S1: Welcome in San Diego. It's Jade Hindman on this episode of KPBS Midday Edition. Could AI kill us or heal us all? We explore the possibilities and the regulations required of each. This is KPBS Midday Edition. Connecting our communities through conversation. By now , maybe you've heard about recent warnings about AI and its dangers. A former researcher announced his resignation from a major AI company with a warning AI could kill humanity by the end of the decade. Those and other concerns raised by or raised those in the AI space have led to discussions about the safety of the artificial intelligence industry and growing calls for regulation. Stuart Geiger is an assistant professor at UC San Diego , where he works in the field of AI and data ethics. Professor Geiger , welcome to Midday Edition.
S2: Thank you. Happy to be here.
S1: Glad to have you here for this. Um , what do you make about these dire predictions about AI? Where do you stand? What do you know?
S2: Well , I think there are a lot of risks. There are a lot of risks in our society , even before AI. And I think that there's a lot of things that when you start to put experimental , untested technologies that we don't quite understand into very critical parts of our economy , our society , our industry , we can have issues. But I also think that anyone who's giving you some sort of quantified , you know , prediction of the extinction of humanity is maybe telling a science fiction story that is distracting us from some of the real , serious present concerns that arise with AI.
S1: I mean , when a researcher says this technology could kill all humans , what what specific chain of events are they actually even imagining?
S2: You know , it's hard to tell sometimes and often they're not very specific , but sometimes they're sort of imagining a kind of Terminator , Skynet like scenario where these things evolve to a degree that they gain some sort of consciousness or. And again , a lot of things that people who I think work day in and day out with these see some of the mistakes they can make , see some of the dangers they can make. But the apocalyptic doomsday scenarios are sort of really premised on this kind of idea of it becoming something super superhuman , as opposed to something that is more just computers making decisions.
S1: Is it even plausible?
S2: Well , it's plausible in a certain sense , in a certain sense , but I think that the other kinds of risk I'm actually far more concerned about authoritarians. I'm far more concerned about authoritarians and oligarchs sort of using AI to do sort of good old fashioned exploitation. I'm concerned about the climate risks. I'm concerned about job displacement. And I think a lot of these apocalyptic scenarios , which again , possibilities. There's always I'm a scientist. I'm never going to say something's not possible in that sort of sense , but I think that those end up distracting us from a lot of the real present harms and concerns. Yeah.
S1: Yeah. So immediately , what do you see as as a present harm that needs to be addressed?
S2: I think there's a huge present harm with a lot of our important industries and decision makers really outsourcing their brains and their thinking to a lot of these. I'm not as concerned with some sort of , you know , Terminator Skynet situation where it autonomously takes control of all of our nuclear power plants , as I am with a nuclear power plant operator who holds their phone up to the control panel and says , chat , what's going on with my nuclear power plant? You know , I think that there's a lot of things with job displacement , a lot of things with climate , and a lot of things where people are putting a lot more trust into these than they should be.
S1: Do you think that AI has the you know that the in it exists , this possibility of a better world for us? And how do you see that working?
S2: Well , I think there's an issue with when we talk about it just as quote unquote AI , because again , it's a very complicated question. What AI actually is. It's a very much a moving target. 50 years ago , spellcheck was considered AI , and by some definitions , like an old school thermostat would count under some proposed regulations as AI because , you know , you delegate some sort of decision when you set it to 72 degrees. It can sense its environment , it can sort of change its environment based on that. Um , but I think the the tech company is actually the big AI company , sort of they advantage that when we sort of put all kinds of data analysis , all kinds of modeling , all kinds of automating drudgery , and we we could have put everything that we can do with automation and computers into one bucket of AI. And then we're told to the society , we have to either take it or leave it , or are we going to regulate this thing or not? Is it good or bad? But in fact , these things , there's there's a huge difference between the kinds of AI models that require massive data centers that are burning massive amounts of energy that , you know , to have these thousand agents forms and the kind of things that ordinary people can run on their own computers to be able to summarize documents or transcribe speech or automate some of the drudgery. Hmm.
S1: Hmm. You know , these recent discussions about the dangers of AI have led to growing calls for regulation of the AI industry. And as you alluded to some saying , some of the major companies like OpenAI and anthropic are moving dangerously too fast. What are your thoughts? I mean , is it in need of more regulation , more rules?
S2: I think absolutely , this is in need of more regulation. I mean , software is one of the most unregulated industries , and if you compare it to anything else where you have people calling themselves engineers , you wouldn't be able to get away with this in aerospace or civil engineering. This is not how we build bridges. This is not how we build airplanes. This is not how we build nuclear power plants. And I do disagree with the framing of this as a race and it being framed as a slowdown. What I think that people are grasping and that people want to make comparisons to the nuclear arms race or other things like that. And I think what a lot of these engineers inside of these companies are really sensing is that something is really messed up in the culture of these organizations , of the incentives in these organizations. We actually have a recent paper out in Harvard Data Science Review , where we sort of review a lot of industrial disasters and the lessons that are that are that the AI industry can learn from it. And from Chernobyl to Challenger to the 737 Max. We do find that there's often , you know , some sort of root technical cause , but these are allowed to happen because these organizations have sort of normalized deviance , as they sometimes say. They have poor risk perception , misaligned incentives this move fast and break things kind of culture. A view of sort of no bad news is sort of allowed. Kicking the can down the road like these are , I think , a lot of the things that these engineers who are sort of calling for concerns and calling for regulations are seeing that their companies , their organizational cultures , you know , aren't really Responsibly developing these and so.
S1: Well , I mean , I know you don't have a crystal ball here , but if you could point out , like what an immediate consequence of this type of culture could produce , what would it be?
S2: Well , I think it's some of the things that we've exactly seen in some of the incidents that we've seen in the in the recent weeks and months where these companies have turned off a lot of the safeguards they have , for example , the recent hugging face hack , where they try to disclaim responsibility and they say , oh , our agents went rogue. But there's always a chain of events that leads back to humans making decisions , even if it's , you know , much farther down the road. This is sort of the what they effectively did. The metaphor I use is sort of like taking a rabid dog to a dog park off leash , and then being surprised and saying that , well , we're not we shouldn't be held responsible for what happened in that park if someone gets hurt. Hmm.
S1: Hmm. Um , there has been talk of the need for a , quote , kill switch for AI models. What is that exactly? And why do you think that's important?
S2: Yeah , I think the kill switch is one of the most basic things we can have. It's from the 1800s from , you know , from when we were building automated factory machinery. Everything has a kill switch , an emergency stop , a scram button. Every industrial industry that's critical has one of these. And these are things that. Yeah. Would these things run on computers? These things those computers are powered by electricity and water and networking. And yeah , the ability to be able to have kill switches is important. Every AI system that I try to run , I have a very strong control over being able to shut it down immediately , because I've seen when I try to use these things for various parts of my own work , I've seen where it can , you know , just go haywire and and do something that I completely didn't intend for it to do.
S1: I mean , okay , we're talking about kill switches. We're talking about all sorts of stuff here. And this technology is something that could shape the lives of everyone on the planet. Why do you think it's not regulated like other industries?
S2: Well , I think software has. This is something that's been the case for software for decades and decades. And , you know , historically , I think one of the reasons behind this is because software was often embedded in different sectors. And so to the extent that we have a lot of regulation around software , it's because it's regulation about software in healthcare or software in finance. And so it sort of takes a lot of those existing regulations. We have a lot of sector specific regulations here in the United States. And so software is sort of a meta sector. It's gone into all kinds of different sectors. And so we're starting what we're starting to see now especially is it becoming more of a sector in its own right. And so oftentimes the regulation has wanted to sort of put it very close to the application layer , where it actually sort of goes in. But they are starting to become , you know , universal , you know , industries for the automating , you know , all kinds of tasks , all kinds of labor across sectors. Hmm.
S1: Hmm. Um , President Trump has downplayed AI risks in all the calls for regulation and emphasized the importance of staying ahead of China. As you mentioned earlier , when it comes to AI. Take a listen.
S3: I say this , but we're leading China in AI. We're the most sophisticated country in the world. And frankly , I want to keep it that way because whoever wins AI wins. And we can put guardrails. We can do this and that. But I think you have a lot of negative forces that are bringing it up that shouldn't be bringing it up. And they're bringing up things that won't happen. But whoever wins with AI wins.
S1: You disagree with this framing of AI as an arms race. Talk more about that.
S2: Yeah , this this is not a race , which is the kind of thing we tell children sometimes. But you know this , it's a race in the sense that maybe we're also racing with China to. Or we should be to electrify our power grid. You know , in the sense of whichever country electrifies there or , you know , gets off , gets off of fossil fuels and onto renewables is going to have an advantage the faster we get off of fossil fuels and entrepreneur rebels , that's going to have a lot of benefits. And so , yes , in a certain sense we might be wanting to race in that sense , but the idea that this is something where the first person that gets to it , or the first country that gets to it , or the first organization that gets to it , is going to permanently win. There's all these memes about being six months away from the permanent underclass or these sort of things like this. Um , those are , I think , science fiction and the the history of technology. And a lot of the developments in here show us that when these things are developed and it's actually really hard to keep them secret. This is the case also with , you know , nuclear weapons. But at the end of the day , like this is just math. And so these things , you know , it's not the case that if one country sort of makes some sort of development , they're going to be able to sort of keep it secret enough and keep it private enough to be able to have sort of a permanent advantage. These things are a lot more incremental , as opposed to something where there is a finish line that you can sort of cross. And once we're there , we're done.
S1: Um , for those of us , um , not working in the world of AI , what advice do you have? How should we be thinking about types of tools like ChatGPT before using them?
S2: Yeah , so I think we should be very careful about these things. I think they have a lot of promise. I think again , we can automate , especially a lot of the drudgery , but especially when it comes to so-called agents , which basically what that means is things that , you know , you give them a login to your account and then it can sort of do things for you. I am very careful. And you see these these consumer products especially are begging you to connect their your inbox and your social media accounts and all these things so they can post an email and order Instacart on your behalf and things like that. I find that terrifying. I basically have a rule for myself that it's me who authorizes any decision , and I have to sort of do that. I also think it's really important to evaluate AI on a very task specific and context specific basis. And so AI is not like doesn't work like humans. It can be good at one thing and then something very close next to it. It can be absolutely terrible at it. And you have to sort of know it's , you know , some , some , some the consumer AI products are really good for research and summarization , but really bad for problem solving , you know , and vice versa. Some of them are really good for coding and terrible for other sort of things. And so I think if you don't , I tell my students , you know , that if you're in a use AI and you don't have a sophisticated understanding of what certain products and certain models are good for , bad for where they fall , where they're where they're good , then you have no business sort of using them. You're you're not literate in that. These things actually , you know , people want to use AI , I think a lot because they're sold this fantasy that we can sort of turn off our brains and it sort of takes away. We can spend a lot more free time. But I actually find that if you want to use these things well and use these things responsibly and actually get benefit out of them and be able to do , you know , larger and larger kind of tasks or things like that , you actually have to spend more of your brainpower. You have to focus. You have to be able to keep so many things in check. You have to , you know , there's so many different moving parts.
S1: Earlier in the conversation , you said this talk of AI doomsday is a distraction. Yes. One thing we should all focus on now when it comes to AI.
S2: Um , well , I think , you know , things around the climate impact , I think. I think we want to talk about existential risk to humanity. I think the climate crisis is an existential risk to humanity. It's a much more slow moving one than the AI doomsday ers have prophesized. But I think that when it comes to things like the massive push , the move fast and break things , the sort of , you know , $1 trillion in capital commitments that are sort of being made to to spin up an experimental kind of technology around this. The , the climate of this is , is , I think , you know , concerning. And then I also think it's really concerning the things that just ordinary decisions that are made in ordinary organizations , like I'm at a university at UC San Diego. I'm terrified of if someone starts using this for admissions or grading or hiring or things like this. I do a lot of work on bias and discrimination that's embedded in these models. And so I think just the day to day way that we delegate tasks needs a lot more careful thought. Hmm.
S1: Hmm. I've been speaking with Stuart Geiger. He is an assistant professor at UC San Diego in the fields of artificial intelligence and data science. Professor , thank you so very much I appreciate you. Oh.
S2: Oh. Thank you.
S1: That's our show for today. I'm your host , Jade Hindman.
S4: Thanks for tuning in to Midday Edition. Be sure to have a great day on purpose , everyone.
