid research experiments that we need to all be aware of. And so we're getting kind of a mix of that out there in the news stories, and it can be difficult to make sense of that. So it's clear to me that the capabilities are there. There will be some cyber security threats. There may be some bigger threats around like bioweapons and so on. So this is a really big deal. But to really consider how we respond to this as a society with the federal government, with state governments, with the companies themselves, there needs to be more people looking at the evidence information than just the companies themselves. It's funny because as we have this conversation, I've tried to explain to people, I do my best to wrap my head around some of this stuff. And I'm now remembering when I was watching Congress discuss the internet when I was a kid. And I remember just thinking like, oh my gosh, these guys just don't get it. And so I'm sure there's like another generation that is kind of understanding this and so like i try to never put my foot in my mouth and that's why i ask you like what what do we make of these headlines so i i guess the extension of that is since we kind of kicked this off talking about kill switches i wonder where your headspace is at right now because you kind of explain what may or may not work and why these conversations are are so important but i i wonder you know in an era where we see all these ai execs or all these tech execs meeting with the president talking about ai like what are you hoping is being discussed in that room like What do you think we should be paying attention to as we move forward in this quote-unquote new world? You know, I categorically don't have a lot of faith in what's going on in those particular discussions because the wrong people are in the room, right? There's people with a vested interest that have billions of dollars. The most valuable companies in the history of the planet are involved in these discussions. There needs to be people that have less of a vested interest in the outcome involved in there as well. And one place to look, of course, is academia for people that study this, of which there are a great many. You know, the other issue is that the federal government has kind of pushed on one particular side of this, and the alternative opinions there are not adequately represented in that room. So I'd like to see those discussions with the evidence in play opened up where congressional committees invite experts from a variety of different areas to discuss what the right regulation we can do. uh some you know regulation is a little bit difficult of course because there are other countries in play so we don't actually have uh necessarily the right avenue to globally regulate this although i actually am somewhat optimistic that china and some other and other places are actually in the same kind of boat like they're actually wondering what to do about this as well so i think there is an opportunity to have those conversations at a global level but we're not it's not clear we're really having that we have kind of the full steam ahead view and then we have the you know doom and fear kind of view and we don't really have that nuanced middle uh that we that we need to have in a transparent way so that's the first thing i'd like to see is just to have that discussion and then what i think would come out of that if we did follow that protocol is uh you know we not not to rely solely on product liability laws which is what's been proposed you know in the same way we don't rely on product liability laws to keep planes in the sky you know you you don't just wait for a plane to crash and then just sue the airline and just hope it'll fix itself that way, you actually get out in front of those issues and put some rules in place. We would need that as well in this area in that the technology is at least as, if not far more dangerous than most of the things we've come up with. Specifically, I think some platform level solutions like what was proposed can be a part of it, but you can't guarantee that all the models, the open weight models, the smaller models, the locally deployed models aren't necessarily going to follow those rules. So you do need some strong language and policies with teeth in them to prevent people from releasing things that give bad actors capability to, you know, take down some company's website without putting any effort in at all. So Bill, I guess the question is, is it too far along to see this? Because it sounds like you're hopeful that we can get places. There's the pessimistic side of me, which probably is from the journalism side of me that says like, hey, if we're looking at another country and we're worried that they're going to advance, so then we naturally have this knee-jerk reaction. We need to advance as well. And it kind of just feels like it's just going to keep speeding up. That's the pessimism that I hear from so many people, especially online and reactions. And we know that's not always the real world. So it sounds like you were kind of offering a little bit of hope as to reining things in before it gets too dangerous. Is there any particular reason that you're feeling that hope or am I reading too much into that, you know, I don't know if I'd embrace the word optimism because I'm, you know, I'm worried like a lot of people and perhaps more so because I am so uh close to the technology but I, you know, I do think that some of the most obvious ways that things can go badly everyone on the planet doesn't want to happen and so that's the basis for a conversation to to be had and perhaps some agreement to be made so you Treating this as if it's purely an arms race, to use that phrase, is not the only way to look at this. It can be about how do we prevent these particular outcomes that we can deduce and get everybody together and work to agree on that. I think we're in for a bumpy road, though. I think there will be capabilities in the hands of folks who, with remarkably low effort, can do relatively high harm, especially in terms of cybersecurity, but also in maybe some other areas. And so that's going to be coming down the pipe no matter what, and that's going to be something that we're going to have to endure. The other big one, of course, is what's going to happen with automation in terms of job loss. And so for that, you know, same thing, get people in the room to think through those solutions. But I personally don't see any other way around some form of universal basic income coming down the pipe eventually. As we automate information work jobs, of which there are a great many, there needs to be some way to account for that. We can't have all the wealth in the world being concentrated in the hands of, you know, four or five companies that have, you know, hundreds of employees each, which we may be headed towards if we're not careful. and i think that is one of the the scary things that we're kind of almost glossing over these days because of all the other things that have come out in headlines so i should ask you know how do you how do you land in this space i gotta imagine that there was something about ai that interests you because a lot of it does seem negative these days but but something had to draw you in here like what what is the the goal or the hope or what drew you into this side of academia uh the glass half full side of this i guess well yeah yeah again i'm being painted as as optimistic it's not it depends on the room a little bit of whether i sound like the other the pessimist or the optimist you know my background was more in in data big data and data science back in the day you know 15 years ago and so on when that discussion was being had and so we were interested in the systems that could help bring more evidence-based uh decision making to both organizations as well as the general public as well as you know higher education and anybody else that they need to work on things we we didn't like the hunch based you know economy that we had sort of been running for the for the previous you know 50 years or something uh but you know once you sort of get all that data together in one place one thing you can do with that now that these this other technology has come about is answer broad-based questions and take broad-based actions without a whole lot of effort. And on the one hand, as a techno-optimist kind of person, and a lot of the folks in computer science are trained in that regard as engineers, they see this as extremely beneficial. You know, we can automate all the things that caused inefficiencies and pain, and we can collect all the right data, we can answer questions based on the data, and so on. But of course, as the capabilities got generalized, what these things can put together is write any software you want, you know, with a handful of prompts. And that includes software that you wouldn't necessarily want written, such as cybersecurity attacks and other kinds of harmful things. So I think we're in this kind of phase where we've trained people to do what they did, which was push the capabilities and advance them as far as they could, as fast as they could, and then recognize that, wait, we succeeded at that so well, but now we need to walk it back. So your question was where my interest was, you know, shifting from kind of the platform level data management into these higher level capabilities was just following where the real problems were. We've solved, we've addressed some of the issues around how to do large scale data processing, and we have not addressed the issues on how to align these capabilities with human interests. And so that kind of drew me back into this area. And I will say, didn't mean to paint you as optimist or pessimist, but it does sound like it kind of depends on what room you're in, Right. As you say, I think part of that is because I never want to feel like I'm the alarmist because it is such a heavy topic. But at the same time, like it does sound like more and more people are kind of sounding alarms in this space. So a point taken. I'd like to just kind of leave the end of these interviews open ended, because frankly, like I don't want to say this too many times in a week. But we were talking to an astrophysicist from UW the other day. So I just refer to myself as a dum-dum. when we have these conversations i'm not the expert you are so so i'll just leave this open and ask you like what should we be paying attention to what do you think people should be engaging in when it comes to ai because i don't want people to just be hurling around unintelligent stuff that they don't know one way or the other uh what should we be doing as we move forward i i think uh you know advocating for that transparency is the thing that i keep that i keep repeating, that this relatively small number of people that are privy to the information that we need to be able to figure out where to go from here, that needs to be opened up and made more public. You alluded to it. The other caution I'd have is try to avoid falling into either extreme where it's very tempting and frankly would be historically justified in saying, ah, people are always being alarmist about these things. It always turns out to be less than they think. you know we didn't we didn't higher education didn't disappear with massively open online courses and you know the internet didn't do x y and z and so on uh however this does feel a little bit different to me because of the capabilities that that we have you know you have kind of a natural language interface into advanced computing which lowers the bar for anyone to get into it and we were relying on relatively rare skill sets being the only people that could do this stuff and now if anyone can do it then we have we have we have to figure out a different way to to to manage that um and so with it without you know but but also not necessarily giving up all hope where we say oh well you know okay i guess we're i guess we're done i'll you know quit my job and move out to the woods and and it's all over you know i don't i don't i don't know if that's quite justified yet either i think there are conversations that we can have i think we can uh find a common ground in what we're trying to avoid and what we're trying to achieve um and address some of some of the some of the questions that are coming up so i think most imminently it's a question of cybersecurity. That's happening right now today. And we have a concern there. And I think slightly, well, only slightly, very slightly further out is the questions about the overall economy and what we're going to do if we end up automating away a lot of these jobs. Oh, yeah, just that, you know, there's so many sides of this. I could probably take up hours of your time, but I will let the viewers know we caught you right after class today. So we appreciate getting the time. Dr. Bill Howell, co-director of the AI Clinic, associate professor at the UWI school. Hey, thank you so much for walking us through this and maybe getting us a little bit more understanding of what is honestly being talked about nonstop these days. No problem. Thank you so much.