Ok, but Would AI Judges Really Be Any Worse?

Transcript

This transcript was generated automatically and may contain errors — the audio is the record.

0:00:00 Clip: You must never, ever give up or give in. You must keep the faith and keep your eyes on the prize. That is your calling. That is you mission. That is moral obligation. That is mandate. Get out there and do it. Get in the way. Change happens because people care. The people have the power to make this life free and beautiful. To make this live a wonderful adventure.

0:00:38 Clip: I still believe with all my heart, America can be a place where no child is left hungry, no community is left behind, and no one gets told they don't belong. Country is always worth it.

0:00:57 Thomas: And welcome to Opening Arguments. This is episode 1103. your host, Thomas Smith, we've got a super interesting one for you today. A moment, I'll be joined by not only, of course, real life attorney, Matt Cameron, but I'll also be joined by another legal expert, Aziz Huck is one of the foremost scholars of constitutional law in the U S. an author. He writes for Politico, Washington post, other outlets. Lately he's been thinking a lot about AI and the law. When I think of that, I think of just, you know, oh, chat GPT helping me with my Thomas takes the bar exam questions and not doing a good job or something. But there's so much more to it. There are a lot of areas intersections of AI in the law that maybe you wouldn't think of and they're really fascinating, really important. Are sometimes opportunities, but also that we really need to worry about. Will be a fascinating conversation. Can't wait to speak to Professor Huck. We do, we'll take our usual break. Please support the show, patreon.com/law. the ads and support conversations like this. All right. After this break, we will be on with Professor Aziz Huck opening

0:02:04 Thomas: Aziz Huck, thanks so much for joining Opening Arguments. How are you doing today?

0:02:07 Aziz: Doing well.

0:02:08 Thomas: Excellent. Yeah, we're excited to talk to you. Question, we were just wondering how you got interested in the area of AI in the law. Not necessarily something you probably studied in school.

0:02:19 Aziz: AI has not been a topic for law or really for public policy prior to around 2010. And my involvement dates from about 2017, 2018. I was a cooperating counsel with the American Civil Liberties Union, the ACLU of Illinois, on a case involving policing in the great city of Chicago. One of the things that I learned in the course of that case, which unfolded over several years and involved at the end of the day, a consent decree with the police department concerning stop and frisk as a policy. One of the thing I learned was that the Chicago police department was using a predictive algorithm to decide who to stop. We dug into this. We looked at the records and we pushed the Chicago police department for disclosure about what kind of algorithm they were using. And what we learned is that they had taken an instrument that was designed by a professor of sociology at Yale that was intended to identify people who were likely to be victims of violence. The police department was using this as a way of predicting who would commit acts of violent crime. And, the theory the police department had for flipping this algorithm around was, well, we think that violent crimes occur within limited networks of people. We think it occurs between gang members, in effect. And we think that if we identify people who are victims we're also identifying people who are culprits now of course that has all sorts of uh slips

0:04:09 Aziz: Between premise and conclusion and we can talk about those if you want but to answer your question this this got me really interested in exactly how government in particular was looking at ai and algorithms and predictive tools generally as a means for either doing things they were already doing or taking on completely new projects that altered the relationship between the state and society so my interest really springs from that first encounter and kind of my sense of surprise at the the nature and the way in which the machine was being woven into this very human practice of Chicago policing.

0:04:56 Matt: I know you from your work with constitutional law, and that seems like actually a natural progression now that you're describing it, that you were looking at civil rights and constitutionality of this kind of thing. Because obviously, I'm sure you're going to tell us that there's no way this can be a neutral model if you're drawing from the kind of data they're talking about and the the presumptions and algorithms they're using, I assume that this would factor in things like what neighborhood somebody lives in and their wealth and occupation. Is that kind of stuff? Are we on the right track here?

0:05:22 Aziz: Think that's right. The original model drew upon largely data that was produced by social services agencies, as I recall, in addition to some data that had been drawn from public records of violence in Chicago. So the predictive model that was being used by the Chicago Police Department was one that in part rested were or reflected training data that itself, on the one hand, with respect to the social services data, embodied in some way some some of the disparities in living circumstances and opportunities that are woven into the built fabric of Chicago, which is a very racially segregated city. And on the other hand, wove in the historical biases that are present in policing in a big city like Chicago in the form of who gets arrested, who doesn't get arrested, what kind of instance of violence are being reported and what kind of instances are not. So I you're absolutely right to say that this, many decisions or many use cases involving AI or algorithms on the part of states and the part if agencies of the state, necessarily builds upon data that itself either reflects perhaps troubling distortions or troubling distortions in government behavior or patterns in a larger society that we might not want to replicate. The other point of the question was, isn't there a continuity between this and work on constitutional law? I wanted to point out that there is, I think, a pretty deep continuity in the following way. When American law scholars think about their often what they're thinking about is, well, what are the institutions that we, the people, at some point in time have fabricated to run, to exist independent of us as individuals, to run almost a sift on their own steam, not just for today, but for tomorrow and for the days after that. So topic of constitutionalism is durable institutions that in some way kind of run independently of the people who fill particular offices. And when you think about the topic Of Constitutional Law as essentially artificial machines that we build, and that's all that government is, it's an artificial machine, right, that's described in a few lines of law, which when you think about it, again, is analogous to the kind of code that is used to make machines. If you think of government as a set of machines that are meant to run beyond the lifespan or beyond the time in office of a particular person or group, then there's lots of similarities between the artificial machines that make up our government and the artificial machines that we now label as artificial intelligence. And absolutely, the Chicago Police Department example shows, I think nicely, or it presented to me nicely, as an instance in which those two kinds of artificial machines, one of which is very old and which we've grown familiar and accustomed to, and one of Which is very new, how those two kinds of artificial can be blended together. But I'd make in respond to your question, Matt, was that is we a lot of experience thinking about autonomous machines and the way that we relate to them and the way that they do things sometimes that we didn't predict and produce outcomes that we don't really want. That's really what constitutional law at its core is about. And to my mind, that project of thinking about machines extends really naturally to the context of artificial intelligence.

0:09:23 Matt: Makes a lot of sense. Oh,

0:09:24 Thomas: Had a question about that specific instance, because I'm wondering now, based on how you described it, it sounds like the problem, again, at least as described, wasn't so much with anything that a machine was doing, but was with the misapplication of this is looking for victims of crime rather than, you perpetrators of crime. And I'm wondering, what would the analysis have looked like? Like, pretend they did have a really good algorithm that was using all of the available, you accurate information to, to, guess, profile would be the word of potential criminals, but it was based on like good data in quotes. How would the diagnosis change there? For

0:09:59 Aziz: Think a premise of the question is that it's possible to write down an algorithm that you can apply to a set of training data in a way that produces accurate predictions about this particular kind of event. That's the basic structure of many artificial intelligence tools. You have, on the one hand, a setof historical data that you learn learn some principle or decision rule from, and then that decision rule gets applied to some new set of data, what's called out -of -sample data. And the presumption of your question is that that process of finding historical training data, using it to train up a new decision rule, and then applying that decision role in a way that generates predictably accurate outcomes is one one that is feasible to apply to the policing context, and in particular, the context of violent crime. The Princeton computer scientist, Arvind Narayan, has a terrific book called AI Snake Whale. And Narayan makes the point that almost all forms of prediction with respect to complex embedded species of social behavior. That is not stuff about how some biochemical system works, not something about whether a certain set of symptoms predicts this particular kind of disease in an individual, but rather how does this complex set of economic, cultural, and social conditions lead to this narrowly defined particular outcome, all of those kinds of predictive tools end up having extremely high error rates. Why? Narayan's argument is that, in truth, it is really hard to get a set of training data that actually captures everything that is going on in a complex, real -world social situation. And it's really hard to extract from that a decision rule that a machine can apply that's going to be applicable in a world that is ever -shifting, ever -evolving, and importantly, ever mutating in response to what it is that the people running the machine are doing. So I think that the right way to answer your question is to, with apologies, push back on the question by suggesting that we ought to think about AI applications in terms of those that are feasible and those that Are less feasible. And we ought To recognize that there's actually a class of things that we really care about, and I take violent crime to be one of those things, that are just not the kind of things that we can get AI to do. There might be other things where AI is a terrific solution, because the kind of training data that we have and the kind of prediction problem that we're trying to solve is amenable to the kind of critiques that Narayan makes. But at least violent crime is probably not one of them.

0:13:24 Thomas: Sure. I to clarify because I was wondering if that was the kind of thing that would have perhaps survived a legal analysis that the first thing you talk about, which is completely misapplying the category and looking for, you of crime rather than perpetrators of crime. If that would have maybe survived some legal analysis that real world thing possibly wouldn't. I I firmly agree with you that would I rather an not be doing this kind of thing, but just in terms of the reality of our legal system, I guess.

0:13:51 Aziz: I that it's helpful to start there by noticing how few challenges there have been to algorithms in the US court system. The most high -profile one in the criminal justice context concerns a risk prediction tool that's used in bail and in sentencing called the COMPAS algorithm. But beyond the challenge to COMPOS, and there was one in Wisconsin. There was one in Michigan. I think that there was also one in Florida, although I'm not certain of that. Those challenges, it's really hard to think of instances in which litigants have brought challenges to criminal justice AI tools. Well, why is that? One of the reasons is that many of these tools are used in contexts such as predictive policing, which is basically what the Chicago Police Department was engaged in, where there is no constitutional or statutory hook that allows a challenge. When someone is arrested because of a predictive policing tool, almost always there's going to be some independent legal basis for the arrest. Almost always there is going to some reasonable suspicion that justifies initially an investigative stop, perhaps also the necessary reasonable suspicion to justify a search, and then there's going to be some basis upon which the arrest can happen. So because of that, because the constitutional and statutory system for challenging exercises of the criminal justice power channels attention toward a relatively narrow set of substantive questions, we just haven't seen many challenges to these policies um of the things to kind of push again not to fight the question but to maybe sort of complicate its premises a bit is is just to notice that outside of the context in which an individual is getting a decision that is immediately being made by a machine it's actually quite difficult to think about or to imagine imagine legal theories in the American context that would permit challenges. So don't see that many. And the reason that we just Don't see that Many challenges is perhaps the paucity of such plausible theories.

0:16:11 Thomas: That's interesting because on one hand, it totally makes sense. On the other hand, I feel like the algorithm that is being used in stop and frisk in New York, to limited knowledge, The is darker skin equals stop and frisk more. And that was that algorithm was installed in the police officers brains. That seems like something that was challenged, even though I the same arguments apply there and how it's used. But there was a way if you show enough evidence that like, hey, the underlying, you I guess algorithm again, the underlying method here is is know, racially biased, et cetera, et cetera. Is there a way to get at it from that angle, perhaps?

0:16:44 Aziz: When New York's stop and press policy was challenged in federal court, the basis for the challenge was the protection clause of the US Constitution. And the theory of the challenge was that the policy was based upon a discriminatory motive. It was based upon a motive. The lawyers in that case, which was called Floyd, were able to demonstrate the existence of a racial motif by pointing to statements that were made actually in the course of the trial, but also that were brought in as evidence in context of discovery to the effect that the precinct level officers in charge of directing Stop and Frisk told their rank and file police sergeants or police officers, go out and arrest people of a different color scheme. So So there was evidence there that there was an intention in the mind of a particular person that had a racial character. Now, in the context of a police department that uses a predictive policing tool of one sort or another, and almost all big cities use predictive policing tools, and there have have been many critiques of these tools as being racially coded or invidious in some way. Almost all of these contexts, there's nobody you can point to. There's no smoking gun evidence that you can Point to a particular individual who has a racial motive in mind.

0:18:18 Aziz: It's almost never the case that the designer, for example, has a Racial Motive in mind, it's almost always is the case that the police department can say, well, we're adopting this because it's going to make policing more efficient. We're going to be able to target our resources more narrowly and more effectively. And those claims are going to not just in many cases on their face plausible, But often they're going to be, in fact, true in some deep way. It that the operation of these tools can have effects that are racialized in character. But showing that a policy has racialized effects does not itself satisfy the equal protection standard for a discrimination violation. You have to also show intent. And in the context where machines are being used as opposed to humans to direct policy, that showing of intent is much much harder to make.

0:19:26 Thomas: Wow. That's devastating. It's like they're offloading their mens rea to a machine and something, you know. Wow. Great point. Is brought to you by rocket money. I use rocket money I've used it for a long time. Rocket Money is incredible because nowadays it feel like everything's a subscription. Every single thing is a subscription You used to buy stuff, now that doesn't exist anymore. Which, which fine, you know, it's good in some ways, but a downside is you often sign up for stuff and forget you did, or you signed up for a trial and then it ends and lose track of that. And then you're paying for something you didn't want. Rocket money can help you with all that and more because rocket money is a personal finance app that helps find and cancel your unwanted subscriptions, monitors your spending and helps lower your bills. So you can grow your savings. Got a great dashboard that gives you a clear view of your expenses across all your accounts. And that's key because I'm liable to sign up for different things in different accounts. That's a good way to lose track, you up for the same thing across two and sometimes even three different places in accounts. Done it. I have absolutely done it and Rocket Money will We'll sort that all out for you. Even have a thing where they will try to negotiate lower bills for you They scan your bills to find opportunities to save and you can ask them to negotiate for you something I would never do on my own. Deal with the customer service so you don't have to. Rocket Money has over 5 million users and has saved a total of 500 million in canceled subscriptions saving members up to $740 a year when using all of the app's premium features. Cancel your unwanted subscriptions and reach your financial goals faster with Rocket Money. Go to rocketmoney.com/opening today. That's rocketmoney.com/opening. there right now. One more time. rocketmoney.com/opening You'll

0:21:12 Matt: Me to what you're saying about sentencing and recidivism, which is what we're talking about with risk assessment, which I know AI has been used somewhat, at least in some jurisdictions, for that. And from what I remember, that did cut cut down on things like gender differences in sentencing. I think there's some kind of issue, and again, I'm just working from memory here, that the judges would ignore when AI was recommending lower sentences and racial biases persisted. And there's also, of course, the fact that if the AI, as we're talking about, is working from the data pool that it's working from, that it's still going to think that maybe people from certain neighborhoods or certain backgrounds are more likely to be recidivists in the same way that a human judge might use that bias. So what's been going on with sentencing

0:21:49 Aziz: The sentencing context, the tools that tend to be used are not technically AI tools. So for example, the compass tool that's used in many jurisdictions is a hand -coded tool that's based upon 137 different variables that are each given a certain weight. And just like in the regression equations that some people may have encountered when they were taking taking college statistics or college math, you add up those variables discounted by each of their weights and you come up with a prediction for a given person. That's not really machine learning or AI, that's just prediction. And prediction tools of that kind have been used in the sentencing context since the late 1960s, early 1970s. They've just become much more more sophisticated as time has gone on. Have been challenges to those tools on the basis that they explicitly take account of gender. And they explicitly take a counter gender because it turns out that gender is highly correlated, in fact, with recidivism risk or the risk of violence. Seems pretty clear, even looking at the raw data of offending rates by and by women. Case here is the Wisconsin case I mentioned, which is called Loomis. And court there declines to find that there's an equal protection violation because it says that the use of gender there is justified, although it cautions that the reason it's doing so is that the is advisory rather than mandatory. Other part of your question is about how do judges take up and use the algorithm's outcomes. Work here is done by a really wonderful data scientist and empirical legal scholar called Megan Stevenson, who is based at the University of Virginia. Virginia, and Megan's work finds that when algorithm is used, there will be a change in judges' behaviors, but often over time, there'll be a return to form. That is, you'll see some change, but there's a kind of bounce back that occurs over time. That's one study. There are probably other studies that are going to be done. There are probably others that have been been done that I don't know about uh but the the general point is that you that each of these algorithms all of the ones that we've been talking about are tools that are um necessarily in particular human context they're a particular institutional context most of us have been in workplaces we know that each workplace has its culture each of those culture produces different kinds of attitudes, beliefs, practices, or patterns. Something new down into a workplace, and it's going to be shaped, woven in, in different ways in different workplaces. And that's just as true for courts as it is for a company or a papermaking company. And it's just as true for algorithms as it is for any kind of new technology. I one has to be really careful about generalizing, but at least the studies that I'm aware of suggest that there are unpredictable effects of putting in these tools into policing and judicial contexts. Mm

0:25:14 Matt: That makes a lot of sense actually that humans would take over again but I actually wanted to ask as well kind of following up from that about places where systemically we might start to see ai taking over I'm not talking dystopian necessarily although it can get there uh in massachusetts where I live we've been fighting uh there's possibility this might actually pass at some point but uh the idea of automated traffic enforcement right that you can get a ticket in the mail from you know blowing the red light or if the toll booth notices you got somewhere you know where the only could have been speeding And of course, is pretty benign, but then it works its way up to AI prosecutors in China making decisions about charging people with no human involved. I think there's something like eight different offenses they were experimenting with. It's something with a 97 % accuracy rate, allegedly. But includes things such as provoking trouble, which is a term for dissent in China. So you can have an AI prosecutor bring charges, assemble the evidence and put together a case I mean obviously that is extremely dystopian but I'm hoping we're nowhere near that but provoking

0:26:08 Aziz: I heard that running a a liberal podcast is soon going to be enough to get you on the making trouble list so you should be careful I

0:26:15 Matt: Trouble for sure we're

0:26:15 Thomas: Oh, yeah. Yeah. Yeah, we'll try to still conduct interviews from the prison phone line. But yeah.

0:26:16 Matt: Trying to do our part I hope

0:26:21 Aziz: I here they'll have uh they'll have recording equipment in prison libraries I

0:26:25 Thomas: Good. Revolution will be podcasted.

0:26:30 Aziz: Think that's a terrific example and I think it's a a thought -provoking one and it raises the following question in any area of of regulation, there are things that are mechanical, decisions that are relatively reflexive in as much as they just depend upon a readily available stock of information. So, for example, the decision as to whether a person is speeding, the position of a person, the decision is to whether somebody turns right on a red light at an intersection where right on on red is prohibitive. Those relatively, if I can call them mechanical decisions, sit side by side with decisions that are a good deal more discretionary. So anyone who spent any time looking at the criminal justice system knows that this is a system that's shot through with judgment calls by police, by prosecutors, and by judges about all sorts of things. So for example, we all know that if a police officer comes across a group of kids who are being unruly on the sidewalk, the police officer has to make a judgment about whether what those kids are doing counts as whatever the jurisdiction's equivalent of disorderly conduct is. So we have a system that has these mechanical decisions and these highly discretionary decisions. And there's lots in in between, and it's really hard to draw a clear line between the mechanical and the discretionary. And so we're at a point where the technology exists to permit us to automate the highly mechanical decisions. And maybe we'd be better off in all sorts of ways if we did that. Maybe we'd better off if we didn't. Maybe we'd do that for the purposes of reducing the amount of speeding, which has enormous human costs, especially in the United States, is which a a country where I think the average car is about one and a half times as heavy as the cars in europe and therefore it's much more deadly than cars in in other

0:28:32 Aziz: Parts of the world I

0:28:32 Thomas: Has one and a half times the freedom, though. You didn't account for that benefit.

0:28:35 Aziz: Didn't I didn't can can you can you fit a a surfboard in that in all the space the freedom gives you I don't know um you have serious social costs that can be mitigated deaths that can be avoided by automation. On the other hand, it's really hard to know, well, go from automating essentially those mechanical decisions to automating things that really require discretion? And it's Really hard to say up at the front end, okay, these are the things that are mechanical, these things that have not. And we as a culture and a society just haven't really put much thought into that. And I don't have a way of drawing clearly. And what your example shows is that if you go from Massachusetts to Beijing, for example, the way that people think about that line changes dramatically. The cultural significance of the mechanical as opposed to the discretionary is really different from one place to the other. And people's tolerance for losing control, losing that discretion is really different in different places. So I think you're absolutely right. There's this very hard line to be drawn. We don't really have a way of drawing it. And even if we did have a great way of drawing it, actually the value choice that we're making when we draw that, when we decide what kinds of discretionary decisions are just mechanical enough that we can put them in the hands of the machine without further review. We just don't really have the vocabulary or the sophistication as a legal culture, I think, yet to think through those things.

0:30:12 Matt: Such a good way to put it and I think I'm coming from a defense mindset as an attorney so I guess to me you know obviously I recoil as I think a lot of people do with the idea of automatically charging anything and certainly the idea of just by rote charging something when you know because to me a free and democratic society does not charge everybody all of the time for everything they could possibly charge them for um think that's a pretty fair statement and I think this that's what I'm really concerned about is if you leave those things in the hands of machines this actually brings me to what I thought was one of your most interesting pieces and points about the right to a human decision I've been thinking about for the last couple of days and the idea that maybe we could start to legislate or constitutionalize, and I know that wasn't necessarily your point, but that is kind of the implication I guess is we have to do that the future, the idea that you have a right to actually have a human review your case, whatever that is, whether it's for welfare benefits or whether you're going to prison think that's something that is going to be in the conversation? Think it should be

0:30:45 Thomas: Interesting.

0:31:05 Aziz: Certainly think it's part of the conversation because it's already present in other

0:31:09 Matt: Hmm

0:31:10 Aziz: Countries' laws. So for example, the European Union's general data protection regulation includes something that can be characterized as a right to a human decision. The AI Act that the European Parliament passed earlier this year in August contains a set of prohibitions on what are called social scoring systems, which are systems like the Chinese social credit system that assign people a numerical rank and then deny or grant them access to certain kinds of benefits on the basis of that ranking. And I think that the intuition behind that prohibition in the EU Act is precisely, no, no. You can't be denied access to, let's say, credit because of this number. You have to have somebody actually think this through and offer you a decision to your face that accounts for the information that you can give to that person down in there. That. The article that you mentioned, which was published about five years ago now, I come to the conclusion that the standard arguments that are made on behalf of the European right to a human decision don't run through. And I think those arguments are based on the one hand on accuracy and on the other hand on human dignity. And that's right. I continue to think that provided provided a machine is sufficiently accurate. We've already talked about moments at which, or applications of AI, where the application really is beyond the domain of what can be done with an AI, and therefore we can't meaningfully talk about accuracy. Provided an AI is capable of resolving a particular task, it's almost certainly going to be, at least in theory, better than the closest human analog. And I'm not persuaded by the arguments that you benefit people in terms of their dignity by putting them in front of another human being like a judge or a probation officer or the like and in part that that is based upon a sense from having been an attorney in a different not a criminal defense context but a slightly related context uh but a sense that in many moments when people are confronted or placed in front of authorities, particularly criminal justice authorities or welfare bureaucracies or the like, those are moments of deep humiliation and deep psychological harm. And replacing those moments with a less personalized and more impersonal alternative will often ameliorate real harms. So I don't think that the arguments for the European right to human decision, which are based on accuracy or dignity, really work. However, since I've written that article, there's a line in that article that I think should have been a paragraph or a section. That there is a reason that's linked to something that was said earlier by one of you for there should be a right to a human decision. And it's this. Of the consequences of moving from a human to a machine decision -making system in criminal justice or in many other contexts is you dramatically reduce the cost of doing something. You make whatever it was the state was doing much, much cheaper. Why? Because, as the economists would say, the marginal cost of extending a machine from doing 100 things to doing 101 things is essentially zero. Whereas the marginal cost of a or a parole specialist going from 100 cases to 101 cases is actually meaningful. So the marginal cost of expanding the scope of a program drops dramatically as you go from humans to machines, which makes it much, much easier to scale up some sort of a governmental program. And in the criminal justice context in particular, as critics of mass incarceration have pointed out for decades, the very fact of the scale of certain kinds of governmental systems, systems, particularly ones that coerce and imprison and inflict pain on other human beings, the very fact that we can do those things at scale is a profound moral problem. So to my mind, there is a really good argument for the right to a human decision, but it doesn't touch upon any discrete personal interests that you or I have. It has to do with the kind of architectures of control that are possible if we have machines, as opposed to human decision makers.

0:36:00 Thomas: Yeah, I mean, we already have mass incarceration. How much more mass? Finally, we can automate mass incarceration and get mass squared. I don't know. Dystopian yeah

0:36:08 Aziz: True that the United States has one of the highest incarceration rates in the world, but again, I don't mean to make China the bogey person of conversation when it comes to technology. I think that's far too easily done, and it often has a kind of racialized cast. Nevertheless, China has automated the policing of dissent, particularly by its Uyghur population, the reports of facial recognition tools that are coded to recognize Uyighur faces, The reports of location and interventions based upon the physical movement of individuals that keep them within relatively narrowly defined physical spaces, I think, and so for very large populations, populations in the tens or hundreds of thousands. Again, in the case of the Uyghurs, I suggests that there are forms of automated population level coercion and control above and beyond those that are currently used in the United States that are profoundly morally troubling. And however morally troubled you are by mass incarceration, the that our system of mass incarceration is open to lots of moral critiques, as I think it is, shouldn't foreclose or shouldn't blind us to the fact that, well, there are actually worse versions of this out there that you can imagine once you layer in this new kind of technology.

0:37:49 Matt: Blood was getting a little colder as you were saying some of this, I was thinking about this morning I had an asylum hearing as I often do with a client who was talking about the worst thing that ever happened to her which is what you do in an asylum hearing for the most part and you know she broke down as people often do when they're telling these stories I was thinking about what you're saying about the profound sort of emotional consequences and humiliation of having to share something as personal as what she had to talk about with an immigration judge but I was also thinking don't know I guess I'm still sort of processing this but but the idea that that could ever be done.

0:38:18 Thomas: If you're talking into deportation bot 1000 and it's like information received you know like that kind of thing it's even more heartless

0:38:24 Matt: It's right? The computer would be like, well, I See You're Crying, but you clearly, this wasn't persecution, what you're describing, or, you know, I don't think it meets the case law. And I know you're not suggesting that we replace immigration judges. I don'T think anybody is, I hope. But I could see the future where we get there, and that is terrifying. And

0:38:40 Aziz: That's right although notice again how much you turns know upon what you think of immigration judges I mean and and I mean there is a there is A View of the World and you're clearly you've got direct experience of immigration Judges that I don't have but can imagine A View Of the World where which says well look you know the the people who take these jobs the people who are doing these things are cruel. I don't know whether that's our world, but it's certainly an imaginable world. Maybe it's a very closely, maybe it's a world that's close to come. But you can imagine a world in which the people who administer the law, who are making those judgments, are doing so in a way that actively inflicts pain and that in no meaningful fashion gives people a sense that their humanity is witnessed so so much depends upon the quality both of the people but also the quality of the machines because I think that part of the force of your concern is that is that somehow we think of machines as insensitive to information which is very peculiar it's a very peculiar trope in the artificial intelligence context because surely one of the things that machines seem to be able to do much better than human beings is account for complexity and not folk home in on one thing and lose sight of everything else. I completely hear the concern. I think the concern is one that is very It's hard to know whether replacing humans for machines in that or other context will in fact make things better or worse. But I think that thinking thinking about that, you know, what exactly will be the change and how exactly will the status quo be altered is a really important set of reflections and considerations.

0:40:26 Matt: That was actually one of the most interesting things that I thought you were saying in that piece was that we shouldn't assume it's necessarily worse course, because we do. Like you said, the tropes are very much there, that computers are insensitive and unfailing. And, you know, I'm certainly open to that. I will say I'm practicing immigration law in Boston, where our judges are generally very humane and coming from the local population. But it might be a very different situation in Central Texas, as it is. So it's very true.

0:40:47 Aziz: Think that the most powerful examples maybe that are counter to the thought that you have there come from non -governmental contexts. Let me concede that they're non -governmental contexts, but the two examples I would give are first, dating apps. So I'm old enough and have been married long enough that the idea of entrusting my romantic future to a machine still to me kind of crazy. It still seems, to me, utterly alien. The end and yet we now know that dating apps are now the dominant uh through which people form lifetime relationships whether it's marriage or something else and I think if you had told people that 30 years ago they would have said that's insane I would never let myself I would ever give over my emotional happiness my most important human decision I make as a person in our society who pair bonding through marriage is the norm I'm not going to give that up to a machine and yet here we are and so that's an example where I think I that there there at one point in time would have been a very strong reaction and it's gone away and I I can only really report here what I what I hear from my students. But my students, when I express this instinctual emotional reaction with respect to dating apps, they look at me incomprehendingly. They're like, what planet are you from? Why would you even think this? So I'd observe that in the course of a very, very short span of time, I think a very basic instinct with respect to what I all concede is an important decision in the human lifespan has been radically altered. Other example I would point to is Japan has a shortage of care workers. And this is particularly important in Japan because the population in Japan has been more elderly since the late 1980s, early 1990s. And so you have a large elderly population and by no means enough people to take care of them. But one of the success stories in Japan, and it's an interesting success story to think about in relation to immigration, for I'll say in a moment, is the use of robots to provide not so much care, but companionship to the elderly. And robots in the context of providing companionship have been overall a success story in Japan hat, that emotionally cathect, to use a word that comes from therapy, they bond with machines and they form emotional relationships and they don't have the kinds of loneliness and the kinds emotional damage that you would see if they were alone. Now, the interesting connection to immigration here is that Japan very much had a decision to either produce domestically companionship for the elderly or import it to bring in workers from the Philippines, from Thailand, from countries that are close by with not too different cultures. They chose to make, not to import. And so here's an interesting case where technology is actually, I think it really is a substitute for immigration um in this way that I again 30 years ago if you told me well you know you will have a robot taking care of your parents I would have said that's crazy but clearly I'm wrong or I would Have been wrong no

0:44:30 Thomas: There's this short uh story today, I Am Paul uh I don't know if you are familiar with it but it's a very very interesting short story uh on exactly those those lines. See, won the Nebula Award in 2015. But yeah, really interesting concept there. In the case of dating, I did meet my wife through OkCupid. And what's funny is I'm a millennial, so I have like, Whereas you might have the skepticism of using any technology in there, I have, like, the millennial version, which is... Okay, but back in my day, it was an elaborate elaborate questionnaire that the computer did that matched you up based on criteria that makes a lot of sense. Now it's just an app and you just like put a picture in there and then, you know, so it's like different levels of the same like skepticism about it. But ultimately it is just a tool because it's not as though yet the machines are forcing you to mate with anyone or to marry anyone. It's still just a tool to meet people, which I think is interesting in the context of of this conversation when you talk about AI or whatever it is as a tool to aid maybe in decision making sometimes. But I guess with the waning time, I did want to ask, and maybe this last answer was kind of that, are there areas in the law where you are optimistic about AI, or where you think it's going to be particularly helpful? Is side to some of this that we haven't talked about yet?

0:45:07 Matt: Swipe. Yeah, Yeah, yeah. That's so true.

0:45:46 Aziz: Think that many people experience the beneficial side of automation in the law when it comes to doing things like writing a will or creating a power of attorney either for themselves or for an older relation. And the software that's used to produce legal documents through things like LegalZoom has become much more sophisticated Over the last decade or so. And I think it is an unmitigated blessing that people my parents' age can go online, click a few buttons, and those really important legal documents that provide them with a lot of comfort and are very, very important for organizing their lives in space of minutes. And maybe at the cost of 30 or 40 dollars rather than having to take time go and see a lawyer pay 500 for something that um know will take months to do um think that that that kind of automation is is it's very hard for me to see uh downside from that I

0:47:02 Matt: Actually, I have one more question about a use case, I think arguably is fairly positive. If you're familiar with Kevin Newsom on the 11th circuit on his concurrences uh

0:47:11 Aziz: I know who he is.

0:47:12 Matt: He's written a couple of really interesting concurrences about, and one was in a civil case and one in a criminal case about what justifying why he was turning to chat gpt and a couple other llms uh look into the ordinary meaning of words and I just thought that was such an interesting and elegant way to do it and he wasn't saying I'm relying on this finding he's just saying this is one more tool and it's actually more effective than the dictionary because for uh expressing like physical restraint right which isn't one word it's two words it's it's you know he wants to get a sense of how that is used in ordinary language and that to me seems like the right direction I'm still sort of turning it over I read those concurrences last week um I I'm sorry if you're not familiar with it I didn't mean to spring it on you um I just want to get your thoughts about that sure no

0:47:51 Aziz: A little skeptical of the idea that there's something called ordinary meaning in the abstract that is detachable from the the specific context in which a phrase or a word is used. And I would worry that the use of tools like ChatGPT, actually like the use of dictionaries in legal opinions has a kind of aura of objectivity and restraint. But as we now know with dictionaries, it really opens up a set of avenues of pretty unrestrained discretion on the part of judges that can be camouflaged because you have a citation to a dictionary or or perhaps now a citation to chat GPT at the beginning of the relevant sentences?

0:48:42 Matt: It's true he does a pretty good job of getting through the possible downsides of it but I just think it's an interesting conversation to have I think I do come down on that too we'll

0:48:50 Thomas: I suppose starting with the premise of we want to figure out the ordinary usage. Idea that like, well, these large language models are literally looking at all of the usage. And so potentially it could be a good tool, but yeah, I share the skepticism of that methodology in general. But what great conversation. Is there any way people should go to follow your work or read more of your work or anything like that?

0:49:13 Aziz: Like many academics, and do most things through formal papers that are subject to some sort of scrutiny, and I can get them read and challenged by colleagues. And I post all of those on a website called SSRN, which is, I think, the pretty standard place where academics post in the legal academy.

0:49:32 Thomas: Excellent. I

0:49:32 Matt: Be sure to link to some of those relevant pieces

0:49:33 Thomas: Where we found you. Yeah. Excellent. About your book or anything? Is there anything you want to plug?

0:49:39 Aziz: Just published a book called The Rule of Law, a very short introduction, which is as the title suggests, suggests an extremely short introduction to the history of what turns out to be a pretty complex and contested topic that has particular meaning in our legal context in the US and in the UK, but it's turned out to become globalized. And there's now different strands of the rule of law around the world. And the book is an effort to introduce a broad array of readers leaders to the origins and maybe the future of what we mean when we talk about the rule of law.

0:50:15 Thomas: Fantastic. Professor Aziz Haq, thanks so much for joining us. This was a pleasure.

0:50:19 Aziz: Thanks for having me, guys.

0:50:20 Matt: Enjoyed it thanks again

0:50:21 Clip: This podcast is a production of Opening Arguments Media, LLC, All Rights Reserved. It is produced and edited by Thomas Smith, who also provided the fabulous intro and outro music used with permission.