Episode 21
AI in the US Elections, and the Laws We Don't Have (Yet)
with Nate Persily, Stanford Law School
Show Notes
A week after the 2026 midterms, we dig into what AI did and didn't do in the election. In Nate's view, AI mattered more as an issue than ever before, especially through the fight over data centers, but the feared wave of vote-swinging deepfakes never came. We talk about what regulation looks like now that AI isn't a partisan issue but still faces a presidential veto, and why, as Nate puts it, “if you go into the OpenAI cafeteria, the food is more regulated than the models that they produce.” He explains why PATA still matters but isn't enough (“it's not enough to just essentially fight the last war”), and what transparency could look like when chatbot conversations happen in private.
We also get into the open-model problem and how to get defenders access to frontier AI, who's liable when AI defames someone or an agent goes rogue, whether we need an FDA for AI, and why “there's pretty much nothing that a young lawyer can do that an AI cannot do better at this point.”
In This Episode
- AI for, in and through politics: what the midterms told us
- Data centers as the outlet for people's AI anxiety
- Where AI regulation is headed: states, attorneys general and a consensus among the labs
- China, open-weight models, and a government-sanctioned pathway for security defenders
- PATA, the “last war,” and new ideas for AI transparency, such as letting users donate their chats to research
- Section 230, AI defamation and who's liable for AI agents
- An FDA for AI, and the government's talent gap
- How AI is transforming law firms, legal training and the courts
Chapter Timestamps
- 00:00Introduction
- 1:48AI as an election issue, especially through data centers
- 3:43Campaign AI, political ads, and whether deepfakes changed votes
- 5:13Everyday AI use and the challenge of drawing regulatory boundaries
- 7:24Bipartisan support, federal barriers, and state-level regulation
- 11:16Industry governance, safety concerns, and data-center opposition
- 14:32China, open models, and access for safety researchers
- 18:31PATA and the case for independent research access
- 23:24What AI transparency should measure
- 30:16Section 230, AI-generated content, and liability for harm
- 33:49AI agents, transactions, and responsibility when something goes wrong
- 39:05Possible federal AI oversight and the need for public expertise
- 41:22AI's effects on law firms, legal education, and junior lawyers
- 45:14Court backlogs, judicial use of AI, and closing reflections
Transcript
There may be transcription errors: we apologize for those in advance.
Opening: Nate Persily — What we didn't see, though, is the thing that people were most afraid of, I think, which is the use of AI to deceive voters to vote in a particular way. Undoubtedly, that happened. There were attack ads and false narratives online, and we could list them. But in terms of an impact on the electorate, I don't think we saw much of that. That's not to say that there weren't millions of examples of AI-generated imagery related to politics. It's just that it didn't tend to move votes.
Rob: On this episode of Won't Fix, I'm joined by Nate Persily, the James B. McClatchy Professor of Law at Stanford Law School. Nate works where democracy and technology meet. He's one of the country's leading election-law scholars and co-author of The Law of Democracy, a leading casebook in the field. He also co-directs the Stanford Law AI Initiative, is a faculty affiliate at Stanford's Institute for Human-Centered AI, and leads the American Political Science Association's AI Task Force. He's well known as the architect of the Platform Accountability and Transparency Act, or PATA, the bipartisan proposal to give independent researchers access to data from social media platforms. A week after the 2026 midterms, we dig into what AI did and didn't do in the election. In Nate's view, AI mattered more as an issue than ever before, especially through the fight over data centers, but the feared wave of vote-swinging deepfakes never came. We talk about what regulation looks like now that AI isn't a partisan issue but still faces a presidential veto, and why, as Nate puts it, "if you go into the OpenAI cafeteria, the food is more regulated than the models that they produce." I think you're going to enjoy this conversation, so please stick around. And as always, please send us your feedback. Thanks. Hi, Professor Persily. How are you?
Nate: Good. How are you doing? Please call me Nate.
Rob: Oh, of course. Happy to do that. So we met a few years ago. I think I gave a talk at your law school class. Was that right?
Nate: That's right.
Rob: It was awesome to have the chance to do that, and it's been good to keep in touch over the years. So we just had a pretty momentous midterm election, and it seems that AI played quite a role there. Is that right? How did you see it? The way I think about it, there's AI being used as a campaign tool, AI as an issue that voters were thinking about, and also just money from these companies influencing these races. How did it play out in your mind?
Nate: Well, that's right. I think you can think about this election in several different domains. There's AI for politics, there's AI in politics, there's AI through politics. And in each, you see AI playing a role. Let's start with AI as a political issue in the campaign itself. Whether you look at data centers or campaign promises made by politicians, AI played a more prominent role in this election than in any other election in U.S. history. In some ways, it's surprising how little a role it played in, say, the 2024 election. But just in the most concrete way, 40 out of 41 initiatives restricting data centers passed at the local level. So that gives you a sense of the anxiety people have about the AI transition, as well as specific concerns about data centers. But it's not just that. AI also played a significant role in the way people were thinking about economic anxiety and the issues in the campaigns. A lot of it got filtered through the data center debate in places like Texas and Ohio. But there's also a more general sense, whether you're tying it to future employment prospects or talking about inequality or the power of billionaires. There is an AI cachet to all of that. Then there's the AI-as-a-tool question, which I think you were gesturing toward: the use of AI in campaigns. The Wesleyan Media Project did a very interesting study that found well over 150 ads that had been run with AI use, and most of them didn't have disclosures that they were using AI, even though in about half the states or more you're now required to have those disclosures. So we're seeing AI used in creative expression, just as we would in every other form of entertainment or advertising, and we saw it in this election as well. What we didn't see, though, is the thing that people were most afraid of, I think, which is the use of AI to deceive voters to vote in a particular way. Undoubtedly, that happened. There were attack ads and false narratives online, and we could list them. But in terms of an impact on the electorate or on vote choice, I don't think we saw much of that. That's not to say that there weren't millions of examples of AI-generated imagery related to politics. It's just that it didn't tend to move votes.
Rob: I'm in Austin, Texas, so a lot of money was spent to show me all kinds of ads. The thing I noticed the most, maybe because I don't watch as much network TV, was the mailers that were clearly AI-generated. They didn't even look like James Talarico, just a bunch of attack ads printed out and mailed. So I find AI showing up in places I don't always expect. It's not just deepfakes. It's also just really bad versions of other things.
Nate: Which is to say, you're seeing it in all aspects of life. You're seeing it in corporate communications. It seems like half the LinkedIn messages I receive have some kind of AI component. So as goes life, so goes politics. The same tools that industry is using, or that individuals are using, or that my students are using, are now going to be used by campaigns and all the people who work in them.
Rob: One of the things I've remarked on before is that people discount, or don't quite understand, especially if they're not frequent users of AI, just how much it can automate and let things be processed or produced at scale. It's like when broadband internet access was first rolling out. People said, "Wow, it's a lot faster." But the other thing was that there was no handshake. You didn't have to dial into your AOL account. It was just there when you wanted it. With AI, that ability to create and iterate on lots and lots of different messages is probably the thing that might have an impact, at least in the future.
Nate: It's also hard to demarcate where normal technology ends and AI begins. If you're using Microsoft Word these days, you're using an AI product. If you're using Google, there's AI in it. If you're using an iPhone for pretty much anything, there's AI in the applications. That's one of the challenges as we think about regulation and the use of AI in campaigning. Every person at an advertising company right now is using AI. They're not necessarily using it in a deceptive way, but it's being used to put together the videos or the text, in the same way that students are using it for their research papers, academics are using it to do their work, and lawyers are now using it. So every component of work that's relevant to a campaign or a political process, including election officials themselves, when they're communicating with voters or using tools to check for fraud, is going to involve a lot of AI.
Rob: You mentioned regulation. How is the likelihood of regulation going to shift now? The market obviously believes it's going to change quite a bit, based on what we've seen in the week since the election. How do you see priorities shifting?
Nate: One of the interesting things about the AI regulation debate, at least among the mass public, is that it has bipartisan support. And frankly, even among elites right now, leaving aside the Trump administration for a moment, even among elite members of Congress, you don't see the kind of polarization on AI that we see in pretty much every other aspect of social and political life. That doesn't mean we're going to get a big legislative response, because at the federal level you still have to get past the president's veto, and he's been four-square in an accelerationist mode. Though maybe the writing is on the wall because of the mass rebellion against data centers, and maybe that will lead to some changes. But there are many different types of AI regulation. The question is whether you'll get a comprehensive law like they have in the European Union, or sector-specific laws that deal with things like endangering kids with AI, or political advertising with AI, or privacy, or something specific to data centers, like their energy use. AI regulation runs the whole panoply of these different sectors. So I suspect we'll get something. The question is whether it's the kind of comprehensive legislation that most people associate with AI regulation.
Rob: It also seems like we'll continue to see a patchwork of state-level activity. And perhaps even more so, I'd think there are a bunch of state attorneys general who are quite interested in moving their agendas forward on this.
Nate: We've already seen it. We've already got about eight AGs suing, I can't remember which AI companies, in a case analogous to the kids case that was brought against Meta, but specifically related to AI companions. And we saw Character.AI is now in bankruptcy because of the lawsuits brought against it. So yes, you're going to see the state AGs. But also, as you said, the laboratories of democracy are going to start experimenting. California is first among equals here, both because the companies are here and because it's been more out front and aggressive on AI. We have SB 53, which is basically a larger set of transparency measures for the AI companies. And the governor has a commission making recommendations, which I believe will be released next week, on the implementation of SB 53. Colorado has its laws, which have gone back and forth. New York has the RAISE Act. So you should expect to see more experimentation at the state level. That's certainly a suboptimal outcome. You don't want ChatGPT to work differently in California than it does in Idaho. But that's what happens when the federal government is unable to address these concerns.
Rob: One underappreciated thing, speaking as someone who has often built technology to comply with regulations, is that having one set of things you can ship across the world, or across a single country, is a lot easier than having to version or shard your product 50 different ways. That seems obvious, but I think regulators sometimes don't consider that if they partnered with some of these other entities, they could get 80 to 90 percent of what they want voluntarily, instead of spending years in litigation or regulatory hell trying to get it.
Nate: I do think that what's happening in the industry right now is that we're lurching toward a consensus among most of the major labs on what the skeleton of a major AI regulation might look like: pre-deployment testing, incident reporting, some kinds of transparency requirements. I'm not saying they're running to embrace all of those. And of course there's a lot of diversity in the industry. Jensen Huang is different from Elon Musk, who's different from Dario Amodei or Mark Zuckerberg. They're not all swimming in the same direction. But particularly after the Hugging Face incident over the summer, there's a sense that the calls for regulation are not mere industry capture or virtue signaling. There's a sense that we need to do something now to chart our AI future before things get out of control.
Rob: Some of the other things we've talked about were information harms: Can you trust this information? Is this real? Now we're talking more about safety, loss of control, agents doing things outside their parameters. It sounds like you're seeing that shift in the conversations regulators are having with the companies.
Nate: I think all of these are happening at once, especially as you get political entrepreneurs who now have the wind at their back because of the elections. You've got some folks who are going to focus on kids. Marsha Blackburn, who's now going to be the new governor of Tennessee, has been out front on kids and social media, and I think kids and AI. You might expect something to happen there. You've got others who are worried about the existential-risk questions and want to go in the direction of AI safety institutes and pre-deployment testing. For the most part, I think that's where the labs are: if there's going to be federal regulation, they're comfortable with that. And then you have others calling for moratoriums on data centers. That's a whole other set of concerns. My general view on the data center debate is that while there are genuine concerns about NIMBYism and maybe electricity rates and all that, it is literally a concrete manifestation of people's AI anxieties, and therefore it's something to protest against. Whether you're concerned about kids or privacy or loss of control and existential risk or the labor effects, the one lever you have is to say, "I don't want an AI data center built in my community." So you channel it toward that. Now, I tend to think that's totally misguided, because there's a lot of misunderstanding about data centers and their environmental impacts, water and all that. It's not really about those issues. That's not to say everybody should want a data center right outside their front door. There are perfectly good NIMBY reasons for that, just as there would be for building a stadium or anything else. But we should deal with the actual problem. The problem people have is with an industry where, if you go into the OpenAI cafeteria, the food is more regulated than the models that they produce. So we should have some sensible regulation that charts rules of the road for model development.
Rob: That certainly makes sense. A question I get asked all the time, including when I was on CNBC recently talking about AI safety: if we slow down on our side, will the Chinese labs slow down? It's a very valid question, and the answer is probably no.
Nate: There are two dimensions to that question. The first is, what about China? I think that's a serious question. It's not clear to me that regulation necessarily means slowing down, by the way. There are ways to do regulation that simply make sure these models are checked before deployment. And I tend to think that, leaving chips aside for a second, which of course are in the news, so much of the advance in Chinese models comes through distillation, through use of the closed models that we here in the U.S. are producing. So while the gap might narrow, it's not clear to me that we'd get overtaken just because we put in some more safety measures. But that's a legitimate debate to have: how much will a law end up slowing development? Most of the AI labs are now saying they have to pace the frontier, whatever that means, even as they keep releasing these models. There's a second dimension to your question, which is correlated with China but a little different, and that's what you do about open models. The thing that frustrates me a little about the AI governance debate is that everyone continues to think about it the same way they thought about social media: it's a few companies, which it is, and if you just regulate those companies, you've solved the problem. But AI is a technology with many manifestations. The rules you apply to ChatGPT are not the ones you're going to apply to autonomous vehicles. And it's a technology that's widely available, regardless of which U.S. companies you regulate. It is unfortunate that the most authoritarian government that is leading in AI is also making the models most freely available. Most of the strong open models are coming from China. Of course there's Llama from Meta, NVIDIA has its open model, there's Mistral, and others are out there. But you cannot just treat open models as a safety asterisk, where the responsible actors in the system have to submit to all the safety testing while anyone, in theory any terrorist around the world, can adapt the open models. You've got to figure out a way to deal with that, and it's a really difficult problem.
Rob: I was listening to my former colleague Alex Stamos talk about this over the weekend. He was saying you can get GLM 5.2 abliterated, with its safety training stripped out, run it on an M5 Mac, and cybersecurity threat actors are going to be doing exactly that. Meanwhile, and this is less about regulation than about safety measures, a lot of researchers and defenders, and I've encountered this myself in our work, are getting refusals from the centralized models, Anthropic or ChatGPT and so on, when we're trying to do defensive cybersecurity or trust and safety work. So I think this sets up a very problematic dynamic.
Nate: And this is exactly why government regulation doesn't necessarily inhibit innovation. We could imagine a government-sanctioned pathway for people like you to get access to models before deployment so you can develop cyber defenses. But because of the antitrust problems and the collective-action problems among the labs, it's going to be very difficult to do that without some kind of government intervention.
Rob: I think that's right. All of these companies have programs you can apply to, but it's unclear what the criteria are.
Nate: And of course Amazon and Microsoft will be first in line. But we need thousands of firms to have access to that stuff, and we need some system in place to make that happen.
Rob: Speaking of regulation, I think I mentioned this when I messaged you: the Platform Accountability and Transparency Act, which you proposed about five years ago. Can you tell us a little about it, how it's moved in the last few years, and where it stands? Is it still relevant in your mind, or are there new things to add?
Nate: I think it's still relevant. Let me give a little background for listeners. Along with Gary King at Harvard, I worked on a project with Meta. We didn't get funding from Meta; we got funding from foundations to try to make Facebook data available to the world's social science community. The reason was that whether you're talking about disinformation or how people use the platform or polarization or hate speech, the people inside the companies have access to the data, but those on the outside don't. And this is some of the most valuable data for understanding humanity right now. Unlike in previous eras, it's locked up in private companies. So we were trying to figure out a privacy-protecting pathway to make the data available to social scientists. We had very limited success, though we spent many years trying. One reason was the privacy issues. And some of it undoubtedly had to do with trying to get this done with one company, as opposed to having a law. So after that experience, I was frustrated. And in Facebook's defense, when we went to Europe, I was trying to convince Facebook that GDPR, the European privacy law, would allow it to make data available to independent researchers. I'd say to the European regulators, "Look, I just need you to tell Facebook that this is okay." And they'd say, "No, what we'll say is that they have a very good argument that this falls under the public-interest exception of GDPR." And I'd say, "You can't say that. You've got to say they're allowed to do this, because there's potential liability for them." Why should they give out the crown jewels to outside researchers? That kind of argument happens all the time, let alone with the FTC or whatever might happen in the U.S. Fast forward, and I thought: if we're going to mandate, not just for Facebook but for other companies, some privacy-protecting pathway for researcher access, how would we do it? The Platform Accountability and Transparency Act has two big goals. One is public disclosure of certain kinds of information related to advertising, algorithms and widely viewed content. The second is a privacy-protected pathway, highly regulated by the FTC, that would allow outside researchers to access the data. It's languishing in Congress. I was never under any misimpression that just because a Stanford professor puts out a piece of legislation, it's going to get passed. I released the first version when Frances Haugen was testifying, because I thought, "All right, you want something to latch onto? Here you go." Chris Coons, Amy Klobuchar and Rob Portman were the initial co-sponsors. Mitt Romney came on board, and it's been resubmitted by Chris Coons with other bipartisan co-sponsors since then. It's been significantly improved from the draft I originally wrote, so I don't want to take credit for all of it. But it's there. Now, what I think needs to happen is that we need to figure out how to add the AI companies to this. It's not enough to just essentially fight the last war, although I think it's still a war worth fighting. And by the way, it's a war the EU has picked up, because the Digital Services Act has a researcher-access provision that's not dissimilar to PATA. But we need to think about what transparency is going to look like going forward, particularly for the AI companies.
Rob: It's interesting that the DSA folks and Ofcom and others are looking at how they can evolve. They now have some experience with how people are using the data, the different access levels, the slowness in implementing trusted flaggers and other mechanisms, and they're trying to improve. So hopefully we can learn from, and maybe leapfrog, some of those efforts. But I'm not going to hold my breath until I actually see it happen.
Nate: Also, not to get too far into the weeds, but transparency with AI is a little different from transparency with social media companies. For all the difficulties with social media, you could observe what was happening from the outside to a greater or lesser extent, though not as well as people inside the firms. With chatbots, the interaction is happening almost always in private. Except maybe Grok, which is a special case, these are one-on-one interactions. You and I, on the outside, really don't know how people are using AI. We get reports, principally from Anthropic, on things like what share of conversations are about politics. Of course, politicians and regulators think there's a hell of a lot of politics happening on these platforms. It's probably a really tiny share. That was always true with Facebook, too. But we need to know how people are using it. Because of my perch here in Silicon Valley, I get some insight from rubbing elbows with people at the companies, on things like how many people are using AI for companionship and psychological support. It's a hell of a lot more than you might expect. That's an area where I thought it might be a so-called tail problem, but people really are using it, particularly younger people, to have conversations, to treat it like a friend. From a public policy standpoint, that's important to know: how people are using it and what you want to focus on. How much imagery is being created with these models? What kinds of responses are people getting? So there's a way forward. You could compel the platforms to offer users an option to donate their data. I don't think you can compel individuals to donate their data, because a lot of it is private, but you could require it as a feature, built into the user experience from the get-go, that would let people donate data. That might provide us some data. And you could require the companies to submit certain kinds of reports beyond model cards and beyond incident reporting, on how the tools are actually being used.
Rob: I think regulations often get hung up on process, like "We need to understand how you do this thing," as opposed to: What are the outcomes? How are people being affected?
Nate: I agree, and I'm glad you said that, because I could not agree with you more. This is the thing that always frustrated me when I talked to academics, frankly, about what a piece of legislation would look like. They'd say, "It's not enough to get transparency on what content is being viewed on Facebook. We want to know the internal decision-making processes at the company." Two things. First, you're not going to get that. Second, even if you could, you'd be changing the way the company behaves. If you get all of the internal memos, those memos won't be written going forward. And this is the lawyer in me speaking: you've got to try to get half a loaf here. If we could just get some information about how people are using the platforms, that would be a huge leap forward.
Rob: And the same argument about how powerful the models are applies here. You can now have many more third parties with the capability to actually process that information. They'll need some funding for compute, and probably not a trivial amount, but the platforms could provide a set of conversations or images or videos, and decently funded academics or nonprofits or other entities could process them and come up with meaningful insights.
Nate: Correct. I think academic access is important here with AI for the same reasons it was important for PATA. But with AI there's also a whole other level of transparency issues, and that has to do with third-party evaluators. Not really academics, but outsiders. This is the METR question from the Hugging Face episode. That was brought in for a compressed period of time to deal with a particular emergency. But will you regularize that, so that you have some kind of input from third parties who have insight into how the models are going to perform, or how they're performing on an ongoing basis?
Rob: I'm an advocate. I think that makes a lot of sense. One of my concerns is that there's probably a very small number of people who are qualified to do this and want to do it, versus going off and starting their own AI training-data company and making hundreds of millions of dollars. Or maybe that'll change soon.
Nate: That's where government regulation can reset the balance a little. If you require it, then it's not just the METRs and the Fathoms and the Transluces of the world, or Apollo. It's going to be Deloitte and McKinsey and Accenture and Ernst & Young. A lot of these auditing companies might dedicate resources to it if it makes sense for them. Now, the challenge, and this is going to be a problem in the EU in particular, is: who are we talking about? Just the foundation models, or everybody in the AI stack, including the apps? This is what worries me a little about the EU AI Act from an entrepreneurship point of view. The rules they have for the model developers, like OpenAI, I think are perfectly reasonable, though I might write them differently. And the companies have basically signed on to the EU's AI code of practice. Where things get really difficult is that the whole regulatory regime Europe has come up with is based on risk tiers, and high-risk systems have to go through all these different requirements. And what counts as a high-risk system or application is a lot of the economy: health care, education, banking, all these kinds of things. I understand we ought to be careful about how AI is deployed in those areas. But I just taught my 200-person AI governance class here at Stanford, and I've got maybe 70 nineteen-year-old engineers in that class who, if they spent the next month trying to develop an application for finance in the European market, could do it. And if they're not subject to the same pre-deployment rules as the EU companies, those companies are going to have to compete with my nineteen-year-old Stanford students, who might put them out of business.
Rob: One thing I did want to ask about, going back to content creation: so much content is now created by AI or with the help of AI, and platforms have been protected for a long time by Section 230. How do you see the regulatory framework around liability evolving?
Nate: That is the trillion-dollar question in a lot of ways. First things first: Section 230 does not really apply to AI-generated content produced by the company. It can get complicated, because where does human agency end and AI intervention begin? There are certain situations, say when Google's AI Overviews are summarizing things on the web, where Section 230 might apply. I'm just being my lawyerly self here. But there are cases right now, like defamation cases involving AI, where ChatGPT says Nate Persily committed murder last week, or Google says something like that in its Overviews. There's a really interesting case out of Minnesota involving a company called Wolf River Electric, where Google's AI Overview basically said the company had engaged in embezzlement. And they can show that since AI Overviews came on the scene, they've lost millions of dollars in business. Section 230 is not going to cover that kind of thing. Then there's the general liability question you were pointing to: for what kinds of harms are AI companies currently liable? This brings up all the tort law issues we dealt with in the social media age. Take the suicide cases, where an AI bot or companion counsels kids toward suicide or self-harm. Will the same arguments that were used against Meta now be used against OpenAI and others? I think we're seeing the start of that. And here the company is more directly responsible, because in the Facebook context, at least, it may have been user-generated content, which would be protected by Section 230. The argument in the social media addiction cases was that the design of the platform is not covered by Section 230, and that the design led to certain harms. Those kinds of design arguments are going to be used against the AI companies: you did something in the design of this product that made it dangerous, and you could have designed it differently. To some extent, we know that's a perfectly fair legal argument. Take an autonomous vehicle that starts going wild and running people over. Liability is going to play a very important role in charting our AI future and assessing the risks. The question is whether liability law on its own is going to be enough, because that's what the Trump administration and some folks in the industry have said: just let tort law handle it. The point is that tort liability regimes can deal with some problems, but not all of them. If you have highly correlated risks that lead to a bioweapon, or existential risks, it's all well and good if you can get compensation. But if we're talking about a virus that's killed millions, that's not terribly helpful. So you need to marry a normal liability regime with greater federal legislation that tries to assess when companies are liable and when they're not.
Rob: The other interesting thing is when people tell software to go do things on their behalf, like buying products. I'm trying to figure out: if I get my Muse AI agent to go buy something on Amazon, and then I decide it bought the wrong thing and I forgot to return it, can I charge back my credit card? What is going to happen in those cases?
Nate: That's a great example, because it's another case where law doesn't inhibit innovation; you just need answers. With some of these agentic questions about who's responsible, in some ways it's more important that we get an answer than that we get any particular answer, because once you have an answer, the market can evolve around it. The most famous liability or contract case involving an agent happened with Air Canada. Air Canada had a bot, a simple voice bot, that was doing some things online, and it sold fares that were much lower than Air Canada wanted. The question was: when a bot does something and makes a mistake, who's responsible? And in Canada they said, "Sorry, Air Canada, you're on the hook for this. If you promise that these fares are going to be half price, your bot did it, you're responsible." Okay. Now we know Air Canada is going to try not to make that mistake again, because it's going to make sure the bot doesn't do that. Similarly, whether it's your Muse agent or whatever bots are interacting with each other, we're going to need some rules of the road that then get baked into the technology about who's liable when things go wrong.
Rob: Here's an interesting quasi-example. I asked my Muse agent to buy something on Amazon, and it was pretty well publicized that Amazon didn't want Muse agents on its site. So I asked Muse whether I could get to the Amazon site through my home network, and how I would do that. It told me all the steps, helped me set it up, and then said, "You might also want to install a stealthier browser than the one I ship with." And it did all of that for me in its virtual machine. I didn't actually buy anything on Amazon, but I was able to access the site and presumably could have. So is Muse at fault? Is Meta? I don't know whether Meta and Amazon are going to get upset at each other.
Nate: Let's think about the different actors there. Actually, four. Who could potentially be liable if something goes wrong? Meta potentially could be. You, as the user, potentially could be. Amazon might be. Or the agent itself, which is a weird thing to say, but we're in that world right now. What does it mean for agents to be liable for things? Especially as you think about open-source agents and how they'll be rolled out, there's a movement afoot where some people say every agent should be registered somewhere and have insurance behind it. I don't think we're going to live in that world, but you can see the new issues arising from the agentic revolution.
Rob: And there's also a degree of non-determinism, just as we as human beings have agency and may or may not decide to comply with rules. When I posted about a version of this, some people said, "I couldn't get my agent to give me that information about going through my home network," and other people said they could. So Meta could say, "We tell it you're not supposed to do that, but people may mess with it, or it may just happen organically, or we don't really know."
Nate: When I teach my students the liability cases, whether it's defamation or other questions, take defamation as an example, because it's front and center. If you go into ChatGPT and ask, "What crimes has Nate Persily committed?" and it says Nate Persily committed murder last week, that's potentially actionable. That's defamation. If you go into ChatGPT and say, "Tell me a story about a Stanford professor named Nate Persily who committed murder last week," I don't know. Is ChatGPT really on the hook for that? Now suppose you have DALL-E or whatever the image-generation tools are these days, ChatGPT or Nano Banana, create an image or video of me committing murder or robbing a bank. Are they on the hook for that kind of defamation? At what point do we say that ultimately the user is responsible, especially when you get into jailbreaking? If the company has done everything it can to prevent users from doing something, but you're a clever guy who's figured out a way to get it to defame someone, maybe that's your responsibility at that point.
Rob: Coming back to regulation: Ro Khanna has proposed something like an FDA equivalent for AI, an AI safety agency that would license frontier models and so on. What do you think of that kind of proposal?
Nate: I think there's a good argument for that, whether we do it that way or another way. There's a lot baked into it. The first question is whether we want to regulate AI at the model layer, the technology itself, or in the applications. The answer is that we probably want to do both. Inevitably we're going to regulate the applications, whether through judges or through new laws on autonomous vehicles or whatever. Then the question is: for the developers, are we going to regulate the models, and if so, to what end? This is where the existential risks and the CBRN risks, chemical, biological, radiological and nuclear, plus cyber threats, come in. You probably need a federal government effort to address those. Whether you create a new agency or locate it in the Department of Defense or in Commerce, we can debate that. But if you're going to have pre-deployment testing, incident reporting and some transparency, there has to be oversight by some government body. Now, every country trying to deal with this runs into the problem you were suggesting with the outside auditors: we don't have the talent in-house, because so much of the talent is in industry. We're trying to provide pathways for Stanford students to work in government on this stuff. The UK has become the leader, I think, because its AI Safety Institute, which started after the Bletchley summit, has become first among equals when it comes to testing models. Singapore has hired a huge number of AI experts into government for such a small country. We in the United States have a real scarcity of AI experts in government. I'm sure that if we had the will, we could pay people to serve in those roles. But it shouldn't be just a small group. There should be at least hundreds of people doing this, especially if you're talking about each of these kinds of risks, and to make sure model developers aren't languishing for half a year waiting for an agency to give its blessing to a model. You're going to have to do this quickly, for the geopolitical reasons you mentioned about China catching up, and because of the competitive pressures inside the industry.
Rob: I agree with that. That makes a lot of sense. One final topic I wanted to ask you about, as a professor at one of the leading law schools in the world. Speaking for myself, I use the models a lot for questions I might once have called an attorney about, and I come from a long line of attorneys in my family. My day job is in trust and safety and cybersecurity, and that industry is being changed massively. But it seems to me that, whether it's already happened or will happen, the legal profession is going to be changed massively too. What are your thoughts on how that's going to play out?
Nate: That's a perfect question to ask me today, because I literally just gave a 90-minute lecture to my students on AI and labor-force impacts. I talked about call centers, radiologists, teachers and lawyers, because that's where we have some evidence. So far, we don't see huge impacts on the employment of lawyers from the rollout of AI. As with so many other industries, people say it's coming. We started the AI initiative I lead at the law school to look both at the law of AI, which is everything we've talked about up to now, like regulation, and at AI for law: What's it going to do to the legal profession? What's it going to do to law firms? What's interesting is that the big law firms are rolling it out with zeal right now, and they're adopting different models. Some are treating it as a technology-procurement issue, the way they treated Westlaw and Lexis, the previous legal research tools. So they go for products like Harvey and Legora, which are legal AI tools. Others are adapting the foundation models to their own circumstances, which makes a lot of sense, because frankly, whether it's Harvey or Legora or another tool, the back end is usually one of the foundation models anyway. So it's not clear what you gain. That's a frequent question in the "SaaSpocalypse" debate, about whether these kinds of applications have a future. Leaving that aside, there are big firms spending a lot of money on bespoke AI tools, because they're treating them as the next generation of Westlaw. Then you have law firms, and we're actually having a session here at Stanford tomorrow, Thursday, November 12, on AI-native law firms, where the future of the law firm is being reinvented. These AI-native firms are very top-heavy: a few partners rolling out hundreds of agents to help them do their work. There's no settled model yet for which future law firms will choose as they integrate this technology. As with everything else in the professions right now, the key question is how you develop apprenticeship and training for young lawyers going forward, because there's pretty much nothing a young lawyer can do that an AI cannot do better at this point. So we need to make sure young lawyers have a pathway to get training, so they can eventually become the partners who use the agents to do their legal work. That model hasn't been worked out yet. And it's affecting legal education as well. I'm teaching a class on using legal AI tools to some of the 1Ls. We really need to rethink how we're training our lawyers.
Rob: That makes a lot of sense. An attorney explained it to me recently this way: they think the amount of work at these large law firms is going to keep going up massively, and there'll be a lot more incoming, because people can use these tools to file more lawsuits that big companies will have to defend against. But they also think there'll be a ton of pressure on the judiciary side of the equation. Just processing all of that incoming work creates a real choke point, and that's one of the things they said they're quite worried about.
Nate: But that's also something AI can help with, and this is one of the challenges. To take it outside the U.S. for a moment: I've done a lot of work in Brazil lately, including meetings with the Brazilian Bar Association, where I learned that the Brazilian judicial system has a backlog of 40 million cases. Not 40,000, not 400. Forty million cases sitting there. And they said using AI is not optional; they have to use AI to try to deal with this backlog. That speaks to your point: if use of the legal system increases, we're going to have to use AI to manage it. Now, we've got to deal with the hallucination problem. We've got to make sure there's human review. We've got to deal with bias and all the other familiar problems. But AI is going to be incorporated by the judiciary just as it is by lawyers, and we're going to have to develop new rules of the road for all the actors in the system.
Rob: That makes a lot of sense. I think the other thing is that, the way you might have a public defender, you're probably going to need public compute so that people can be defended, or can use these tools. There are a whole lot of implications. But I appreciate all the time, Nate. It was really great chatting, and I really enjoyed talking about all of this.
Nate: This was a really great conversation. My pleasure. Great to see you.