The AI Innovation and Law program at the University of Texas School of Law explores the complex legal and policy questions surrounding artificial intelligence, which Kevin also covers on his own podcast, Scaling Laws.
This episode was recorded at the Ai4 conference. Listen to our other Ai4 conversation with Paypal’s Srini Venkatesan and Zscaler’s Sam Curry.
Connect with Kevin on LinkedIn and X.
TRANSCRIPT
(STACK OVERFLOW INTRO MUSIC)
Ryan: Hello and welcome to the Stack Overflow podcast, a place to talk all things software and technology. I'm your host, Ryan Donovan. We're coming live from the floor of the Ai4 conference. And today we are talking about, not the engineering side of AI, but the law. And my guest for that is Kevin Frazier, who is a professor at the University of Texas School of Law. So, welcome to the show, Kevin.
Kevin: Thanks for having me, Ryan. And to all those folks who thought that the engineering side of AI was the most fascinating part...
Ryan: (Laughs)
Kevin: I'm here to tell you that the legal and policy questions are pretty fascinating, too.
Ryan: Yeah. Yeah. Well, I mean, in a way it's social engineering, right?
Kevin: Yes, very much so.
Ryan: So, let's do a, like... A quick overview of what's the major issues with AI and the law.
Kevin: Yeah. So there are a couple different buckets. The bucket that is most salient for most Americans is data centers.
Ryan: Yeah. Mhm.
Kevin: So, even when you say AI governance, the first thing that a state lawmaker will then say is, they won't talk to you about data, they won't talk to you about how a model is going to be released and tested, so on and so forth. For them, the most salient thing that they hear from their voters is just data centers. So even though you haven't said anything about it, they'll go straight that way.
Ryan: Right.
Kevin: And when we're talking about data centers, there's permitting concerns, there's zoning concerns.
Ryan: Right.
Kevin: There's the regulation of utilities. And all of those are very weighty, very difficult legal questions that involve questions of federalism, the power allocated between the states and the federal government.
Kevin: Right.
Ryan: You get municipalities involved, you get counties involved. So that can all be gnarly. Then, if you press them and say, okay, I'm not talking about data centers, I'm talking about AI specifically. They will then not go directly to the tech. They'll pivot to something like workforce. And concerns around the loss of jobs. So, there are proposals far and wide now about how to address perceived and/or real consequences of AI being more ubiquitous in the economy. Today, we're talking on... August 7th, 2026? Gosh, time flies. Representative Casar in Texas introduced the idea of a tax on AI companies. We know that Senator Warner, for example, has been studying how best to try to measure the number of jobs that are being cut specifically because of AI at major corporations. So that's another bucket. And then you start to get into concerns related to the tech. And that's where you'll see issues with respect to deepfake content provenance. And then you can get a little bit further into the weeds, and folks will have issues around child safety and cognitive de-skilling and related questions around how AI is being used.
Ryan: Right. So for the data centers, that is a big one that's getting a lot of protests. You talk about the, sort of, permitting, land use. There's social effects too, right? There's effects on utility prices, consumer hardware prices. I think there's noise issues around it. I had this conversation with another podcaster friend of mine, about whether the people who are protesting are crazy. And he says, well, a little no, a little yes, a little no. And I tend to be a little more on the protesters' side. Just because there are social effects and, you know, not much is being done about them. What is the law doing to shift to protect and/or enable data centers?
Kevin: Yeah. So, you know, I think that if you ask anyone, hey, do you want a little bit more congestion on your roads? Do you want, perhaps, your view to be obstructed? And do you want perhaps an annoying hum? Very few people are going to say, just based off of those facts alone, yes, sign me up, I would love that, compared to the alternatives. As with most public policy issues, the question or the difficulty arises from the fact that the benefits that result are diffused. And are not immediately felt by the people who instead are saying, hey, well, it takes me 15 minutes to get to work now. It used to take me ten. I used to have this great view, now I don't. So on and so forth. So, how to measure those and how to quantify those is something that's been really hard. Because it is really easy to see when your view has been obstructed.
Ryan: Right.
Kevin: It's a lot harder to see, hey, my county now has more tax revenue than it's ever had in the past few decades. The library's open longer. That pool project is finally going to be completed. The museum hired more staff. You don't feel that immediately, but it's really important for a community. And so, I don't think anyone is necessarily crazy. I think they're responding to the information that's right in front of them. And unfortunately, in many cases, that information is not necessarily accurate. You know, with respect to water usage, for example.
Ryan: Right. Right.
Kevin: Are there bad actors? Are there bad data centers? You betcha.
Ryan: Yeah.
Kevin: But it's in the interest of hyperscalers to be as efficient as possible.
Ryan: Sure.
Kevin: And so, they're coming up with ways to use less and less water. And, you know, if you're gonna die on the hill of water usage, please tell me you're not drinking any almond milk, that you don't use almond butter, that you've given up golfing. Same goes with respect to land use. Yes, we are seeing it that these data centers are taking up more and more space.
Ryan: Yeah.
Kevin: You know what takes up significantly far more space? And I'm borrowing from Andy Moseley here, who I highly recommend following. Christmas trees.
Ryan: Oh yeah.
Kevin: Right? Way more Christmas tree farms for something that no one really needs.
Ryan: Christmas trees, and parking lots too.
Kevin: Parking lots. Right. So, the issue that I do really buy into is, number one, just, the visual effects, and the lived experience of a data center, of loud construction, of a little bit more chaos on the roads, busier businesses, which is in some ways a positive. But for...
Ryan: Yeah.
Kevin: If you're a regular at the diner down the street and you can't get a seat, that's annoying. But electricity prices.
Ryan: Yeah.
Kevin: Is a real one.
Ryan: Yeah.
Kevin: The issue is that it's not as easy as just saying, it's Google's fault, it's Meta's fault, it's Amazon's fault. Utility law is centuries old, and it's really complex about the mechanisms that are in place that attempt to diffuse the costs across rate payers.
Ryan: Right.
Kevin: Now, generally, we support that, because it makes sure that no one's bearing too high of a price.
Ryan: Right.
Kevin: But we never had hyperscaler data centers on the grid in the way we do now. And so, what may have been a good regulatory paradigm in 1920 doesn't necessarily work in 2020 and beyond. And so the question is, how do we make sure that these ratepayer protection pledges get codified in law, such that, it isn't average Joes and Janes who are paying for...
Ryan: I think a couple of years ago, we saw, sort of, the effects, benefits of those sort of shared rate payment, when Texas had a huge storm, and Texas was not part of any of the other ratepayers, and electricity went through the roof, right?
Kevin: Exactly. There's a lot of benefit to the interdependency and the redundancy of a national grid system.
Ryan: Right.
Kevin: And so, having a lot of folks who are contributing to that grid and building out new power can be a really big net positive.
Ryan: Right.
Kevin: But sharing that burden is a tricky math problem. But it's not necessarily going to be solved by pointing fingers.
Ryan: So, a lot of times these sort of, like, big corporate build-outs are sold through increased jobs and increased tax base. But there's often tax breaks given at the same time. How real are those jobs and tax base raises for data centers?
Kevin: Yeah. So, I will give you a standard law professor answer, which is wildly frustrating.
Ryan: (Laughs)
Kevin: The answer is, it depends.
Ryan: Right.
Kevin: Right? So, for savvy counties that have strong negotiating skills. For whom this isn't their first ball game, right? They have folks who used to work at Fortune 500 companies who are now county commissioners. They have gone to battle with big corporations previously. They know how to draw out the full benefits. And when they do, they are very real. So, there are parishes in Louisiana, for example, that were able to so thoroughly negotiate with these hyperscalers that their teachers are getting $50,000 bonuses.
Ryan: Wow.
Kevin: That's a big deal for a rural community that probably is struggling to recruit and retain...
Ryan: Right.
Kevin: ...The best educators, to be able to lure those people in with that big of a bonus? That's a game changer for a community.
Ryan: Yeah.
Kevin: Now, if you're less savvy...
Ryan: (Laughs)
Kevin: ...And you're not extracting all of that money out of the hyperscaler...
Ryan: Yeah.
Kevin: Things can be a little bit more dicey, right? Because you're not certain if the costs are worth the benefits. And the costs are real, as we discussed earlier.
Ryan: Yep.
Kevin: The benefits, in terms of a jobs prospect... It's important that we go to the most reliable sources of information about these kinds of jobs.
Ryan: And who is that these days?
Kevin: So... (Laughs) Unions, right?
Ryan: Aha. Yeah.
Kevin: So, who are the people who want these jobs? They're electricians. They're plumbers. They're construction workers. And you can go talk to the IBEW, the International Brotherhood of Electricity Workers.
Ryan: Yeah.
Kevin: And they'll tell you, yes, we want way more data centers.
Ryan: Yeah.
Kevin: So when people say yes, but, you know, they're short term, you only have two years or whatever, so on and so forth. The union will say, well, great. Point us to any construction job that isn't short term.
Ryan: Right.
Kevin: And so, yes, I hear the complaints that these are not jobs that last decades. But for the construction worker, or the electrician or the plumber, their whole career is stitching together job by job by job. Then the question becomes, what does it mean for the community as a whole?
Ryan: Yeah.
Kevin: And that's where you can have a pretty nuanced location by location analysis of, is this going to be a data center that requires more of an on-premise staff than others?
Ryan: Yeah.
Kevin: And then relatedly, how is that maybe going to increase what we call secondary employment? For every one data center worker you bring in, you may bring in their spouse.
Ryan: Yeah.
Kevin: Who needs a job and goes and goes and starts something.
Ryan: Right.
Kevin: That's another two seats at the diner.
Ryan: Right.
Kevin: That's another three seats at the movie theater.
Ryan: That's the money going into the community. Right?
Kevin: Exactly. Yeah. So it can be a really complex calculation.
Ryan: So, what has been the sort of legal response to this complexity around the data centers?
Kevin: So, unfortunately, the response we're seeing with increased frequency, led by Governor Hochul in New York, is a moratorium on hyperscaler development. So there is a one-year ban on hyperscalers being built in New York. Other jurisdictions have explored things like temporary pauses, or they have otherwise delayed the implementation of planned data centers. And a lot of that is grounded in justified concerns. I will say, for example, I am a strong advocate for, all of this should be occurring above the board. The fact that people are being kept in the dark about how these negotiations are proceeding...
Ryan: Right, NDAs on them. Yeah.
Kevin: ...Ludicrous. No one should have to be guessing about how and when a data center is going to come in and what the terms are going to be. Similarly, when we hear about hyperscalers finding loopholes, whether explicitly or implicitly, that allow them to run through regulatory processes in a way that maybe doesn't align with the community's expectations.
Ryan: Yeah, yeah. (Laughs)
Kevin: That may be legal, but it doesn't leave people with a good taste. And so it's important that...
Ryan: Right.
Kevin: ...From a sort of, how is the strength of our democracy doing? Making sure that this is transparent and done in a reasoned format. Now, I'm not saying we should, you know, lean into NIMBYism and throw up barriers just for the sake of barriers.
Ryan: Right.
Kevin: But making sure people feel like they're a part of the process is important. But the issue with respect to these moratoriums and bans. When we think about data centers, they need good water.
Ryan: Mhm.
Kevin: They need it to be relatively cool. They need open land, and they need good infrastructure. Well, New York, especially upstate New York...
Ryan: Yeah.
Kevin: ...Sounds pretty dreamy. Especially if you're concerned about the environmental concerns we've mentioned previously.
Ryan: Yeah.
Kevin: But where are these data centers going to get built now? I don't know, Nevada sounds pretty good.
Ryan: (Laughs)
Kevin: If you're not going to let me build in New York. So, this is a fundamental question of, not to go full law professor, but, what does it mean?
Ryan: What's your job again? (Laughs)
Kevin: Yeah, yeah. What does it mean to be a union? And that involves sharing the costs of national infrastructure. And so when New Yorkers say, you can't build it here, it's not as though New York is suddenly going to see administrative access to AI.
Ryan: Right.
Kevin: Their data is still going to bounce around on data centers throughout the country. Just not in New York. And so, is that fair? Is that free riding fair? And the traditional answer has been no. And that's my concern.
Ryan: Well, so, I think the data center fight has been kind of a proxy fight for the AI slopification of a lot of things, right? For most consumers, the way they see AI is, worse results. Your feeds are full of garbage, low-quality garbage, and things don't quite work right anymore. And they have no recourse other than this, you know, land use proxy battle, right?
Kevin: Yeah. No, that is exactly the point. Which is... Again, it's that question of, concentrated costs and diffuse benefits.
Ryan: Right.
Kevin: We know the costs. You can see, your LinkedIn feed sucks.
Ryan: (Laughs)
Kevin: Because everyone is now just maximizing their content. We have all those stories. And it's important to note that most Americans, number one, haven't received meaningful AI training.
Ryan: Right.
Kevin: And number two, aren't accessing the latest and greatest tools.
Ryan: Right.
Kevin: So, only 2.2% of American households have a paid subscription to an AI tool. And even when they have a paid subscription, they're not paying for the nth degree. They're getting the bare minimum.
Ryan: Right.
Kevin: And so, when it comes to discovering the benefits of, let's say, scheduling a task every week that's going to scan your email, identify, you know, the kids carpooling. Identify the best grocery store timing, and so on and so forth.
Ryan: Yeah.
Kevin: Americans haven't seen that. And they also haven't heard about, in many cases, the real benefits that AI is deriving, for example, in healthcare. There's a company in Texas called Zauron Labs that is working with radiologists. And I say with, because there are more radiologists using these tools to improve the accuracy of their scans. The founder, Kyle Clark, he'll tell you, that when... He's a radiologist. When radiologists have their own images scanned, they always ask for a second look from their friend.
Ryan: Right.
Kevin: Most Americans don't get that second look.
Ryan: Right.
Kevin: And that's where AI can, number one, improve accuracy, eventually drive down costs of care.
Ryan: But the... The AI in these cases is usually a more traditional machine learning model. It's not the generative LLMs that...
Kevin: Sure.
Ryan: Yeah.
Kevin: Yeah.
Ryan: So what... Move to the next one on your list.
Kevin: Okay.
Ryan: The workforce concerns. Now, we've obviously heard a lot of layoffs being blamed on AI. So, what is the legal response to AI agents being sort of treated as employees?
Kevin: Yeah. So, legally, there's a bit of a question as to how we're going to measure these workforce development impacts. We know that some states, for example, like New York, have amended what are referred to as their WARN act statutes. So there were laws enacted generally in the 1970s in response to massive layoffs at factories.
Ryan: Right.
Kevin: And the idea was, hey, if you're about to lay off 30,000 people in this small town...
Ryan: Right.
Kevin: We need to know about it.
Ryan: Right.
Kevin: You have to notify state and local authorities so we can prepare.
Ryan: They're going to apply for unemployment.
Kevin: Exactly. And let's bring in the cavalry. Let's make sure we've got the workforce assistance in place. There's efforts to amend that for AI, which is to say, hey, if you think that AI is going to cause this sort of layoff, you need to notify us. Well, there's a lot of skepticism among economists and public policy folks that there might be a benefit right now to blaming your job losses on AI, right? Because you're not saying, hey, maybe we weren't great at management.
Ryan: Right.
Kevin: Hey, maybe we missed this wave for that product.
Ryan: I mean, a lot of people are saying overhiring from Covid, or the end of zero-interest financing, right?
Kevin: You could also name, I don't know, several wars...
Ryan: (Laughs)
Kevin: Inflation generally, domestic political cat... There's a lot that we can point to. So it's relatively convenient to have a scapegoat.
Ryan: Yeah.
Kevin: The question for me is the question we faced in prior eras, which is to say, how do we make sure that the US is the spot where the jobs of the future get created first? So the Bureau of Labor and Statistics came out with a report about a month ago identifying that they think there's going to be 3.1% increase in jobs by 2034.
Ryan: Okay.
Kevin: That's pretty good. That's not nothing. But it's specific jobs that are going to see the largest surge in that employment.
Ryan: Yeah. So what are those jobs?
Kevin: So, it's data scientists.
Ryan: Oh, okay.
Kevin: It's sales engineers. In some cases, it's folks in the medical profession. It's the care industries, the things that are innately human. And then, of course, it's going to be jobs we haven't labeled yet.
Ryan: Yeah.
Kevin: So the question is, number one, how do we help create those jobs? And number two, for the disruptive period - and this is really important - that disruption is going to occur. And it's going to hit some communities and some industries harder. Talk to anyone in gaming, they're not feeling great right now.
Ryan: Right.
Kevin: Right? We know that there have been a lot of layoffs...
Ryan: Yeah. A lot of creative industries, yeah.
Kevin: ...Because of AI. So. How do we help those folks pivot to meaningful new opportunities? And that's where I think lawmakers need to be vastly more creative.
Ryan: Right.
Kevin: We need to start thinking about the equivalent of the CCC in the New Deal. How do we help people find meaningful short-term work?
Ryan: Right.
Kevin: Where they're learning new skills, helping their community? That's something we can do.
Ryan: Because one of the other jobs that's come up has been sort of a, you know, training, data labeling, like, you do your job in front of the AI and it learns it, sort of like the service industry job for white-collar people.
Kevin: Right.
Ryan: And it seems like there's a path to less satisfying jobs there.
Kevin: And that's where we have to avoid the prior instances of, workforce training efforts have generally based their policy review, whether we're going to continue to fund that workforce training program, so on and so forth, based off of the job employment within six months.
Ryan: Yeah.
Kevin: Well, hey, if I'm a workforce agency and I'm being assessed on if I got you a job within six months...
Ryan: Yeah.
Kevin: I'm just gonna find the quickest, fastest thing, even if it ends up being morally demoralizing.
Ryan: Right, right.
Kevin: And, you know, it's just menial work. We need to create the right incentives such that we're finding the programs that aren't just finding people work, but finding them meaningful, enduring work. And that's a heck of a lot harder.
Ryan: Right. And then in terms of, you know, we talked about protecting the existing employees, or maybe laid-off employees. Is there a future where, you know, agents are treated as actual employees? Like, there's L and I attached to your cloud installation?
Kevin: You know, I would suspect that we're not going to see AI suddenly qualify as employees under the law, such that we see them regulated by the Fair Labor Standards Act, for example. That may come about in several decades. The Overton window right now...
Ryan: Right.
Kevin: ...Is not shifting that far. We're still questioning how we treat independent contractors. And trying to apply laws that really...
Ryan: (Laughs) One step at a time, right?
Kevin: Right. So getting all the way out there to what is it to appropriately treat an agent in the context of employment...
Ryan: Yeah.
Kevin: ...I think it's going to be a minute. I think it merits attention. But the law... We've got so many other issues that we still have to figure out.
Ryan: Sure.
Kevin: Yeah.
Ryan: And speaking of the next issue on the list, kind of the data provenance. Which has a lot of, you know, copyright, IP, and then also liability on what's produced. So there's been a lot of movement on this, and it... I'm not sure, but it sort of seems like there's a settling of issues here. Because it seemed like there was an argument with, like, fair use versus copyright protection. Right?
Kevin: So I would say that there has been a decreased amount of attention to data as fair use on the input stage, right? For training. We have seen that there's a sort of recognition that, this is how fair use was intended to operate, right? When you think about the purpose of the IP clause in the Constitution.
Ryan: Right.
Kevin: It is to spread and generate knowledge.
Ryan: Right.
Kevin: Well, AI, for its many flaws, if it's good at one thing, it's really good at diffusing information, making expertise more generally available. Of course, are there pitfalls and does it hallucinate from time to time? Yes.
Ryan: Yeah.
Kevin: So do lawyers.
Ryan: (Laughs)
Kevin: But that is a really powerful tool, and I think the general assumption is twofold. Number one, fair use for input will additionally... It will eventually be recognized as the law of the land. Whether that reaches the Supreme Court is an open question. But if enough circuits, the intermediate courts, reach that opinion, that'll be settled law for all intents and purposes. The related question, too, is, seeing how more and more companies are leaning into licensing agreements.
Ryan: Yeah.
Kevin: So the folks who would be most likely to sue...
Ryan: Right.
Kevin: ...Are now finding, hey, if you're paying us...
Ryan: Right. Yeah.
Kevin: ...We're fine with that. And so the issue has been slightly mooted.
Ryan: For some of the fair use concerns, I've sort of always wondered, like, how is not, like, if you take, ingest this book, and add it as part of your training data, how is that not a derivative work?
Kevin: So, derivative work is usually going to have to be much more narrowly confined to its relationship to the original work itself. When you are looking at the fair use prong here, which is transformation.
Ryan: Yeah.
Kevin: Right? Are you seeing a transformation of that data? And, you and I and your listeners could get super technical about embedding...
Ryan: Yeah.
Kevin: And seeing how the data itself is being transformed.
Ryan: Yeah.
Kevin: But then at a more general sense, or abstract sense, we can say that the data is being used in such different myriad ways. Than saying, for example, reproduce the entirety of Harry Potter, but with a slight new chapter or a new character. That might sound like a derivative work, right?
Ryan: Right. It's not Pride and Prejudice and zombies, right?
Kevin: Exactly, exactly. So, you know, there's ultimately also a discussion of, what sort of society we want to build, in terms of, how do we compensate and reward creatives? And that's a really important conversation. I will say, when I talk to folks in creative sectors, I generally say, there's one rule for our conversation. You can't say copyright. And the reason why is if you talk to a musician... A normal musician. I'm not talking Taylor Swift, I'm not talking to Beyonce.
Ryan: Right.
Kevin: If you go talk to a okay artist, or author.
Ryan: Yeah.
Kevin: Financially. They're excellent, of course.
Ryan: Sure, sure.
Kevin: Financially, they may not have even registered their work for copyright.
Ryan: Oh, okay.
Kevin: Which, to enforce a copyright, you have to have registered your work. If they have registered their work, they're not in a position to then go and enforce that copyright if there's any sort of infringement, because lawyers are really freaking expensive.
Ryan: Yeah.
Kevin: And so copyright isn't the thing that is traditionally resulting in renumeration and a livelihood for the creatives we really care about. And so the public policy question is a lot different once we say, okay, we're not going to hyperfixate on copyright. How do we create new schemes? Like, for example, I was talking with one individual at an AI lab about, what if we had the equivalent of Trump accounts - these are the new saving accounts for younger Americans - for artists. Such that the government could pay into this savings account or this, uh, wealth fund. Patrons.
Ryan: The NEA for all, or something.
Kevin: You know, NEA for all. Exactly. Patrons. Apparently we're going to see tons of them as soon as these labs IPO.
Ryan: Yeah.
Kevin: They could pay into these accounts to help support and sustain artists. So, I'm not saying that's the best way forward, but I want us to be a heck of a lot more, pardon the pun, creative...
Ryan: Yeah.
Kevin: ...About how we think about supporting the arts, than just saying copyright, copyright, copyright.
Ryan: Yeah. And I think for copyright, a lot of folks may not know about the registration. Like, there's ideas of poor man's copyright, where you just sort of mail it, or you just put copyright, you know, 2026 on something, and you're like, I'm good, right?
Kevin: Right. Not quite. Yeah.
Ryan: Yeah. So there's also sort of issues on the output side. The Frontier Labs have accused some of the Chinese companies of distilling. Taking the output and using that to train their models. What's the legal maneuvering around that?
Kevin: So, distilling is generally viewed right now by lawyers as a matter of terms of service violations. Right? And if you are Anthropic, for example, which, you've said you cannot distill on our models, and then you distill on the model, that is a contractual dispute.
Ryan: Right.
Kevin: For which there is law that can be responsive to that. We know that distilling has been used by most of the major labs.
Ryan: Sure.
Kevin: Go ask Elon, and he'll say that's some... That's a tactic they're using. We know that distilling is important for the open source community domestically, in terms of being able to develop and expand access to these tools. So really, the dispute right now is seen as a TOS issue. Congress, though, is weighing if there should it be additional steps taken to limit the extent to which we see distillation. This has big geopolitical questions, right?
Ryan: Sure.
Kevin: Where, we may support and really want to champion open source domestically.
Ryan: Right.
Kevin: Because it's great for the research community, it's great for red teaming, it's great for safety and academics. But when we start to think about that in a, hey, China's now reaping the benefits of our hard work...
Ryan: Right.
Kevin: That gets a lot sketchier. So I suspect that we will see a more formal legal response to distillation in a foreign context. How that gets enforced...
Ryan: (Laughs)
Kevin: ...Is going to be very difficult.
Ryan: Yeah.
Kevin: Yes. That's the current lay of the land.
Ryan: Alright. New Geneva Convention for AI. Maybe.
Kevin: We will see. And we know how well the Geneva Convention's...
Ryan: (Laughs) Yeah. So, when you touch on the last one, you talked about the sort of effects of using AI. Effects on kids, de-skilling,, AI psychosis. What sort of law is springing up around that?
Kevin: So, the most important law, and this is existing law that folks should know about, is unfair and deceptive acts or practices. So, these are often referred to as viewed UDAPs. And they exist at every state. So, there's a UDAP statute that basically says, look, if you say one thing about your product or your service, and then it doesn't behave in that way, or if it behaves in a way that is, just substantively wrong or icky or unfair...
Ryan: Right.
Kevin: ...That is a cause of action that usually the attorney general in your state can enforce against the company. So, for example, in Texas, Attorney General Paxton went after a company called Pieces. Uh, that was alleging, a incredibly high rate of accuracy for an AI tool that was used in a medical context. We didn't need a new law.
Ryan: Right.
Kevin: We didn't need some specific bespoke AI thing that says, you can't say this about your rate of hallucination.
Ryan: Yeah, yeah.
Kevin: The UDAP was the lever that led to a settlement between the company and the state. With respect to child safety, there is a strong argument to be made that attorneys general are underenforcing their UDAPs with respect to how kids are using these tools. Because we are seeing that a lot of these companies are issuing press releases and blog posts that say, we've done this to protect kids. We've done this to prevent teens from using our tools. And existing law can be used to hold companies accountable. Now, my concern is that some folks, and I'm going to be a little, uh, blunt here...
Ryan: Uh huh.
Kevin: ...Are profiteering off of the interests and the attention you can gain from being the advocate for kids.
Ryan: Right.
Kevin: And they're proposing things like...
Ryan: I mean, every politician loves to say it's, think of the children, right?
Kevin: ...Think of the children. Think of the kids. Let's think about something like a banner. That pops up 90 minutes into a user session with AI. That's being bandied about as a really smart move.
Ryan: (Laughs)
Kevin: To help kids use AI less.
Ryan: Sure.
Kevin: And my concern is that, the so-called champion of kids are going to say, look, we did the thing.
Ryan: Right.
Kevin: We passed the bill. Well, go talk to any teen and they'll tell you, hey, when I see that banner, I'm excited about it. Oh, look! Look how much time I've spent on this tool doing this thing.
Ryan: Yeah. I mean, that's sort of like legal theater, right? Like...
Kevin: Exactly. And so what I want to see is a more robust conversation. With therapists, with psychologists. Who, right now aren't pervasive enough to help kids with true mental health issues.
Ryan: Yep. And you're talking real therapists. (Laughs) Not AI therapists.
Kevin: Not AI therapists. But how can we design these tools and enforce performance-based regulation?
Ryan: Right.
Kevin: Right? So, performance-based. What are our true concerns? Is your concern that a child is leaving a conversation, or a teen is leaving a conversation, less happy or less optimistic than when they started? Well, that may be a real thing, but guess what? We can measure that.
Ryan: Right.
Kevin: And we can look at data such that in the aggregate we can say, hey, is your AI tool statistically leading to less optimistic children or less optimistic teens over this period of time? If so, you need to show new interventions that you're imposing, and we're going to pay specific scrutiny to you. This is the difference between the state saying, this is the specific intervention you have to impose.
Ryan: Right.
Kevin: And the state saying, hey, be creative.
Ryan: Right...
Kevin: Find what works best...
Ryan: ...Let's do some resources out there.
Kevin: ...Yes. Exactly. Let's study this. Let's bring in experts and see what interventions actually work, and celebrate and spread those, rather than just hoping that a banner, for example, works as intended.
Ryan: Yeah. And I think it's interesting you talk about, like, existing interventions. Like, I wonder about, you know, is a child allowed to agree to a term of service, right? Are these things... Can you prosecute the AI companies if they act as lawyers, or doctors, or give advice that's otherwise restricted, right?
Kevin: Right. Well, but the concern there is... So there was a bill, for example, in New York, that was going to make companies liable, or prohibit companies from releasing tools that were substantively... Or their responses were providing substantive answers in several professions: the legal profession, medical, therapy, psychology. But then things like veterinarians...
Ryan: Okay.
Kevin: ...Were subject to this. Plumbers, I think, were somehow included.
Ryan: Oh, okay.
Kevin: And the issue is that, the status quo is not adequate.
Ryan: Right.
Kevin: We have too few people who can find access to attorneys. We have too few people who can find reliable mental health support.
Ryan: Yeah.
Kevin: So I'm not advocating for these tools to step in fully.
Ryan: Right.
Kevin: But the question isn't, okay, if we deny them, a user access to an AI tool that is occupying, or in some ways providing information in those domains, it's not as though you say no, and then suddenly they can find a lawyer or a therapist that's affordable and available.
Ryan: Sure.
Kevin: So we need to find a more happy medium and a more nuanced solution. But again, that's harder.
Ryan: Yeah. And I think there's a... I said, you know, not to get too, too, too, big and broad on this one...
Kevin: Yeah.
Ryan: ...But there's a part of this that is the growing loneliness of America, the isolation of everybody. And some of the AI companies may benefit from that. And I think that the... That, you know, you can argue that causes, right? The whole bowling alone follow-ups. But there is, you know, a sort of isolation that's come on a policy level.
Kevin: And this is where I think we need to be more creative about how AI can be used. So, for example, I've called for a community in the loop. Where imagine if you, Ryan... I apologize. Where do you live?
Ryan: In New York.
Kevin: New York. So there's your local neighborhood council.
Ryan: Yeah.
Kevin: In theory, has tons of information about new events that are happening, new businesses that got started, things that may be of interest to you. Let's imagine that there were more streamlined connections between your council, your small businesses, your nonprofits, and the AI you're using.
Ryan: Right.
Kevin: Such that you, Ryan, could turn on that, let's just call it community in the loop.
Ryan: Yeah, yeah.
Kevin: And you're saying, oh, gosh, you know, I'm tired again. ChatGPT.
Ryan: Yeah.
Kevin: Claude, I'm just not feeling too great about things, so on and so forth. Hey, Ryan, you know, you told me last month that you were really concerned about this new intersection.
Ryan: Right.
Kevin: And the traffic signal there. There's a hearing on this next week.
Ryan: Right.
Kevin: Do you want me to register you to go testify, or to go attend? These are the sorts of subtle, small interventions that I don't want this to go into paternalism.
Ryan: Yeah.
Kevin: But it's an example of just, surfacing information in a way that...
Ryan: Government information is pretty hard to find a lot of times. Exactly. Yeah.
Kevin: And even for new nonprofits.
Ryan: Yeah.
Kevin: If you are... You know, let's say, Ryan, you care a lot about the East Side Parkway.
Ryan: Yeah.
Kevin: And there's a...
Ryan: Absolutely.
Kevin: There we go. There's a cool new roller skating event that's happening on the East Side Parkway. You gotta be there. That information getting to you on X or TikTok or whatever may be really hard.
Ryan: Yeah.
Kevin: But if we can have a better information flow that's hyperlocal, that's specific about people with your shared interests, I think there's some chances that your agent can now go find that information and surface it for you. So, that's the sort of interventions that I want to explore, and that I don't think we're talking enough about.
Ryan: Yeah.
Kevin: I would say that, there needs to be vastly more conversation about AI literacy and access to AI. We're talking in Vegas, surrounded by people who can tell you about how many agents they've spun up in the last five days.
Ryan: Yeah. The most literate of AI people.
Kevin: The most literate.
Ryan: Yeah.
Kevin: And most Americans aren't at that level. And most Americans haven't had a chance to really mess around with these tools and see how easy it is to start a new business, to launch a new website, to explore new topics. And so, I'm really encouraged by things like Claude core; I'm really encouraged by free programming that Google's doing on AI literacy. But that has to be way more ubiquitous.
Ryan: Yeah.
Kevin: So that more people can be a part of this revolution, and not just feel like it's something that's happening to them. And so that we can see this be something that furthers our republican form of government, lowercase-r.
Ryan: Yes.
Kevin: Rather than something that concentrates power.
(MUSIC PLAYS)
Ryan: Well, I thank you for listening, everyone. I have been Ryan Donovan. I edit the blog, host the podcast here at Stack Overflow. If you have questions, concerns, topics to cover, you want to virtually yell at me, email me at podcast@stackoverflow.com. And if you want to connect with me directly, you can find me on LinkedIn.
Kevin: And, Ryan, thanks for having me. I'm Kevin Frazier, the director of the AI Innovation and Law Program at the University of Texas School of Law. You can find me eating breakfast tacos in Austin, or on LinkedIn or X, or I have my own podcast that's not as good as this one.
Ryan: (Laughs)
Kevin: Called Scaling Laws.
Ryan: Flattery will get you everywhere, Kevin. Thanks for listening, everyone, and we'll talk to you next time.
(STACK OVERFLOW OUTRO MUSIC)