Scott Shapiro (Yale Law): Law as Code & the AI Power Shift

(upbeat music) - John, we're talking about AI in the law today. Curious, what are you most worried about or? - What are you most excited about? - You know, I'm worried that AI's gonna take all of our jobs. I'm optimistic that it will help meet the huge market of undemanded demand for legal services. - Yeah, it's the contrast in most professions, but as lawyers, we sort of want to hold on to our jobs as client service professionals, we're happy, I guess, that our clients will get better service. - I don't know. (laughing) - Or maybe there will just be more clients. That's that I think is the optimistic hope. - Or maybe we'll all just become AI entrepreneurs, I don't know. - Or AI entrepreneur supervisors. - Yeah, we'll become babysitters to our own replacements. - We had some conversations about who would be good on this topic and, you know, there's a lot of interesting perspectives and I think our guests today will bring a broad one and an interesting one. Who are we talking to today? - We're talking to my Yale Law School colleague, Scott Shapiro. He's a professor of law and philosophy at Yale Law School and I think what's really interesting about him is that he thinks about both AI and the basic nature of law. He's written books about what law is and he's most recently written a history of hacking and cyber security. - Yes, the founder of the AI lab at Yale Law School and he's also a founder of a new AI legal technology company that we're gonna hear more about. - Okay, so it seems like it'll be a great guess to ask the impossible questions of. - See, he's used to getting tough questions as a legal philosopher and I think as someone who thinks about the basic nature of law, I hope you have an interesting perspective on what AI will do to the law. - All right, well, why don't we bring him out? - Well, we'll pepper him with as many difficult questions as we can come up with. - He's a law professor, he's used to it. (laughing) - Oh man, I'm so ready, go ahead. Let's start with the singularity. (laughing) - First of all, thank you so much Joel and John for having me on your podcast. I'm really excited to talk about this subject which is really something that's of grade, I think not only intellectual interest and social importance but it's the topic that's really near and dear to my heart because well, in my work, I think about the law is a kind of social technology. It's this incredibly sophisticated mechanism for enabling communities that have many people conflicting opinions and complex goals enables them to cooperate and coordinate their, like a civilizational operating system. - Exactly, as exactly what it is, it's like an operating system for, I would say law self conception is an operating system for the just and good communal life though, of course, in practice it doesn't always turn out that way. Of course, it's often an instrument of oppression but the self conception is one that is a technology that enables us to achieve those things that we would never be able to achieve if we didn't have these techniques. And what I think is super interesting is that society is becoming so complex and the kinds of things that it does are so complicated that the law we're producing to enable that behavior and that kind of coordination is becoming so complex itself that we like need another technology to help us apply it. So I think of like AI is like the next step in social technology which enables us to coordinate our behavior in ways that would be somewhat like very difficult and very costly without it, if impossible. So AI for law becomes like the next step in laws evolution. - So the law is too complicated for its own good or too unwieldy to even manage. - Yes, I mean, we get to the point where that which has enabled complexity has become itself too complex and that suggests what if there was an alternative technology a supplementary technology that enabled human beings to use this social technique going forward. And so whereas there are many things to be afraid of in terms of AI, I think it's inevitable for people like us that is lawyers who are essentially social engineers. It's a new technology for us in order to engage in social engineering. - So is the hypothesis then Scott that the law has to become increasingly complicated in order to enable increasingly complicated interactions among people and that AI will help us to comprehend that complexity. - Yes, that is exactly what AI will basically simplifying, summarize the laws so that ordinary people can understand. - Yeah, yeah, let me say none of this of course is inevitable, right? The technology has to be built and has to be built for certain purposes, whether it helps people or whether it helps really rich people, or bad people. - Yeah, that depends on like what we, what we build, but the idea would be exactly what you said, which is that the kind of complexity starts to fold on itself and that AI is, I believe, a very valuable tool for people like us, especially who are lawyers, but also producing them for non-lawyers enabling a form of democratization of legal information. - Okay, I'm talking to two Yale law professors, so I guess in this, I'm the lightweight. So why don't I move it to a more concrete example? So as I'm browsing the internet, I'm clicking terms of service, I'm agreeing to these multi-page contracts. I'm imagining a world where my AI or my AI agent is able to review all those for me, flag any major concerns and maybe push back where needed, maybe even without my input. That's possible, of course, but that's not the way it's gonna work. The way it's gonna work is your AI agent's gonna say, well, actually, you just signed over everything, and we're, but like, but I wanna watch that TikTok. I mean, the issue with terms of service or two, one is that like, obviously they're so complicated, nobody reads them. And number two is most of the things that some people care about, most people don't care about, which are like privacy. And so like, okay, so they have access to my phone, but again, I wanna watch the users who are gonna rate it content. And then finally, so what are you gonna do? Like go to the other TikTok. I mean, I guess there are other, there's reels and there are various other things, but the point about terms of service is they tend to be contracts of adhesion. So like, it's not like I think my AI agent is gonna renegotiate a thing with Oracle. It's just gonna tell me the terrible things, and I'm gonna click I agree. But then even it will be capable of telling me the terrible things. - Yeah, so it will be, but the thing is, you already have that ability. It's called reading. You could take a look at it, but none of us, it's just not worth it to any of us to do that. - Well, maybe if we dream a little bigger, maybe our AI agents will coordinate with seven million other ones and pressure the TikTok TOS to change. - No, that's right. I welcome that antitrust lawsuit to see how that plays out. I look, the future, my senses that was driving, clicking behavior in this case is very low cost to most users and the inability to change it anyway, even if you knew what it was. - John and I have talked about AI as a complexity amplifier. It sounded like you were describing it as a complexity reducer. What use cases are jumping out at you regulation? So it just turns out that most of the law applied in the United States and probably in many other countries are not applied by officials, they're not applied by lawyers. They're applied by employees, managers and companies, officers and companies who people at call All centers who answer questions about how to use your insurance or whether you can withdraw funds from your flexible spending account to reimburse child care. These are all complex legal questions and they're being answered by people at companies who tend not to be tax attorneys and even if they were tax attorneys the price of our, the things that we buy from these companies would go way, way up. So actually think that like in industries that are heavily, heavily regulated, which so many in the United States are because they tend to be high tech rate industries, people have to apply legal rules and statutes and regs all the time. And it seems to me that certain forms of artificial intelligence act like calculators in the days before when you had accountants who added things up by hand. I think now we'll be able to use sophisticated rule engine machines to determine what regulations require. In fact, actually I think there's going to be a whole new area of legal work where people are going to use AI to break laws, to break contracts, to break or to escape. Yeah, right. I'm using break in the kind of the the coder hacker sense of break that is you have a system that's designed for something and you hack it that is you you engage in some behavior with the system that was in some sense unexpected and it produces a result that was not the intended one. And it seems like we can either use AI to figure out how to why comply with my legal obligations. I could also use it in the other way. I could use it to see how to avoid my legal obligations by engaging in what a hacker would say exploitation behavior. Are you talking about our brethren in the tax bar? Yes, exactly. I mean, let's just say our siblings in the in the tax bar are use their human pattern reasoning abilities to figure out where the holes are, but we're now going to very quickly turbocharge that by using algorithmic methods for testing holes and contracts and statues and regulations. And I think it's going to open up a whole new world of legal chicaneery. I'm really intrigued by this analogy between law and code. One of the main uses of AI right now is encoding. And what I understand you to be telling us is that for the same reasons AI is useful in coding, it's also useful in law, which is itself a kind of code. We explicitly call it a code. And so just as it's usual to be able to kind of use AI to describe what you want to do in natural language and then have it produce it in kind of official code, you can kind of do the same thing with law. You can say there's this official code and ask your AI to tell me in natural language what it wants. Or I can ask it in natural language to do something or to allow me to do something that I want to do. And the AI can figure can translate that natural language query into code. That's exactly right. Think of legal codes. Think of that as a programming language because as a programming language, you can either run it. That is you can see, oh, what am I supposed to do according to the law? Or you can try to exploit it. That is instead of figuring out what the law requires of me, I say, I want the law to let me do this. Well, can I throw a branding on this free shot? This is like a loop holds for the middle class. Well, yeah, right. Except for the people who are going to be able to figure out how to exploit these things are going to be the ultra rich. They'll have access to better models or what do you. Yeah, it's got some people have always found ways to exploit the law since, you know, homerobby's code. I'm sure. And so the question is, how will AI transform that tendency? And I think what Joel is suggesting is that in the same way that I, John Morley, as a person who's never programmed anything in my life, can use AI to do things that I wouldn't otherwise be able to do. Similarly, ordinary untrained people will be able to use AI to kind of hack the legal code as it were, which is, I think Joel's impulse behind saying it would be kind of middle class. Yeah. So like, I want to get, if I want to get citizenship in Portugal through my mom's, you know, genetic or whatever heritage, I could hire a law firm to do it or to help me do it. But maybe I could get chat to walk me through the steps and say, well, because of this, you can file this and be, if they don't respond, then you can file something else. Yeah. So I would just, I would say, yes, there's always that. But like realistically, you know, it's not like people, it's not like you could just type into an LLM, find me the loophole, right? I mean, to find a loophole, you have to kind of create a formal representation of what the law is trying to do. And then you try to figure out a way to exploit it. Let me put it this way. Like sure, not people who don't work on Wall Street could type into chat. But if you want to do a general check, you can do a general check. And then you can do a general check. all generated by these legal documents and these legal documents, if you speak to these people who write these things, I mean some of the terms are scrutinized and others are not. And then you sign these things and you know you just care about the payments schedule and things like that. But in in some cases there's been mischief where lawyers have gone back, looked at the debt instruments and say, hey actually there's things that we can do and end up screwing out the screwing the creditors. The maybe some people have heard of the J screw where J Kru used a debt instrument to kind of screw it to cleverly refinance their debt at the expense of their creditors by exploiting some clauses in the debt instrument. Now imagine throwing AI at those debt instruments. That's not it. You know, like it's like you have computer code out there that's been around for decades and nobody's ever looked for bugs. That's what I think, that's what I think that's what I think the private debt markets like right now. It's not just that the AI will come up with the ideas for exploitation. It's also that it could kind of run through the ideas at scale. I can download a million debt contracts and have an AI bat bought, combed through them on, identify the key, the ones that hang zip it to flaw. Yeah, I mean, I kind of just say it's like, like fishing a barrel. Like I said, if I wanted to become like super wealthy, very quick, that's what I would do. Well, I do want to become super well. Yeah, let's talk about this after. So Scott, you were talking earlier about how you'd need to build a model of the legal world. So this is kind of what you're doing at your company, right? Yes, sure. So I've started up with my colleague and co-founder, Rouge Capiskessch, in Yale Computer Science. And what we do is the name of the company is called Lidenitz AI. And after Godfrd Lidenitz, the great philosopher, logician kind of inventor of the idea of artificial intelligence, but also he was a lawyer. And he dreamed of one day automating legal reasoning. And I think actually we're at the point of time that we can automate legal reasoning. So one of the things that we do at Lidenitz is we take natural, let's say a legal code that's written, let's say in English, in natural language. So section 125 of the internal revenue code, which regulates employee benefit plans. What we do is we take that legal language. So it's the statute and also the administrative code. Exactly. The regs that the Treasury Department puts out actually guidelines. There's a whole kind of hierarchy of documents in a regulatory space. And so what we do is we take that and we convert that into a mathematical formalism. In our case, first order logic. And that we then use what are called these theorem proofers, which when you give it mathematical formalism, it can tell you what the implications are of those rules given some factual scenario you give it. So you could say it's interesting. So it's like a code calculator for employee benefits. So what it does is it takes the code and this is exactly how it works. It takes the code, converts it into logic, then then theorem proofers show all the possible combinations of moves you can make that is creating a legal decision tree. And then we use that to create a like a workflow, like set of questions that if you answer these questions, you'll get your answer provably. And it'll need the shortest route to it. And it will always give you the right answer. And what it provably in the sense that it can demonstrate to you each step in the chain of reason. Exactly. So when you're finished, it tells you exactly what rules it needed and gives you a proof as to why that you got the answer that you got. So interesting. Yeah. This is a different type of AI. This is a different type of tech than what we're usually talking about with the LLMS. So this is why I think that when we talk about AI is very misleading to people because there's many different kinds of AI and the AI that everyone's focused on nowadays is so called generative AI or gen AI, which though is incredibly powerful and mind blowingly powerful has the downside of hallucinations, which is creating plausibly sounding claims that turn out to be false. And it's very, yeah, it's very, very, which in a rule finder wouldn't be, would be, you know, if you ran into one of those hallucinations and it was your taxes, then, no, they might find yourself in trouble. That exactly right. But this kind of technology, this is the kind of AI, which is often called automated reasoning or formal methods. And what it is is it's like a calculator for statements, for logical statements. So if you can translate the law into logical statements, you can use it to generate automatically workflows that answer your questions without any hallucinations. That's, I think, one of the great promises, promises of this technology, which is being able to deliver answers to very complex questions in an explainable and reliable way. Now, what I'm a legal philosopher. And so one of the things I care so much about, you know, is making sure that we don't over claim that we don't say that AI can do things in the legal domain that it would be irresponsible to make it do. So I think this procedure of taking regulations and statutes and converting them into logic and then into like questionnaires to answer your questions automatically. That really makes sense in those domains that have very complex codes that are heavily regulated. It would be silly to do this to the constitution. It would be silly to do this in torts to say, like what is the necessary and sufficient conditions for proximate cause because proximate cause negligence, you know, do process all reasonable man standard reasonable person standard, all these things right. They are, they lack precision. And therefore to use theorem provers on them is this kind of silly thing to do. This is not the kind of reasoning that we expect from lawyers. The reason that we expect when it comes to employee benefit plans are deduction. Read the rules, no English, apply it deductively. But when it comes to other types of questions that we are as lawyers are asked, those kinds of techniques don't work. So if we have a will case right and say, was this person under undue influence? Like you're not going to use a theorem prober for that or I should hope you don't. And so what one of the things that I'm really, I mean, it's so weird for somebody like me, I'm a law professor, I'm legal philosopher, I started a company that automatically reasoning and I think part of what really motivated me, not only for the fact that actually what Leibniz said in the 17th century, we can finally do, but it's critical that some that people do this in a intellectually and ethically responsible fashion. And that means that what the technology and the tools that we build should be appropriate to the use case and we should not over claim. And so when we do a rule based reasoning, the tools we build use deductive logic use automated solvers. But when we deal with case based reasoning, which I haven't talked about, but like involve things like proximate cause or was there undue influence, things like that. There you have to use very, very different techniques in computer science. And one of the things that we try to do is make sure not to not to not to cross the stream, so to speak, not to use techniques that make sense for one thing like LLMs for case based reasoning and then apply it to regulation, which their LLMs are terrible at. Is there anybody of law that is truly deterministic in the way that you're describing? I mean, it seems to me like your technology is best suited to a situation where there are obviously correct answers and it's possible given a set of facts to deterministically apply the rules. But you know, creative lawyer can always make hay with any word and any rule. And I wonder, is there any really perfect use case for this technology? Yes. So the issue should never be that the tool. So remember that what the law is in any given case is a philosophical question. Because how do we interpret the law? That depends on how you're supposed to interpret the law. How you're supposed to interpret the law depends on the kind of legal system you're in. And how does the kind of legal system you're in determine how you're supposed to interpret the law? Well, that's a philosophical, jurisprudential question. That is your philosophy roots are coming out here. Yeah, right. So I want to say to John, of course, of course, some clever lawyer is going to come along and say, actually, I can use this, my flexible spending account to withdraw and pay for vitamin C that I can get reimbursed from my flexible spending account for buying vitamin C. Yeah, they'll make some argument super clever is that's all right. That's always possible. But nobody would ever try to do that because the stakes are so low and the odds are success are so low. What we care about is something that gives the right answer 99.99% of the time, but which does not choke off the possibility of dissent. So the tool should tell you exactly why it got, it gave you the result that you got, but it's always up to human beings to say, even though this is the deductive consequence of the plain meaning of the text, that's not the law. The law outruns that. And that's a philosophical claim that is not only possible, but the tool should enable the user to make if they want. I want to follow up on John's point and get your reaction, Scott. On the other end, I'm curious how the LLMs, for example, could be used to exploit this. So you could say, make the best possible case that this is within the safe harbor rule, or make the best case possible. Going towards what you were suggesting earlier that this is a loophole driver. First of all, I want to say that I think being a dom, having domain expertise or at least domain competence matters a lot here. So I think it would be a terrible mistake to use this kind of rule-based decision-making pipeline that I described. You take the code, translate it to logic, build decision trees, then build a workflow. That would be a terrible idea for SecRags, for security regulation. Why? Because the Rags are, they have a lot of vague, baggy language. What matters is what the SCC thinks, what matters with the second circuit thinks, what matters with the Supreme Court thinks. It's really not the kind of pace domain that let's say employee benefits are. It only sounds that way, but in reality, it's not at all. So what you have to do is you have to figure out what is the right way to handle this. Now I think using large language models makes a lot of sense in these contexts for the following reason. It is in the nature of this kind of reasoning that is very inductive rather than deductive, meaning you see lots of cases that share certain features in common and legal reasoning tries to figure out what is that pattern in these cases. That's something like what you're taking examples and trying to figure out what's in common. That's something that pattern matches like neural nets of which large language models are species of. They're really good at that. What you want to do is you want to use large language models as pattern matches to try to see how things relate to each other. We also build tools that do that as well, but what we do is that when the LLM prints out a result, what we do is we always check it. We always check to see does the case exist? If it does exist, is it about what the response says it's about? If it's a hallucination, you know, flag that. So there are ways, if it's not a hallucination, when you hover over it, have it pull up the case, so you can read the case. People often say in AI, we always need to have a human in the loop. I hate that formulation, not because I hate the idea. I like the idea of having somebody in the loop. We're all pro-human. Yeah, we're all pro-human here, but what I think we need to do is we need to think of legal AI tools as making a loop for the human. That as the human should sit in when they're staring at their terminal, these AI tools should bring them all the information they need in order to make a determination given their clients needs. And so it's okay for using an LLM that can hallucinate as long as you have ways of checking whether these are hallucinations. And so I think LLMs are great tools, but we better have systems in place to make sure that whatever they say is checkable. I will put the following way is that, you know, what is a hallucination to one person, to one party is the other party's legal creativity. And so one of the things you want in an LLM is something that does go beyond the law because that's the way in which the law develops. That's the work of a creative law. Exactly. That's the work of a creative lawyer. So you're not making up cases, but seeing new possibilities for novel argument. Right. And notice the idea of a hallucination makes it seem as if the law is determined in advance. But I think most people, I think most lawyers think, I know the laws work in progress and certain answers are certain ways of developing the law better than others. And LLMs would be really good at that because that's not a question of hallucination. If we all frame legal developments hallucinatorally. We all say the law is this, where what we really mean is the law should be this. It's just that what makes it not a complete hallucination is really that does make sense for a policy perspective. That's the way the law should get developed. Let's think about how this plays out in the marketplace. So what you're saying is there are different AI technologies that are appropriate for different legal use cases. And somebody, some human being has the exercise judgment about which technology is most appropriate. So do you see a, I could do that. John. So do you see a series of separate technologies marketed by separate companies? Do you see one or a handful of large companies like Lexus, Lexus and Westlora Harvey bundling these into different into services that get packaged together? How do you see the market playing? Or just one of the frontier labs itself. Yes. Yeah. That's a great question. And let me give you obviously my guess. Okay. So the first thing I want to point out, let me just get the most obvious thing, which is the great big frontier labs like Anthropic, Deep Mind, Open AI. You know, these, the ones that produce these great chatbots like Claude and Cheshire beauty. One of the things that I think really holds back models like them from being great legal LLMs is the fact that the vast majority of the highest quality legal materials are in proprietary hands. That is the, the, the memos that lawyers write or now long emails that lawyers write analyzing a state of the law and then are used for particular purpose and then just sit on hard drive, on share drives in law. Have you ever done any, even back in the envelope calculation on that, like what percentage of legal paper is publicly accessible versus private? That's such a great question. I would think that like what do lawyers produce that are, that's public? So I'm going to court filings. Right. So the court filings brief. but not memos. Okay, what's on Edgar? What's on Edgar? And then also let's not forget statutes, regulations and court opinions. They're written by lawyers and their public. But what do you really want from a lawyer in a firm that's doing litigation? You know, you want some kind of really smart take on the law that enables the lawyer with the client to engage in risk assessment. Yeah. And all that attorney client privilege stuff. That is all attorney client privilege. And what does the LL what does chat GPT have access to? It scrapes the websites of law firms, which they do not put their work product on that on their on their front pages. And cases and statutes and some other stuff that gets that are in the public domain, maybe some law review stuff. But the work product is not there. And so what I so that's why I'm a bit pessimistic about large language models commercial frontier language models being able to do great legal reasoning anytime soon. I do think that the up one way I can see this working out is that with the new Trump administration AI plan, which is really pushing open source models. But I think your anti-wope models, did you mean? I was open open source models that that let's say practice groups and law firms can download and train themselves on their own or somebody train those models for them. So that the models that they're using are trained on their tape on the law and can mine the maybe decades worth of experience that this practice group has had. And so I like scat and AI or yeah, exactly. LLM it would be it would be even more narrow. It would be scat and 10 be five commodity defense. White collar commodity defense. Like that. If there's a group and they do that because they have expertise in that, then they probably have tons and tons and tons of material, which could enable them to engage in an awesome training example. And so what you would be seeing was that these firms could synthesize their past work into bespoke models, which either there could be either they would keep them themselves or they could if there were some privacy preserving way of ensuring that there was no data leakage, they might be able to share it on a marketplace where other law firms could maybe rent, lease their models for particular purpose. You know, I mean, there's so many incredibly interesting things that can be done now. And the least interesting things are exactly what's happening, which is people putting wrappers on chat GPT and then calling that league lay eye. One implication here is that anyone in possession of private data in the form of the written word is now sitting on something that was much more valuable than it was three years ago. Oh, absolutely. I think this is. It's data is now hugely useful for training. That's right. Like basically we're our our employer, John is like Yale University. Like what do we do at Yale University? We produce unbelievably great legal training material. And we throw it out. Yeah, what do you mean by this is a hot take you guys were chatting about this a little bit before. I mean, so maybe the law school as a data mine. Yeah, of course. What do we do that is we engage in reinforcement learning and reimbursement training. So what happens is is that somebody produces an output. It's called their homework. It's called their draft brief draft memo and they give it and then somebody writes back detailed comments and the person then produces another draft and the other person writes another detailed comments. And this is amazing training material. And then what happens to it we throw it out. It gets deleted. But retained in the brains of the student. No, that's try that's true. It's exactly it's retained in the brains of our students, but it could be used to train a legal AI model in a way that really builds on the strength of what a university is as opposed to a frontier AI lab. A frontier AI lab throws 10 trillion tokens or something at a model and sees what happened. We do the complete opposite. We take, I don't know, 2500 tokens, 3000 tokens and then we give them another 112 tokens in response. You know, we're we're giving extremely targeted and hopefully high quality rich high quality legal advice. This is some way that we could maybe plug up funding gaps that we're losing from the federal government by using our assignments as ways of training an AI model and then licensing it out. And if I could go a bit meta here, I'm thinking about this, there's kind of a new economy of information. Anyone who's sitting on a trove of private information that's legible to an AI model now has a massive, it was much more valuable than it was three years ago. Anyone who has just personal knowledge of information that's not legible to AI, that's also in a way more valuable because it's something that you can't use an AI model to do. So if I have personal experience, it's not known to the lawyer at another firm, that's a reason why a client would engage me. It's less valuable is knowledge that's already been used to train an AI model. You know, like if if I have a deep knowledge of the tax code in Scotch, Shapiro has just trained life and it's AI to understand these those same elements of the tax code, that's less valuable now. What it means is that the value of really good lawyer ring is going to get much higher because the less valuable lawyer ring already got used to train the AI. But it sounds like you're saying the value of really good lawyer ring, but it's really the value of these private data troves. You could be using an AI to do the lawyer ring in some ways. I think what Scott's talking about is kind of a third category of information, which is personal knowledge that's not rendered in a form that's legible to AI. Like suppose that I haven't written my personal experiences down and those personal experiences are unique to me such that there's not, they don't appear elsewhere in a way that could train AI. That's really useful now. Yes, that is extremely useful and the thing is that what I'm suggesting is that right now universities have and people like John and I produce it every day, this incredibly useful information for training models to do some professional services. There's almost like a poetry to it that the law schools are teaching the lawyers, but maybe they could be teaching. Well, it's a great model to replace the lawyers or put it this way. What we're trying to do is we're trying to produce models that perform at as well as they can and then figuring out how to use these models for socially beneficial purposes. Okay, that's what we should be doing as universities. Now that will mean that some of models will be replacing some forms of legal work in some way for some people that will be obviously not good news, but I also do think though that given the crisis and legal services in the United States, if, let's say, yeah, law school can build a model that it leases out to four profit entities, but then allows, you know, nonprofits to use it, it should be a way of democratizing legal reasoning. And so, like everything is every technology can be used for good or for bad. And people I think are obviously focusing on how is it going to hurt us, how is it going to hurt people we know? But of course, it has this enormous possibility to help. And we should be thinking about not just how this is going to mess things up, because it's kind of obvious how it's messing things up already. We should be thinking about how we can build technologies that help people and that add. And so, what if universities could use their materials to train models, use it for socially beneficial purposes, and plug up funding holes taken away by the federal government, that could be a real, a real win. It's just there are lots of possibilities here. I love that, Scott, but what's more realistic? Who's going to pay more for this data? Is it going to be a pro bono LLM, or is it going to be one of the frontier labs that really wants to access to this training? Yeah, so I just say is that like, if Yale is using Yale work product and students are donating it, and I would suggest it would be donation a purely voluntary scheme, are they donating it because they think that this money is going to be you, this data is going to be used for the good, and they trust university to use it for research and teaching, then I think we could actually compete against these frontier labs who don't have what we have, which is they don't aren't filled with teachers and students. I like it. It's sort of like a collective alpha-go that AI that was able to win all the games, maybe like alpha pro bono. Yeah, right. I don't need our data to create this winning AI. Right, or alpha goes like a giant reinforcement learning thing. Well, you know what, that's what we do, reinforcement learning. You get an H. You get honors if you've produced of something, you get a P, otherwise. I pass otherwise. But more importantly than getting honors appears that all is that all the emails, all the comments that we write for on drafts for our students, this has to be from a reinforcement learning perspective, the highest possible data that that available on planet earth for law. It's got, tell us a little bit about your AI lab at your law school. Yeah, so, which is kind of along these lines, right? Yes, exactly. So what we try to do is we try to figure out techniques for addressing the needs of low-income households to provide actually services to other clinics to build AI models for other legal clinics at the law school. So let me tell you the two things that we've done. One is to take this rule-based decision-making, a pipeline that I described earlier and trying to use it for either enabling low-income households to claim social benefits that they don't claim because the rules are so complicated or answer their questions when they don't have access to lawyers. So one of the things we're exploring is can we create an eviction tool which will explain to users in particular jurisdictions what is necessary to defend against eviction proceedings? So that's one thing we do. And another thing we do is we try to build AI tools for other clinics at the law school. So that the promise here is if we can create AI tools that increase efficiency for clinics, they will be able to handle more clients and therefore engage in more pro bono services. I also, one of the things that we do at the Legal AI Lab at Yale is we teach students not only how AI works but also what they are undoubtedly going to be doing in the future which is training AI models. And if you tell somebody, oh, you know, train this AI model, you know, one person out of 100 law students will know what that means. And so what we try to do is we try to teach students what it would mean to train an AI model, how to do it and what to expect from it. You're developing AI technology to help lawyers at Yale law school or law students and the lawyers who practice in the clinic. This sounds really similar to the efforts that law firms are making at developing their own internal AI. I think one of the big questions that's surfacing in the legal AI spaces, how much of it will be developed by law firms for their own proprietary usage and how much will be developed by external labs whether the big ones or smaller law specific ones. Does your experience developing technology for your law school help to answer that question? You know, there are very excellent people who work in innovation departments at these law firms. I've met with many of them. I mean, they really know what they're doing. They're really competent experienced professionals. That being. They're computer science professionals. They're not. Right. They're the information technology professionals. Which I guess that what I mean to say is that when I think of computer science, I think about that as a kind of more academic, more theoretical study, the kind of things that do in a computer science department, what professors do. And just like professors, computer science department can't install their operating system, many of them. The trick is in these law firms, because these are very, very complex questions. In fact, to develop this technology is not just enough to like code the app. You have to develop theoretical insights and engage in science. And that's super hard to do outside of a research community. And you know, one of the things I really like about the Yale Legal AI labs that we work, the lawyers and the computer science professors and grad students work together. One of the things that the reason why I'm a bit pessimistic that it's going to come from within the law firms is that I think that actually these are research questions. That is they are issues that require, I think, a degree of time, resources, leisure to develop that may not seem to be that feasible within a law firm. That having been said, I don't think engineers ever are going to come up with the tools. It's going to have to be lawyers working directly with engineers, because engineers have no idea what lawyers need. Lawyers don't know what engineers are able to do. So what we have to figure out is some very, very tight working collaboration design partnerships between lawyers and engineers, software developers, IT professionals and computer science researchers in order to build these tools. Hi, 100 percent agree. I think lawyers should have engineers on their deal teams to think about how to automate processes and how to make client work more effective and more efficient. Yeah, that's right. I think innovation is extremely good about bespoke tools for particular practice groups, but whether we're going to see a tool take over that came out of, I don't know, I just picked them because they don't exist anymore. But that I'm not sure of. Ultimately, I hope that, am I expecting them? Hope. But if I had to bet, I would say that it should kind of come out of law schools working closely with computer science and engineering schools. I mean, look, this is exactly what you're doing. Yeah, I like to think that that is what I'm doing in also what the law school has been really supportive about, which is that what are we supposed to do? Are we supposed to just talk about legal AI and how it's going to affect lawyers? Are we going to start getting our hands dirty, growing up our sleeves, mixing our metaphors, and building tools that make sense because you know what, even though we're not 100% practicing lawyers all the time, we know what lawyers need. And we know the reason why when I did the case-based tool, I had us always check whether the case has existed, because what idiot uses an output that they didn't check. I just thought like what lawyer- -Alaisy one. Yeah, right, exactly. But that's not the kind of tool I want to be building for my clinic or somebody's clinic. I wanted that, like I can say, yes, this is intellectually and ethically responsible use of AI. Scott, the last time we talked about this, you mentioned that there are I think four different domains in which AI was going to change legal practice. What are they? Yeah, so the first would be what I call the business of law, you know, so like how you track your clients, how you do your billing. There, there's been a lot, a lot of interesting developments on how people run their firms, their practice groups. The second one would be legal process. That is, like if you do litigation, you know, document discovery, filing status, things litigation support. Things that on the litigation side, you want to automate, pain points, you want to automate. Obviously, the document discovery is probably the biggest pain point. And then on the transaction side, it's contract negotiation drafting, things like that. Then the third would be legal reasoning. And there's like, okay, so what is the law? And what I suggested was that one of the that it's really important to distinguish between rule based decision making versus case based decision making. And I think having one tool for all kind of legal reasoning doesn't make sense because all the legal reasoning is not the same thing. And the last thing is legal prediction. And this is predicting what a jury might do, what the SEC might do, what the Supreme Court might do. And this involves kind of techniques, a big data techniques using machine learning algorithms. And it's pretty cool, Scott, right? This is not only what's the right answer, but what this particular judge is most likely to like. Yeah, I mean, look, we know every single thing that this judge decided. You know, that's one of those public truths. Yeah. So like, this is what I was saying, like, things are so exciting in this field. Why are we wrapping everything? Why are we wrapping G.P.T. in some kind of interface? There's so many cool things that we can do that are just crying out for application. I feel like discovery is the most obvious use case. Everybody wants to jump towards writing briefs, but I mean, that's going to be transformative. Yeah, I mean, that is the, that has, I think, been the use case that since the beginning. People have focused most on. And I think obviously with, you know, kind of rag based approaches, retrieval, augmented generation. So, you know, I think that's where you use semantic embeddings to search through enormous corp, corpora of texts is an incredibly promising way of getting at a document discovery. And then just using LLMs as to to to to ingurgitate large texts and then find things in them. So that's, this spreads me kind of what you're getting at at the start of the conversation when you were talking about law as perhaps getting unwieldy and overly complex and beautiful serving the simplification engine. Yeah, that's beautiful. That's a great way to put it. Like, I was talking about how complex the codes are, but like, let's just talk about how complex due diligence is because you have to be diligent on so many different things because everything is because there's a zillion things. And so, in the library of congress in some of these cases, right, I didn't know if that's true, but that that that's a, I mean, I'm not doubting that that's true. I just meant that, yeah, I just meant to say that I can't validate, but that's a perfect example of like how we supposed to use a technology that allows us to create situations where they're basically not administerable. I do expert witnessing and every time I try to kind of learn the facts that are relevant to my tiny sliver of a case, I think with these associates who spend hundreds of hours mastering the facts, they're incredibly valuable. And then they just move to some other form. Yeah. So much better to have a technology that sits inside your firm and masters the facts and doesn't forget them. Right, right. And I mean, here, I mean, and we all know that this due diligence document review is so ridiculously imperfect, like for human beings. So it's like it's not good for any human. Yeah, I would say like it has the, the it's more efficient and is more accurate. The weird thing is that humans are weird. Humans are much, much less forgiving when computers make a mistake. Then when humans make a mistake, even when they make mistakes so much less frequently than humans. Like if a self driving car hits a human being, like that's the end of the world. Whereas even if it's one tenth is likely exactly and that has to do with the fact that one of the things about regulating AI, which I'm very wary of is that human beings are weird. And we don't actually know yet how we will respond to AI. And just because we think we'll respond in a certain way doesn't mean we will. And I would just want, I want to wait a little longer to see how humans respond to AI. What what what the use cases are going to be because it's really hard to predict. And humans are we're just we're we complain about neural nets being unexplainable. I mean, I do like 100 unexplainable things a day. Yeah, we're not as rational as we'd like. No, I know we're way more irrational than we realize. If anybody wants to see Scottsy irrationally on this irrationality on display, you can go to his blue sky and Twitter. That is that is true. I mean, it's I constantly jeopardize a job that I have no business having. So that is that is definitely your tenure, right? I am so that's I am tenured. So like why would I do something that could even jeopardize my tenure? That is what I do. That's what I do on social media. But you know, I wanted to just jump in on something you said about how doing this document review or due diligence isn't good for people. And I think about my my prior career as an M&A lawyer sitting in a document room manually reviewing hundreds of con we have contracts looking for needles in a haystack every contract right down whether it has a change of control provision. That's all automated or automatable at least. You know, it would have made my early associate labors much less mine. No, I think this is a lot of the automation in law could could really contribute to associate satisfaction. And maybe accelerate their development. Yeah, my document review story was it was during the AIDS crisis in the 1980s and it involved an insurance company. And I was directed with another summer associate to go up to the C suite to read through all their documents looking for a smoking gun. That was in that didn't make any sense. So I was like, what would count as a smoking gun? You'll know it when you see it. Yeah, exactly. And this was during, and say we're particularly interested about AIDS and HIV and like what about it? Like any mention of it. And so anyway, it was like I spent the summer looking through these things. It was completely wasted. And that something should have been automated. I see lots of opportunities for legal scholarship here. I mean, there's so many economists who want to study variation in legal documents or the effects of legal rules on things. And one of the biggest obstacles to doing that is the difficulty of coding up the contents of all those legal documents. I went time to study, for example, on the contents of closed-end fund declarations of trust and certificates of incorporation. What was their governance like? Well, so we'd like downloaded like 3,000 of these things and then had some Columbia University undergraduate RAs go through and code it all. How much faster to just have an AI bot code all of that stuff? This is just going to render so much of the world legible and therefore available to study. Yeah, I think that's exactly right classification or something that neural nets do really well. Classification is something that undergraduates don't do particularly well. And this is this is great again. Notice how we're not talking about like a chat cheaply for law. Like it's just like those are not the those are not the things that you let me ask you a question. Is there a let's say there was a chat cheaply for law. Would anyone use it? That is. Well, what about just chat GPT I'm curious when you say it's for law, you mean specifically trained on what I mean is I don't know if you've ever tried to use chat GBT to actually write. Oh, yeah, I've definitely used it. It's it's like always at the cost of being good enough but never quite there. That's my experience with it. Yeah, I think it's I think it's surprisingly good on certain things and then you're right. Sometimes it's completely unhelpful or it's not certainly not maximally helpful. But you know, if I drop drop in a document and I say, you know, review my changes and suggest any that I may have missed. It usually has helpful feedback for me. Oh, absolutely. I'm I'm I'm a big I'm a big user of LLM's what I was just meaning was that a chat you be deep for law. I meant was that you type it in and you get non hallucinatory responses that are correct. I'm not actually sure that there is a legal that there's a use case for that. That is do lawyer. I mean who do lawyers want that answer. I would say no. I would say they want one of two things either they want an advocacy piece. Okay, that that argues for certain position even if it you know kind of is somewhat hallucinatory or they want a very dispassionate analysis that doesn't come out on one side or the other necessarily but kind of tells you what are the risks here. And like that's not what the chat you be T for law does. And so again, I think if you're like I get what you're saying. Yeah, I get what you're saying. But maybe the dispassionate stuff is going to be where some of these other types of that's is and that's why you want to trained on your own work material because that'll give you a sense of what your what your risk assessment risk assessments of your practice group. What kind of things they take. We talked about how this in some ways or at least I think of this moment as sort of a golden age for the human lawyer because we have these amazing tools that we never had before. And as you mentioned technology up into this point has been a complexity amplifier. And now with these models were able to kind of. D scale some of the data problems that we've had in the past. You know, if somebody back in the day in a litigation would say I'm going to bury you in paperwork. Well, maybe that's not quite as threatening today when you can have a model parsing. But you know, I wonder what you think about going forward. So we talked about AlphaGo and how it is. What do you call that a reinforcement model? Yeah, reinforcement learning model. And it was you know, it was faring from every. Every conceivable play that had been seen in in prior chess or prior go games. And then they decided to let it just go off in its own and and run. Unobstructed against itself and it came back more brilliant, more undefeatable. And I wonder if you know where a few years away maybe but if you could see a world where the AI itself. Is the best legal mind out there the AI in it in itself. It doesn't need the private data. You know, it can create its own. Seven million iterations and evaluate each and decide what's best. So I think you're raising an incredibly deep question. Because on the one hand, you ask me for prediction. And I don't know if that's the case, but I want to accept the possibility. And I think the really interesting question then becomes what does that mean for us? What does it mean for the rule of law? Let me give you an analogy. Math has gone through the last 10, 15 years bit of a crisis where there's something called computational mathematics where. Now computers can generate theorems that human beings can't follow the proofs. Okay. Now, now you're like, okay, well, is that mathematics? Is it not mathematics? How should we think about that? Whatever. That's math, but the law is different. It's literally above the pay grade of the top sign, top map, but you read exactly. I mean, because it's just too long. They're too long. It's too complex and maybe the proofs not particularly elegant, but it gets the job done. You just can't follow it. That's math. But what do we do for law? Law is not about getting the right answer. It's about getting the right answer in a way in which we can demonstrate two members of the community that it's the right answer. Okay. And what happens when law outstrips the ability of human beings to follow it? Can it be law at that point when the function of law is to guide human conduct? And so what you're what I want to just raise is if there were such a model, it kind of raises tricky questions about what's the status of that? And how are we supposed to handle? You know what I had said was, oh, okay, well, we have complex social interaction. We need this technology called law, which enables it to happen. What happens when law gets so complex? Well, we need another technology to help us figure out what the law requires. But what happens if that technology, the explanation that it gives? It is so complex that it's not essentially outruns human reason. Have we now gotten to the, if you will, the limit point of what law can do? That's deep. The end of law. But I mean, you know, that's the next chapter of your what is law? Yeah, that's right. Yeah, right. Yeah, I never thought that. Yeah. Anyway, yeah. So that's what I think it kind of those questions that I think are just a very natural thing. I think if you think about this from the perspective of jurisprudence, it just becomes like, okay, if we did get to that point, what do we do about it? Given the fact that the function of law is to guide conduct? Yeah, maybe some of these logical checkers, you know, logical calculators can help. Yeah. So I feel that way. I think that that's the things that we're building, I think, really help with that. I do think though what happens when you're having case-based reasoning and they're very complex patterns. And somebody's saying like, you're happy with these patterns in these cases. Well, these patterns apply in these cases too, though you can't see it. Yeah, I mean, you know, we have human lawyers like competing against each other on how pattern matching should be done. It's not pattern matching doesn't have obviously right or wrong answers. I don't know, maybe you would disagree. I let them put it this way pattern, you can match patterns more or less precisely. I think the point that you're raising. is that we don't know which of the patterns matter. Yeah, and so I just don't see a judge ever trusting an AI bot saying, trust me, this is, yeah, this pattern is a better match than that one. Yeah, so this is one of those things that happens with doctors. When doctor with the AI says, oh no, this is cancer. And they're like, no, it's not cancer. What do they do? And here it really goes to the heart of what is just to be a professional. Can a professional say, I don't think this is right, but that's what the computer says. So I'm going to write your, I'm going to put you through a procedure. Or is it like, I'm bound to only act in ways that I feel I can justify to you. And so that really raises a deep question about what it is we do as professionals. And how we're bound ethically to make decisions vis-a-vis our, our, unless we're fully disintermediated and clients are going directly to the model. Exactly. I mean, I'm also just cognizant of the fact that especially in science, we're always reliant on our instruments. You know, like I can look through a microscope and see a bacteria in a culture, but I'm reliant on the microscope to represent the bacteria accurately. And it feels to me like there's an analog here where I could say, look, this model has been scientifically demonstrated to be effective and accurate. And so even though I don't understand because it's been scientifically validated, I will accept its conclusions. It feels to me that the law is harder because how does one scientifically validate the accuracy of analogical reasoning in a dispute about the meaning of a traditional opinion? Right. And I, now, I think that's absolutely correct when, I was, we taught this course, me, my, my collaborator, Roshka Piscache, we taught the course called law logic and large language models. And we did the day where we went through benchmarks. Benchmarks are a big part of large language models to try to to evaluate how good they are. And I said that when it comes to laws, super difficult to evaluate benchmark legal LLMs because it's really hard to benchmark legal reasoning. It's in the normative realm. It's always been harder to understand what objectivity and verification is. I mean, suppose that you were to take a thousand appellate court cases and run the briefs on both sides of the cases through an LLM and have it write opinions in all of them and then compare them to the actual opinions wrote by the judges, written by the judges. And suppose that the opinions produced by the bot differed 30% of the time. Or the judges wrong was the bot wrong. Yeah, well, right. Because the thing is, is that it depends on what the law is. So what's your theory of law? So it's not like you can outsource this to computer science. This is a jurisprudential question. It's not a technical question. It's a philosophical one. Yeah, a total logical question too, because the judges make the law so in a sense. Well, yeah, but like after the fact, but not before the fact, right, they make law by making mistakes. Right? Right. Are you talking about our Supreme Court? Yeah, I won't have that on this show. To make mistakes implies that you're trying to get it right. Well, Scott, I know we're about at a time. Maybe on that note, actually, maybe when we end on something positive, what are you most excited about? Oh, I am so unbelievably excited about the fact that there are there's enormous amounts of legal regulation that up until this point, we're really inaccessible to vast majority of the population. Now is within reach of not only getting incredibly accurate results, but explainable results. That is, the tools can explain why they got the result that they got. That is a breakthrough, because it then becomes that these tools can act as trustworthy algorithmic attorneys. And that in a world in which almost nobody has attorneys, that's a game changer. Got your pure self-made professor of law and philosophy at law school. Thank you so much for joining us. Thank you so much, Sean Joel. It was great to be here and talking with you about League of Legends. That was awesome. Okay.