The future of AI and the law
Law professor Daniel Ho says that the law is ripe for AI innovation, but a lot is at stake.
Naive application of AI can lead to rampant hallucinations in over 80 percent of legal queries, so much research remains to be done in the field. Ho tells how California counties recently used AI to find and redact racist property covenants from their laws – a task predicted to take years, reduced to days. AI can be quite good at removing “regulatory sludge,” Ho tells host Russ Altman in teasing the expanding promise of AI in the law in this episode of Stanford Engineering’s The Future of Everything podcast.
Transcript
[00:00:00] Russ Altman: 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. 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, coworkers as well.
[00:00:48] Dan Ho: Did a couple of studies of legal hallucinations and we showed that general purpose language models like ChatGPT, when faced with a roughly eight hundred thousand benchmark queries actually hallucinate sixty to eighty percent of the time.
[00:01:03] Russ Altman: Oh wow.
[00:01:04] Dan Ho: Um, 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, uh, between one fifth to one third, uh, of the time.
[00:01:32] Russ Altman: 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. Before we get started, another reminder to rate and review the show, particularly if you've found it useful or have learned something new.
[00:02:26] You know, large language models have gotten a lot of attention in the last couple 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 need 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. Well, Dan Ho is a professor of law, political science and computer science at Stanford University, and he is an expert at the use of AI to analyze legal documentation. Dan is gonna 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.
[00:03:36] Dan, what led you to study the law and AI together as a major focus of your academic work?
[00:03:45] Dan Ho: Well, gosh, I had long been interested in both of these topics, but honestly it goes back, uh, uh, even, uh, to growing up when my, uh, uh, parents had grown up in post-war Hong Kong, that generation had fled from mainland China. Uh, my parents moved to Germany. I grew up, uh, in Germany at a time when you had a generation of folks question the choices of, uh, the earlier generation. Uh, and um, I have this vivid memory of actually, uh, uh, being a young kid when the Berlin Wall, uh, came down. And so I'd always had the sense of, uh, how fragile our social institutions can be. And so it had this deep interest in really understanding, uh, our, uh, social institutions and government in particular, and then always had in mind, well, what can we do to really, uh, help make sure our institutions adapt and are able, uh, to, to change, uh, with the times.
[00:04:52] Russ Altman: 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, uh, many of us know that it's adversarial. Uh, we know that there are cases that are that serve as very important precedents. Uh, 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, uh, expectations of society. What would you say we need to know as we about to dive into some of your work?
[00:05:31] Dan Ho: Well, gosh, uh, there's so many things one could say, but I would highlight two. One is, you know, we might, and this will be, this will, uh, maybe be too dated to 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, uh, uh, relevant. Um, and then the second is that, uh, law is inherently contested, uh, due to this adversarial, uh, system as you, as you note. Um, 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, uh, setting where you can find the right Wikipedia article, get the right answer, and you can know whether that's, uh, a correct answer or not. Uh, uh, legal reasoning and legal argumentation, uh, is much more, uh, complicated and that has made, uh, really, uh, for fertile research, um, for AI.
[00:06:44] Russ Altman: And, and again, right before we get into some of the really compelling applications, um, one of the things we, many of us have played with these large language models like ChatGPT and others, um, 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, and I guess a high-level question is when if, if somebody goes into ChatGPT and asks for legal research advice, either a lawyer or a non-lawyer, um, out of the box, how, how are these performing? Do they, have they shown good, I don't know, legal instincts?
[00:07:15] Dan Ho: Well, we've seen remarkable advances in the ability of large language models, uh, to actually provide answers and have a lot of legal knowledge encoded in them. And I would say when you're talking about, uh, uh, querying a model like this for widely available legal knowledge, they actually do, uh, pretty darn well. Uh, the challenge is that it's precisely in the instances where you need lawyers, where the law might be uncertain because you are in a particular state that has a different rule or you have novel circumstances, uh, 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, uh, legal hallucinations and we showed that general purpose, uh, language models like ChatGPT when faced with a roughly eight hundred thousand benchmark queries, uh, actually hallucinate sixty to eighty percent of the time.
[00:08:18] Russ Altman: Oh, wow.
[00:08:19] Dan Ho: Um, on, uh, basic legal facts. Uh, then we did another study of legal AI when, uh, uh, sort of using, uh, systems known as retrieval augmented generation, where you try to retrieve the relevant legal document and then provide an answer. Um, those have substantial improvements but can still hallucinate, uh, at non-trivial, uh, uh, rates, uh, between one fifth to one third, uh, of the time. And so I think one still has to, uh, really, uh, watch out for the potential pitfalls of hallucinations. One theory paper, that's a general theory paper shown, unless the answer is directly contained in the training data itself, well calibrated AI systems will necessarily hallucinate, uh, due to the way that they're trained to generate.
[00:09:09] Russ Altman: Okay, great. So that, that, 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, um, looking at, um, large corpora of legal documents for, um, statutory, statutes, uh, or regulations that are of great interest to a lawyer in, in a certain case. Or are, um, something that needs to be purged from the record because they're out of date and anachronistic. So can you tell me about that project? What, what motivated it and what'd you do?
[00:09:40] Dan Ho: Yeah, we have, uh, well, let me start, there's, there's sort of two projects that are, uh, 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. In 2021, uh, California enacted a legislation that required all of the fifty-eight counties to establish processes to go through their property deed records and identify and redact them for what are known as racially restrictive covenants. Uh, what are those covenants? When we, uh, 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. Uh, and these have been unenforceable since the mid twentieth century due to a US Supreme Court case, but they still persist in many of these deed records because they run with the land. And so California enacted this piece of legislation, uh, which was well-meaning, but in a county like Santa Clara, that could mean that you have to sift through eighty-four million pages of property deed records that date back to the eighteen hundreds, uh, until we started our collaboration with, with the county.
[00:10:57] Russ Altman: Was it simply to, I, forgive me, but was it simply to identify these or to expunge and rewrite them?
[00:11:04] Dan Ho: Uh, both. To identify, uh, and preserve them for the historical records 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 rerecord the deed record to omit the unenforceable, uh, racial, uh, covenant.
[00:11:27] Russ Altman: Right. And this could affect, uh, house, house transactions every day because we all, anybody who's bought a house knows that there's these title searches and it's all very kind of, uh, I don't know, choreographed exactly. And you, you, this, this well-meaning rule I could imagine, might disrupt that choreography.
[00:11:45] Dan Ho: 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 ninety thousand pages manually to identify four hundred of these covenants. And that was the worry is, is this is just gonna take us forever. Los Angeles contracted with a vendor to do keyword searches, uh, uh, for eight million dollars to conclude the process in a matter of seven years. And we felt this is exactly a place where we can help build out an AI system. And so, uh, uh, we, uh, curated, uh, kind of good benchmark data set, um, 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, uh, five million of these deed records, uh, and actually, uh, accelerate the redaction process and bring it down, uh, to be able to identify these things in just a matter of a couple of days.
[00:12:52] Russ Altman: Can that be exported to the other fifty-seven counties?
[00:12:55] Dan Ho: We are very much, uh, interested in doing that and are currently working with several other counties, including San Francisco and Yolo County, uh, here to basically bring that technology, uh, to accelerate the process elsewhere in California.
[00:13:09] Russ Altman: And was part of your job to digitize these or were you blessed with previously digitized documents?
[00:13:15] Dan Ho: In Santa Clara County, we were blessed with previously digitized documents, but actually what we found is that the existing OCR, optical character recognition, uh, system was not great. And that's been another interesting kind of discovery in building out the system, is that actually the ability of multimodal models to do really good character recognition. It's, it's, what we found evidence of is almost a kind of leapfrogging of conventional custom OCR technology through multi-modal models that both had lower error rates and were actually cheaper than conventional OCR solutions. So we actually built that into our pipeline. So we basically ingest just the page images, do the full OCR process, and then have a, a built out, um, a fine-tuned, uh, large language model to actually be able to identify where on the page one of these racial covenants exists.
[00:14:08] Russ Altman: Great. 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.
[00:14:21] Dan Ho: Oh, sure. Yeah. What we, so realized, uh, after that is this is, uh, you know, one example of a pervasive problem in legal reform. Which is that so often what you're overwhelmed with, uh, if you're trying to engage in a legal reform effort or trying to, uh, litigate cases or, or do legal research, is just the sheer volume of materials in front of you. Uh, and reminded of then law Professor Ruth Bader Ginsburg, who hired a small army of Columbia law students to search for fifty-nine keywords across the entire US code, which in present days, over thirty million, uh, words. And they went through each one of these to read every single code provision to identify. All instances of gender bias in the US code. And that formed the litigation plan for how, uh, uh, Ruth Bader Ginsburg began to really, uh, litigate, um, and, uh, uh, reconceptualize equal protection, uh, in the United States.
[00:15:26] Russ Altman: Yeah. So it really all started with the data.
[00:15:28] Dan Ho: Yeah. And it was an incredibly time consuming, uh, research, uh, process. And so we, what we built out in this other research, uh, uh, project is a statutory research, uh, assistant system, we call it STARA, that actually is able to ingest something like the US code, represented in a way that we would teach statutory interpretation, uh, 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, uh, did.
[00:16:05] Russ Altman: But in this case, it's general, if I'm understanding, it's general purpose. So you're not specifying the kinds of statutes to look at. You're creating a platform with which a, a, a legal person could to, could do many different searches.
[00:16:17] Dan Ho: Exactly. Exactly.
[00:16:18] Russ Altman: So, and what are the types of searches that are, well, you already told us about these covenants, these, uh, uh, racist covenants. Are there other example searches that are kind of, of contemporary, uh, relevance?
[00:16:31] Dan Ho: There's so many. In the 1980s, the Reagan, uh, department of Justice tried to enumerate all federal crimes, and the person who spearheaded that effort after, uh, uh, two years of an effort gave up and said like, it's, it's impossible for, for a human being to, to do this. Um, the, uh, example of, of present day is that, uh, you know, there's a lot of reflection these days about why we have so much, uh, procedural bloat in a lot of our regulatory processes. So after we built out, uh, STARA, we partnered with the San Francisco City Attorney's Office led by David Chiu, uh, really to explore these different use cases.
[00:17:13] We looked at a range of different things, but what we really zoomed in on were these, uh, legislatively mandated reports, uh, 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. And of course, at a time when the, we're demanding more and more of the civil service, uh, these can really weigh, uh, our, uh, civil servants down, uh, to complete reports that are not necessarily serving, uh, present day policy purposes, I think.
[00:17:53] Russ Altman: So this is a situation, for example, where they, they're spending a 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, what's working well and what's doing, but, but this accumulates now over a couple hundred years of government.
[00:18:11] Dan Ho: Yeah. The challenge is that you have things that were enacted like this eighty years ago, uh, and uh, are not necessarily relevant in, in present day. And these things can consume a lot of time. Uh, in Neil Gorsuch's book, he points to one report by the Social Security Administration, uh, that, uh, required it to write up a report on its printing operations and he reports that that took ninety-five 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, uh, uh, to read. And what the Washington Post reports is, is, is very few of these reports are, are really good. And so what we did with the city attorney's office is really, um, think about that process in San Francisco, we ingested the San Francisco Municipal Code and Resolutions. Uh, you might be surprised to find out that it's on the same order of magnitude as the US code.
[00:19:08] Russ Altman: Oh, my goodness.
[00:19:09] Dan Ho: Uh, it's, uh, about sixteen 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. Uh, we're able to identify around five hundred and twenty-eight, uh, 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, uh, useful to do. Uh, um, and proposed in a three hundred and fifty-one page resolution to delete or modify over a third of these reporting obligations.
[00:19:48] Russ Altman: And, and, and, and who is the decider on that? Is that something that has to then go to, I, I don't know, the mayor or the, or the council or who, who says yes, let's, let's delete that stuff?
[00:19:58] Dan Ho: Yeah. The resolution has been introduced and will go to the board of supervisors for consideration, uh, you know, uh, in, in the same way that, uh, sort of other pieces of legislation are considered.
[00:20:10] Russ Altman: And, and before we move on to another topic, how did you know, you said something like five hundred plus reports, how do you know if you didn't miss any? Uh, and did you find some that actually when you looked more closely, they actually weren't reporting requirements?
[00:20:23] Dan Ho: Yeah, it's, it's, so, it's such a great question. One of the things that we did in the underlying STARA paper is really benchmark this, uh, 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. And then we sampled, uh, the additional ones to get a sense of, uh, the hit rate on those, and the hit rate, uh, due to the performance of large language models is actually quite high, uh, 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, uh, an array of false positives.
[00:21:12] Russ Altman: This is The Future of Everything with Russ Altman. More with Dan Ho next. 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 are uncovered using AI. He's also gonna tell us about the debate between open AI systems and close the AI systems, and he is gonna end with some advice and the outlook for a chat bot that might replace your lawyer. Bottom line, not so fast with that plan.
[00:22:10] 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?
[00:22:20] Dan Ho: Oh, sure. Uh, I mean, uh, you know, so much of what the city attorney's office did after getting this list of five hundred some of these reporting requirements try to go through which one of these are obsolete, and, uh, which ones of these are really, uh, still serving a, a, a present policy purpose. And we found a lot of what you might call regulatory sludge. Uh, so the, regulatory sludge.
[00:22:45] Russ Altman: I love it.
[00:22:45] Dan Ho: The kind of stuff that really just, uh, um, exists. It may, there may have been a good reason at one point in time to put it in there, but it just doesn't make sense.
[00:22:55] Russ Altman: It's not lubricating the wheels of justice.
[00:22:57] Dan Ho: It is doing the opposite. It's slowing down the wheels of justice. So I'll give you a couple of examples. One is that, uh, uh, the Director of Public Works has to regularly file a report, uh, on so-called fixed pedestal zones for newspaper racks. Uh, those are quite literally pieces of concrete that were elevated to have, to be able to sell, you know, copies, physical copies of the San Francisco Chronicle. Uh, that was a live issue at some point, uh, decades ago. But those have largely fallen into disusage. Not a good use of the Director of Public Works' time, uh, to, uh, create a long, uh, uh, report, uh, on this.
[00:23:38] Russ Altman: So, so let me ask about that. 'Cause, because when you were talking about it, I had, so do they just ignore it realizing, I mean, I'm guessing that they just ignore these requirements and they're not that worried about being taken to task, even though in principle they're in violation of a regulation. Is that the way to think about how a responsible bureaucrat deals with these things?
[00:24:00] Dan Ho: That is the tricky position you put bureaucrats in. I think if you, the Washington Post had this fantastic piece on it, where one agency official sort said, you know, we've been doing this every year. It's been taking us so much time. What if we just don't do it this year? And, uh, you have a ton of, uh, uh, staff members that themselves just described being inundated with these reports and using them essentially as door stops. Uh, and um, so, uh, nonetheless, you also have a lot of instances of agencies continuing to dutifully file these reports.
[00:24:34] Russ Altman: 'Cause it's their job. It's their job.
[00:24:35] Dan Ho: Because it's, it's a statutory requirement. So the Federal Reserve, I'll give you one other federal example. The Federal Reserve each year has to file a report on the Presidential Dollar Coin program. Uh, and uh, each year the Federal Reserve has filed this, but also noted that the Presidential Dollar Coin program ceased to exist in 2011. Could you please relieve us of this reporting obligation? And, you know, those are the kinds of things that really should be, by default, sunsetted, or at least up for reconsideration because it is just not worth the time of the Federal Reserve to keep filing something about a defunct program.
[00:25:12] Russ Altman: Okay, so my assumption that they would just do benign neglect is really a big assumption and that I shouldn't assume that that's, that these reports are actually, in many cases, actually still being created and written.
[00:25:23] Dan Ho: Yes.
[00:25:24] Russ Altman: Thus, creating the sludge. Was there anything else from San Francisco that you wanted to pass on?
[00:25:28] Dan Ho: Oh, I mean, we see provisions that are eighty years old. There's an eighty year old, uh, uh, requirement that the Redevelopment Agency file quarterly reports, uh, on things. And, you know, those are the, the kinds of things that really should be updated. We have a lot of reports on, uh, an entirely defunct, uh, uh, programs. Um, uh, but of course there are also reports that still are quite important. Uh, and uh, in some of those instances what the city attorney proposed to do is to, uh, try to actually consolidate where it makes sense rather than having a handful of different reports that are all about the housing stock. Would it make sense rather than having five different teams do that to have an annual housing inventory report that consolidates all of those reporting requirements. You have it in one place.
[00:26:14] Russ Altman: Well, well, thank you. Those, those examples were as, um, satisfying as I was hoping that they would be. I, I wanted to move to actually a more theoretical issue, and it's a little bit technical, but I think that, uh, we all need to think about it, which is, as you know very well, there are debates about these large language models and whether they should be opened or closed. Open models would be ones where it's transparent, the entire model, its weights, its architecture, maybe even all the data it was trained on. And I know those are very different things, are publicly known and available and then can be scrutinized and, and, and evaluated. Um, and then there are others who say, no, no, no, that is closely held intellectual property, the weights, the architecture, and the data. And, and this is, as you know, very well, very common for lots of the most famous LLMs where we don't have full transparency. Um, and I'm wondering as a lawyer who's, who's using these all the time, and, and, and you know, at some point a judge might ask a lawyer on what is your claim based? And can I examine it, or can the opposing counsel examine it? So, does this open closed thing, does it get on your mind and, and what's your attitude or thoughts about it?
[00:27:22] Dan Ho: Yeah, I, I think I have, think it is probably, uh, less salient, uh, within the legal system as it currently stands in part because, uh, these systems are being integrated into legal search provider systems that actually have a history of being fairly closed. Uh, if you look at systems like, uh, Westlaw or Lexis, they have just not historically made much available in terms of how search even operates.
[00:27:52] Russ Altman: And these are the research tools used by practicing lawyers every day to find the presidential cases and regulations.
[00:27:58] Dan Ho: Exactly. And it's quite rare, um, at least in those instances for, uh, uh, judges to want to know more about that underlying form of technology. That could of course, uh, change. Uh, there have been these, uh, incredible efforts, including by one of our own, uh, Stanford alums, Peter Henderson, to track all instances of legal hallucinations in cases. We have, you know, over a hundred and forty of documented, um, instances where there are filings with, uh, uh, legal hallucinations. And that, those are of course the instances that are really gonna raise the eyebrows, potentially lead to judicial sanctions and the like. But the more general question you're asking about open versus closed, I think is such a central one for the innovation ecosystem.
[00:28:42] In the last few years of under the Biden administration, there was a really pronounced focus and concern that open models, uh, could, uh, really raise, um, catastrophic risk. There's a particular concern in the Biden executive order on AI that really centered around bio risk. And there the concern is, well, if a model is capable of, actually, uh, uh, leading to the proliferation of bioweapons, maybe we should actually have an approach that favors more closed, uh, approaches. In a collaboration, uh, that involved quite a number of folks, but, uh, co-led, uh, by Percy Liang and Arvind Narayanan here, uh, Arvind Narayanan is at Princeton, we really, uh, sort of looked, took a look, uh, at, uh, the sort of evidence base for open versus closed.
[00:29:36] And really, we think in, in a kind of white paper that resulted from several convenings, we really, uh, tried to, uh, stated as what we called a kind of marginal risk framework. What we need to know is what is the marginal risk of open models versus existing technology, or closed, uh, models. And, um, uh, the evidence for that, uh, at least, uh, uh, at that time in terms of the bio risk was not, uh, at all, uh, uh, clear. Uh, there was a great study done by RAND, uh, where they actually had, uh, teams that were randomized to have access, either to the internet, to, uh, the internet, and a large language, uh, uh, model, uh, and to create a bio attack plan. And then those plans were, in a blinded way, scored by experts.
[00:30:28] And what that ran study found was, uh, one, um, there was no statistically significant difference, uh, in the, uh, bio attack plan between those with access to LLMs, uh, versus not. And number two, uh, the kinds of information, uh, that LLM systems were able to provide was not distinguishable from widely available information that you could get from the internet. And so we really have to be careful, um, and, and have careful assessments of, of marginal risk. Uh, what's also tricky here is that just as you know, openness is a spectrum. Uh, so Irene Solaiman has a really good paper on the gradient of, of openness. And so what that means is you can have a llama model on the, on one end, which some would say is not even fully, uh, uh, open because they're, they're,
[00:31:21] Russ Altman: And, and for background llama is the model that Meta, Facebook, uh, puts out for free.
[00:31:26] Dan Ho: Yeah. And, but there is, there is a kind of, uh, license that, that people do have to sign that has, has restrictions. And then you can have something that is, doesn't provide the model weights, but does allow you to do forms of fine tuning on a kind of platform. And, um, uh, one study by Peter Henderson found that actually all of those safety, uh, sort of alignment efforts that, uh, come along with a more closed approach, uh, can actually very easily be stripped with just a few cents of fine tuning. And so, uh, you know, as you're traveling that gradient, even something that is, it doesn't have the weights exposed, can still potentially actually, uh, present a, a fair bit of risk. And so I think what we really need in this, um, uh, ecosystem is much more analysis of, of both marginal risk and marginal benefit of open versus closed.
[00:32:14] Russ Altman: Great. Great, thank you. So in the final minute, I just want to ask you a basic question. Um, people are, uh, thinking about their legal issues and like Google, I'm sure people are hoping for a general purpose, legal chatbot. Maybe they might even pay money for it. What is, what is the future of a general purpose legal chatbot that could give people quality legal advice?
[00:32:37] Dan Ho: Yeah. Well, it's, uh, I think President Carter at one point said, uh, we are over lawyered and underrepresented. So many people, uh, are caught up in, in, uh, the civil justice system with insufficient, uh, legal representation. And that's of course the huge potential for, uh, this technology to democratize access to, uh, legal knowledge. Um, uh, I think there is a, a, a lot of potential there, but one of the, the misconceptions is that a general purpose chatbot will be an end-to-end solution to any and all, uh, uh, legal, uh, problems. Uh, uh, in some of the earlier studies I mentioned, right, these, uh, general purpose chatbots have a propensity to hallucinate at alarmingly high rates. And what we show in that study too, is they hallucinate more frequently for in exactly the types of cases that are likely to have underrepresented litigants, uh, at trial court.
[00:33:39] Um, things that don't involve, uh, uh, uh, uh, sort of appellate litigation. Um, uh, there, uh, instances or where, where these systems are really prone to error is if you, if you ask a question based on mistaken premises. They are, they exhibit a quality, um, that we, in that paper called model sycophancy. They're likely to just reify what the user said. And oftentimes underrepresented litigants don't yet know what the right question is to necessarily ask. And so that is a particular risk. Um, and, uh, I think in, in my view, uh, one of the real paths forward is to focus on specific use cases that are more tailored to the kinds of problems, uh, uh, that, uh, users need solved. Um, because in those specific settings, it's also much easier for us to evaluate both the benefits and the risk profiles, uh, of these kinds of systems.
[00:34:40] Russ Altman: So the chat bots may still come, but they will probably be high, tell me if I'm right, they'll probably be highly specialized and very narrow in their application domain. But in exchange for that narrowness, you might have less hallucination and more of a, a confidence that it's giving you good, good, good advice.
[00:34:57] Dan Ho: Yes, that's right.
[00:34:58] Russ Altman: Thanks to Dan Ho. That was the future of AI and the law. Thank you for listening. Don't forget, we have 250 or more back episodes of The Future of Everything, and you can listen to discussions on a wide variety of topics at the touch of a button. Please remember to hit follow in the app that you're listening to right now. That'll ensure that you're always updated about new episodes, and you never miss the future of anything. You can connect with me on many social media sites @RussBAltman or @RBAltman on LinkedIn, Threads, Bluesky and Mastodon. You can also follow Stanford School of Engineering @StanfordSchoolOEngineering, or @StanfordENG.
