The future of AI and the law

0:00Daniel Ho:This is Stanford Engineering's The Future of Everything, and I'm your host, Russ Altman. I thought it would be good to revisit the original intent of this show. In 2017, when we started, we wanted to create a forum to dive into and discuss the motivations and the research that my colleagues do across the campus in science, technology, engineering, medicine, and other topics. Stanford University and all universities, for the most part, have a long history of doing important work that impacts the world. And it's a joy to share with you how this work is motivated by humans who are working hard to create a better future for everybody.

0:37Daniel Ho:In that spirit, I hope you will walk away from every episode with a deeper understanding of the work that's in progress here and that you'll share it with your friends, family, neighbors, co-workers as well. We did a couple of studies of legal hallucinations and we showed that general purpose language models like ChatGPT, when faced with roughly 800 ,000 benchmark queries, actually hallucinate 60 to 80 % of the time on basic legal facts. Then we did another study of legal AI when sort of using systems known as retrieval augmented generation, where you try to retrieve the relevant legal document and then provide an answer.

1:17Those have substantial improvements, but can still hallucinate at non-trivial rates between one-fifth to one-third of the time.

1:32Daniel Ho:This is Stanford Engineering's The Future of Everything, and I'm your host, Russ Altman. If you're enjoying the show or if it's helped you in any way, again, my favorite bar, a low bar, anyway, please consider rating and reviewing it so that others can see what you think of it and can consider tuning in. We love to get a 5.0 if we deserve it, and we also love to hear the comments. Today, Dan Ho will tell us that large language models like ChatGPT can be very useful for law, but you have to be careful about a lot of hallucinations, and they might not be ready yet to replace your lawyer. It's the future of AI and the law.

2:11Daniel Ho:Before we get started, another reminder to rate and review the show, particularly if you found it useful or have learned something new.

2:25Daniel Ho:You know, large language models have gotten a lot of attention in the last couple of years. They are amazing at taking large amounts of text, summarizing it, searching through it for the pieces of information that you need. And many of us are seeing them in everyday life increasingly. Well, that includes lawyers. As you know, lawyers need to be familiar with a lot of text information. They need to know the law, they need to know regulations, and they need to review previous cases to make sure that they're making arguments that are consistent with precedent. Well, there's a great opportunity for AI to help in law, especially in finding useful laws or regulations that help your case, but also in finding out-of-date, even racist regulations and laws that need to be removed from the public record.

3:13Daniel Ho:Well, Dan Ho is a professor of law, political science, and computer science at Stanford University, and he's an expert at the use of AI to analyze legal documentation. Dan is going to tell us how his work has advanced legal reform by going through huge corpora of legal documents and finding the ones that just need to be changed. Dan, what led you to study the law and AI together as a major focus of your academic work? Well, gosh, I had long been interested in both of these talks, but honestly, it goes back even to growing up when my parents had grown up in post-war Hong Kong. That generation had fled from mainland China, my parents moved to Germany.

4:06I grew up in Germany at a time when you had a generation of folks question the choices of the earlier generation. And I have this vivid memory of actually being a young kid when the Berlin Wall came down. And so I'd always had the sense of how fragile our social institutions can be. And so it had this deep interest in really understanding our social institutions and government in particular, and then always had in mind, well, what can we do to really help make sure our institutions adapt and are able to change with the times?

4:52Daniel Ho:Great. So we're about to talk about a lot of really interesting legal applications of AI that you've dove into. And it occurred to me that I should ask you, what does the regular person need to know about law and the practice of law in order to understand what the opportunities are for AI and computation? So many of us know that it's adversarial. We know that there are cases that serve as very important precedents. We have a rough idea that there are statutes and regulations that are not laws, but that are still in some way part of the expectations of society. What would you say we need to know as we are about to dive into some of your work?

5:30Well, gosh, there's so many things one could say, but I would highlight two. One is, you know, we might, and this will maybe be too dated a reference, but the popular understanding of what lawyers do might be a Perry Mason kind of standing up in the courtroom. But so much of what lawyers really do day to day is legal research and writing, identifying the sort of operative case law and statutes that are relevant. And then the second is that law is inherently contested due to this adversarial system, as you note. And that has actually meant that over the past few years, as we've seen AI beat so many of the conventional benchmarks, that law has become a really interesting terrain for AI research because it is not a simple kind of question answering setting where you can find the right Wikipedia article, get the right answer, and you can know whether that's a correct answer or not.

6:33Legal reasoning and legal argumentation is much more complicated, and that has made really for fertile research for AI. time.

6:45Daniel Ho:And again, right before we get into some of the really compelling applications, one of the things many of us have played with these large language models like ChatGPT and others. And there's always a question about whether they're ready to go out of the box or whether they need a lot of work. So I guess a high level question is, if somebody goes in to ChatGPT and asks for legal research advice, either a lawyer or a non-lawyer, out of the box, How are these performing? Have they shown good, I don't know, legal instincts? Well, we've seen remarkable advances in the ability of large language models to actually provide answers and have a lot of legal knowledge encoded in them.

7:27And I would say when you're talking about querying a model like this for widely available legal knowledge, they actually do pretty darn well. The challenge is that it's precisely in the instances where you need lawyers, where the law might be uncertain because you're in a particular state that has a different rule or you have novel circumstances, that these are not things that can just be readily learned off given materials of the Internet. And so these systems can perform poorly. We did a couple of studies of legal hallucinations, and we showed that general purpose language models like ChatGPT, when faced with roughly 800 ,000 benchmark queries, actually hallucinate 60 to 80 % of the time.

8:18Oh, wow. on basic legal facts. Then we did another study of legal AI when sort of using systems known as retrieval augmented generation, where you try to retrieve the relevant legal document and then provide an answer. Those have substantial improvements, but can still hallucinate at non-trivial rates between one-fifth to one-third of the time. And so I think one still has to really watch out for the potential pitfalls of hallucinations. One theory paper, that's a general theory paper, has shown unless the answer is directly contained in the training data itself, well-calibrated AI systems will necessarily hallucinate due to the way that they are trained to generate.

9:09Daniel Ho:Okay, great. So thank you for that very helpful background as we now go into some of the really interesting things you've done recently. And the first one that came to my attention is looking at large corpora of legal documents for statutory statutes or regulations that are of great interest to a lawyer in a certain case, or are something that needs to be purged from the record because they're out of date, anachronistic. So can you tell me about that project? What motivated it and what'd you do? Yeah, we have, well, let me start, There's sort of two projects that are directly on this topic. Let me start with one that was really motivated by a well-meaning but difficult to implement piece of California legislation.

9:53In 2021, California enacted a legislation that required all of the 58 counties to establish processes to go through their property deed records and identify and redact them for what are known as racially restrictive covenants. What are those covenants? When we purchased our home here in the Bay Area, I had to sign a piece of paper that said this property shall not be used or occupied by any person of African, Japanese or Chinese or any Mongolian descent, except for in the capacity as a servant. And these have been unenforceable since the mid 20th century due to a US Supreme Court case, but they still persist in many of these deed records because they run with the land.

10:40And so California enacted this piece of legislation, which was well meaning, but in a county like Santa Clara, that could mean that you have to sift through 84 million pages of property deed records that date back to the 1800s. until we started our collaboration with the county.

10:58Daniel Ho:Was it simply to, forgive me, but was it simply to identify these or to expunge and rewrite them? Both, to identify and preserve them for the historical record so that people could understand this bit of local history and state history, but also to go through a formal legal redaction process that involved the recorder and legal counsel to re-record the deed record to omit the unenforceable racial covenant. Right. And this could affect house transactions every day because anybody who's bought a house knows that there's these title searches and it's all very kind of, I don't know, choreographed.

11:39Daniel Ho:Exactly. This well-meaning rule, I could imagine, might disrupt that choreography. Oh, the worry was that because it's such a huge volume of deed records, you're asking each of the county recorders to go through, it could lead to a, quote, near shutdown of county recorder offices. Until we started this collaboration, they had a team of two that read nearly 90 ,000 pages manually to identify 400 of these covenants. And that was the worry, is this is just going to take us forever. Los Angeles contracted with a vendor to do keyword searches for$8 million to conclude the process in a matter of seven years.

12:22And we felt this is exactly a place where we can help build out an AI system. And so we curated a kind of good benchmark data set of these kinds of racial covenants and managed to actually collaborate with the county to build out a model to be able to go through 5 million of these deed records and actually accelerate the redaction process and bring it down to be able to identify these things in just a matter of a couple of days.

12:52Daniel Ho:Can that be exported to the other 57 counties? We are very much interested in doing that and are currently working with several other counties, including San Francisco and Yolo County here to basically bring that technology to accelerate the process elsewhere in California. And was part of your job to digitize these or were you blessed with previously digitized documents? In Santa Clara County, we were blessed with previously digitized documents. But actually what we found is that the existing OCR, optical character recognition system, was not great. And that's been another interesting kind of discovery in building up the system is that actually the ability of multimodal models to do really good character recognition, what we found evidence of is almost a kind of leapfrogging of conventional custom OCR technology through multimodal models that both had lower error rates and were actually cheaper than conventional OCR solutions.

13:53So we actually built that into our pipeline. So we basically ingest just the page images, do the full OCR process, and then have a built out, a fine tuned large language model to actually be able to identify where on the page one of these racial covenants exists. Great.

14:08Daniel Ho:So that was a beautiful example. Thank you. Of where the AI system really was useful. It really addressed something that needed to be done. In fact, it sounds like it was mandated, but nobody had a plan. And now there's a plan. You said there was a second example. Oh, sure. Yeah. What we sort of realized after that is this is one example of a pervasive problem in legal reform, which is that so often what you're overwhelmed with if you're trying to engage in a legal reform effort or trying to litigate cases or do legal research, it's just the sheer volume of materials in front of you. I am reminded of then law professor Ruth Bader Ginsburg, who hired a small army of Columbia law students to search for 59 keywords across the entire U.S.

14:58code, which in present day is over 30 million words. And they went through each one of these to read every single code provision to identify all instances of gender bias in the U.S. code. And that formed the litigation plan for how Ruth Bader Ginsburg began to really litigate and reconceptualize equal protection in the United States.

15:26Daniel Ho:Yeah, so it really all started with the data. Yeah, and it was an incredibly time-consuming research process. And so what we built out in this other research project is a statutory research assistance system. We call it STARA, that actually is able to ingest something like the US Code, represent it in a way that we would teach statutory interpretation to law students, and then be able to apply the power of large language models to do systematic scans of statutes and regulations, much in the way that then law professor Ruth Bader Ginsburg did. But in this case, it's general, if I'm understanding, it's general purpose.

16:08Daniel Ho:So you're not specifying the kinds of statutes to look at. You're creating a platform with which a legal person could do many different searches. Exactly. Exactly. And what are the types of searches that are, well, you already told us about these covenants, these racist covenants. Are there other example searches that are kind of of contemporary relevance? There's so many. In the 1980s, the Reagan Department of Justice tried to enumerate all federal crimes. And the person who spearheaded that effort after two years of an effort gave up and said, like, it's impossible for a human being to do this.

16:47The example of present day is that, you know, there's a lot of reflection these days about why we have so much procedural bloat in a lot of our regulatory processes. So after we built out STARA, we partnered with the San Francisco City Attorney's Office, led by David Chu, really to explore these different use cases. We looked at a range of different things, but what we really zoomed in on were these legislatively mandated reports or reporting requirements that the Washington Post in one piece on this called a kind of congressional black hole in the sense that Congress has a propensity just to demand many, many reports.

17:34And of course, at a time when we're demanding more and more of the civil service, these can really weigh our civil servants down to complete reports that are not necessarily serving present day policy purposes. I think -

17:53Daniel Ho:So this is a situation, for example, where they're spending a big bunch of money on some activity and they say, oh, and we should, as part of our stewardship, we should have a report every year or every quarter on how things are going or what's working well and what's doing. But this accumulates now over a couple of hundred years of government. Yeah, the challenge is that you have things that were enacted like this 80 years ago and are not necessarily relevant in present day. And these things can consume a lot of time. In Neil Korsuch's book, he points to one report by the Social Security Administration that required it to write up a report on its printing operations.

18:33And he reports that that took 95 federal employees four months of time to compile a report that, for instance, included things like the serial numbers of forklifts and printers, all nicely curated for Congress to read. And what the Washington Post reports is very few of these reports are really good. And so what we did with the city attorney's office is really think about that process in San Francisco. We ingested the San Francisco Municipal Code and Resolutions. You might be surprised to find out that it's on the same order of magnitude as the U.S. Code.

19:08Daniel Ho:Oh, my goodness. It's about 16 million words. And so that's a really difficult thing, a difficult research process for any human being to undertake. And we ingested it into the Stara system. We're able to identify around 528 reports. And then the city attorney kicked off a consultation process really to identify which of these no longer serve a present policy purpose, which are obsolete, which are just not even useful to do, and proposed in a 351-page resolution to delete or modify over a third of these reporting obligations. And who is the decider on that? Is that something that has to then go to, I don't know, the mayor or the council or who says, yes, let's delete that stuff?

19:58Yeah, the resolution has been introduced and will go to the board of supervisors for consideration, you know, in the same way that sort of other pieces of legislation are considered.

20:10Daniel Ho:And before we move on to another topic, how did you know? You said something like 500 plus reports. How do you know if you didn't miss any? And did you find some that actually, when you looked more closely, they actually weren't reporting requirements? Yeah, it's such a great question. One of the things that we did in the underlying STAR paper is really benchmark this against known tasks. And what we were able to show, for instance, in the federal context where there have been these attempts to try to enumerate all of the reports, we're able to get a near perfect recall of known reports and then find thousands and thousands more.

20:49And then we sampled the additional ones to get a sense of the hit rate on those. And the hit rate due to the performance of large language models is actually quite high on those. And so it really allows you to focus the city attorney's time on the kinds of things that actually matter the most as opposed to being swamped with an array of false positives.

21:12Daniel Ho:This is the future of everything with Russ Altman. More with Dan Ho next.

21:31next.

21:31Daniel Ho:Welcome back to the Future of Everything. I'm Russ Altman. I'm speaking with Professor Dan Ho from Stanford University. In the last segment, Dan told us about the basic opportunities for legal AI, some projects that he's worked on, and the problems with hallucinations in law. In this segment, he's going to give us some specific examples about some of the regulations that they uncovered using AI. He's also going to tell us about the debate between open AI systems and closed AI systems. And he's going to end with some advice and the outlook for a chatbot that might replace your lawyer. Bottom line, not so fast with that plan.

22:10Daniel Ho:Dan, you told us about a really great study where you were looking through old regulations, old statutes, and you found all kinds of stuff. I can't help but ask, can you give us some good stories? Oh, sure. I mean, you know, so much of what the city attorney's office did after getting this list of 500 some of these reporting requirements is try to go through which one of these are obsolete and which ones of these are really still serving a present policy purpose. And we found a lot of what you might call regulatory sludge. Sludge. Regulatory sludge. I love it. The kind of stuff that really just exists.

22:50There may have been a good reason at one point in time to put it in there, but it just doesn't make sense.

22:55Daniel Ho:It's not lubricating the wheels of justice. It is doing the opposite. It is slowing down the wheels of justice. So I'll give you a couple of examples. One is that the director of public works has to regularly file a report on so-called fixed pedestal zones for newspaper racks. Those are quite literally pieces of concrete that were elevated to have to be able to sell copies, physical copies of the San Francisco Chronicle. Wow. That was a live issue at some point decades ago. But those have largely fallen into disuse. It's not a good use of the director of public works his time to create a long report on this.