Prediction Markets and Decentralized Courts: A New Architecture for Online Justice

Transcript

Hello hello. How are you? I'm Federico Ast. I'm founder and CEO of Kleros. This is like my eighth talk at the TCC. So, I'm super happy to be here again. Um, we're going to speak a bit about prediction markets and the law today. Um, I don't know if anyone recognizes this place. Is there an Argentinian in the room? But, this is the University of Economic Sciences of Faculty of Buenos Aires, University of Buenos Aires. And, um, this is where I studied. And, after studying, I stayed, uh, for more time connected as a, professor.

And, in the early 2010s, I was teaching, uh, very obscure subject called Epistemology of Economics. And, this is about how people, I mean, uh, create economic knowledge and how we understand economic behavior. And, in that, uh, course, I used to teach a text by Hayek, um, Friedrich Hayek is an Austrian economic economist. And, um, here I had this text about the use of knowledge in society that he used to explain how people make transactions, they manage information in markets, and how this is better than centralized options like central planning was the um, word by then.

And, this got me interested in researching more to give my students a bit of, uh, more tools. And, I got, uh, to know this, um, text about collective wisdom that was edited by Elan Landmore and John Elster, uh, who are experts in collective intelligence. And, in this text, the first chapter was about some, I mean, obscure also, um, mechanism called, um, prediction markets, um, by a professor called Emile Servan-Schreiber. And, this is the first time I got to to meet, I mean, to meet this. And few time ago, um, time later, I, wrote this at my This might be the first blog post ever to be written about

prediction markets in Spanish. Mercado de predicción and I the thing is I wanted to talk about political this was about how you can use these new tools for predicting the results of elections. This was look July 19, 2015 and a few I think a a month later I published the first version of the article. I made an edit to say look there's going to be launched a new prediction market called Augur. So this was like in August 15. So look at how how early this was and uh note to self I buy more more Ether if you come back to that moment.

And so prediction markets what are they? I mean I don't need to explain much in this audience because I guess you know um this is about speculating on future event and you can buy assets that represents um outcomes and there is financial incentives and this can be used to produce information that can be very accurate for predicting different different things um and then I kind of forgot about prediction markets for a for a while um got to uh be founder of Kleros so I got away from prediction markets. Kleros is a dispute resolution mechanism that can be used by people that make for example a freelance um agreements. This woman she's in France she hires someone from

another country and if there is a dispute in the uh contract there's a jury picked by Kleros and this jury is going to make a decision on who who is right about the case. And there is a token that is used for coordination called Pinakion and if you don't like the idea of freelancers think about agents interacting with themself and a jury of agents resolving the dispute. This is certainly what is what is coming. And um I got back again to the prediction market space as Kleros became oh uh sorry an oracle for prediction market disputes. There was by by then a prediction market called Omen that made different predictions about different markets and different topics and uh Kleros was

the tool, the method that was used for resolving disputes in case there was there was one and there was one of the most famous cases of um all time that Kleros was Do you remember that the election between Trump and Biden? There was a election and then there was a dispute about who won and then Kleros had to decide this case and this um well Kleros decided that Biden won and there is this famous tweet by Vitalik about Kleros proving its sanity by correctly deciding that that Biden won the election. So, this was my first, I mean, coming back to prediction markets after uh leaving them for a for a while.

And then 2021 in my talk uh in at ECC but back then uh I also came back to prediction markets but in a different context. In this case, it was about um content moderation. So, by by then we had um this idea of well, the rise of fake news, the rise of um Do you remember COVID and all of the vaccine and well, there was lots of fake news in online. And that was the topic. And um one of the main things that happened at that moment was about the time that Twitter had um well, uh banned Trump after a number of of tweet. Well, you remember that. And um a question I asked myself is can

prediction markets help create an efficient and transparent method for content moderation? A method that is credibly neutral and that's not controlled by a single agent. And there was this possible methodology or this possible mechanism where you have people making posts in a social network. They make a post and they can deposit with the post. And people can review uh the post uh they can upvote it if they think the post um complies with the rules if and they can downvote it if they think it doesn't comply with the rules. And then, um upvotes and downvotes are like bets on whether the moderator would find the post acceptable if it had to to evaluate it, in case there was um to be a moderator decision.

And you can see uh some examples from this is what it would look could look like, an impersonation. Uh this is what it could look like, like a some scam post on some social media. And people upvote and downvote. And this is kind of a a prediction market where upvotes are like bets on uh the the post complies and downvote does not. And most of the times you don't need to have to resolve the case. Most of the times you don't need to go to the moderator. Uh but sometimes you will have in cases that have many upvotes and downvotes, cases that are very polemic, this could go to Kleros and Kleros will make a decision. And uh jurors will evaluate it. And then, uh if Kleros downvotes the post, people who uh had downvoted the post in the um

original uh moment, they um get money from the submitter who did not comply with the rules. And the people who thought the post was like um uh acceptable, okay? So, it was like a prediction market. If you want to learn more about this, you can learn uh read this post by Vitalik, prediction markets for content creation DAOs. This is uh this mechanism is based on on that post that Vitalik wrote a few years ago. And then, we started having prediction markets becoming like more and more massive as you what uh a bunch of new projects came around. And this is becoming now way more mainstream than what it was when I started with uh studying this type of mechanisms.

And some people are using them to like predict some legal events. Like Supreme Court uh rulings. These are still high-profile things that people are paying a lot of attention. But can actually can we use these mechanisms to actually improve how law works? And I might be surprised that this is not a very new topic. There is this gentleman called Oliver Wendell Holmes that he's a um uh was a justice of the US in the early 20th century.

And he was a legal scholar, one of the most important legal scholars in the history of American law. And he has this very interesting book. It's based on a talk he gave called The Path of the Law. It's from 1897. And in this book he proposes the idea of the predictive theory of law. For Wendell Holmes, law is about the prediction of what courts are going to decide in different cases. And he has this very famous uh quote, "The prophecies of what the courts will do in fact, and nothing more pretentious, are what I mean by by the law."

And this gave rise to what we know of the realist theory of law. This theory uh is not interested in that much of the moral aspect of law. It's more interested in how um the law works uh and the results of trials. And it's based on how bad men see the law. A bad man, he has no moral conscience. He only cares about thinking, "Okay, if I rob this bank, what's going to happen to me? Well, I I there's some chance I rob it and all good. And there's a chance I get to jail. And if I get to jail, I can get this penalty, okay?"

Um this is the idea of how law works according to Wendell Holmes. And so what would a legal case look like in a prediction market? So let's see. It looks a bit more a bit like what we discussed before, like in the content moderation case. There's Alice and Bob who are both have a dispute between them and both think they will win and they will spend years and a lot of money into try to figure out what the judge thinks. But what if there is a market about their case? What if there is a market where people, traders who can be private investigators or people who have um knowledge about this type of cases can trade on who they think will win the the case.

And um let's say that based on this prediction market, Alice has a 90% odds of of winning. Well, maybe we don't need to do the actual trial, right? Maybe uh she can be just declared the the winner, right? If market thinks that she uh is uh so much likely to to win. And then other cases could be where there is um way more difficult decision where odds are more similar and it's not clear who would win. And these cases maybe could actually go to a to a trial and and the jury will decide uh who is right.

And if uh Alice wins, then well, there is a payout and well, people who thought she would win will get the money, okay? And there could be different applications of random trial quality control checks where you could send a small percentage of cases uh randomly selected for a actual trial review. And then if a trial agrees with the market, the decision is confirmed and the system continues. If the trial disagrees, then the market is overturned the decision and you could start um adjusting the system based on uh how the trials are uh agreeing or not with the with the market.

And if you get uh too much of a disagreement, you increase the number of trials that you you do uh for for the quality control. So this is a mechanism that is actually quite fast and it's cheap. It's could be a high accuracy system. It can be a self-correcting system based on this idea of um the trials um for correcting the parameters. And then and this is most important, this can solve like millions and millions of of cases and at very very high speed, okay? I I see your faces and you Of course, this sounds like a a bit odd, maybe, which I agree, but um you know how the US criminal justice system has this tool of the plea

bargaining? Um in this tool, um what happens is that the prosecutor goes to the defendant and makes an offer of, "Okay, if you plead guilty, you will go to jail like 5 years. Otherwise, we will go to trial and we will press for 20 years." Um so, now the defendant needs to think, "Okay, um should I accept a certain probability, 100% probability of 5 years, or should I like roll the dice, go to trial, and see what happens? I might get let out or get 20 years."

It's like, you know, one of these TV shows where you have like the the box of the incognito and you need to decide. Um And by the way, what percentage of cases do you think go to trial in the US courts? For how much? Five? 5%? More like 2%. Yeah. 2% of cases go to trial. Vast majority are decided through a process of of plea bargaining. And uh this is why some people call the system more like a system of pleas rather than a system of trials. And there's a lot of research about how this how this works and the how this mechanism is problematic because

what happens usually is that the defendant uh is advised by many times a public lawyer, a public attorney, who has no time, he's overworked, and the defendant doesn't have any knowledge about this. So, it happens very, very often that this uh backfires a lot. And there's a famous article by Jed Rakoff, he's a a judge, where he actually studies how many people end up pleading guilty for crimes they didn't commit just to avoid the risk of going to a trial and actually being found guilty and end up with a very very high high penalty. This is a very serious thing in the in the US system.

Prediction market could give a better estimate of the probability of a conviction with that information the defendant could make a more informed decision on whether to accept the the deal or go to the trial, okay? And time is prediction markets are perceived as fair and accurate defendants will probably rarely decide to go to court. Maybe they will just prefer just to accept the the decision by the prediction market as they see that it's a very high likelihood to correctly predict what would happen in court and it would do it in like just a few days instead of like having to go to years.

So, to conclude prediction markets are very new tool with many new applications. You already know many of them in politics. I'm sure in predicting wars and things like that, but they also have the potential of helping access to to justice. So, um there's obviously lots of friction to overcome. There's this is a very new mechanism. Law is a very conservative topic, but I think it would be worth experimenting with this type of of mechanisms. Um at Cloudera we're just launching a new product called Cloudera foresight that we could at some point you say to me because we might do some tests of of of this. Of course, not in criminal cases. The criminal was just an example.

Cloudera is for more like commercial type of disputes, um but this is something we are exploring quite um closely now. And let me just close with another quote by Oliver Wendell Holmes that is about the future of law. For the rational study of the law, the black letter man may be the man of the present, but the man of the future is the man of statistics and the master of economics. So, thank you very much. You have the proper questions. Yeah. Are there any questions?

Yeah. Thank you, Federico. In the plea bargaining system, how does the plea offer uh decide how much uh like um years or what what really to offer because that's the key key if if he has a prediction market, he also maybe uh like offer a higher plea because he's sure that he's going to get like killed or not whatever, right? So, in the in the mechanism, so there is a defendant that is arrested and there's pending trial and then the the law under which he's going to be judged is the says that this is 15 to 20 years. So, the the prosecutor can tell him uh based on experience

experience and based on well, now people are even starting to use AI to to predict. Uh okay. We offer you 5 years or take your chances in trial and uh prosecutors typically give um high um like um so, they they they they tend to scare the defendant because they the system is built in such a way so try to be efficient and not go to trial so to save costs and all that. So, the the what happens is that the prosecutor gives the the defendant a threatens with a very high penalty in case it goes to court with the hope that the defendant will just okay, even if I'm not guilty, I prefer to just plead guilty and go two

or five years rather than risk 20 years. In a situation where there is also an asymme- an asymmetry of power between the prosecutor who is an expert uh who knows the law very well, the defendant who in many times is uh poor person that mean all is defended by a public defendant because doesn't have access to well resources. So there's a big asymmetry. And many times the the the defendant will just try this is what the this is of the article about. People plead guilty to avoid the the risk of, you know, So that's how it works.

We have another questions. How do you protect against market manipulation? Cuz in the justice system that's a little bit different, but if someone's having a commercial dispute, they could just fund, you know, their position with the money that they're saving from not having to go through the regular lawsuit process. Um but what type of market manipulation you mean? Like if I'm if me or you in a dispute about commercials about this talk and I'm like I want to win, so I'll get people to fund me being the winner and that would manipulate the market into thinking that my decision would be the right one.

Yeah, I mean that is a type of of manipulation that happens in every prediction market. So the the standard manipulation anti-manipulation things and there and there's also Claros in potentially even if Claros is is not used for for this type of things, you know, clear you could have like a verification that there was no like misbehavior, there was no like collusion. There's even a new discipline called computational antitrust that studies, you know, analysis on chain analysis to understand if there was manipulation or not. So there's lots of different, you know, tools that could be could be used. I think the I think the important thing here is that this gives an estimate that is public.

It's not something that was done like by the prosecutor and negotiated like just in private rooms. So this can give everyone the idea that this is a public uh um like a market where people can participate and then you could also decide not everyone can participate. You could you could say it's only the lawyer accredited lawyers from whatever. So So you could also structure it in different ways so to avoid different types of manipulation. Yes. Okay, a quick one. We have just 1 minute.

Got it, thank you. You'd mentioned people using AI agents to kind of make judgments. Like I guess what do you think about the involvement of AI in these kinds of rulings and this kind of prediction market? Well, what's my talk from last 80s? Okay, this is used this is already old. I mean using AI to predict decisions in the legal system is old. It's there's companies like Premonition, Lex Machina that has been doing this for like more than 10 years.

Now I don't know how much time but we are actually at using uh AI jurors to understand how they would vote in different cases and this is this is coming and you you can see based on our experiments that different jurors, different AIs can reach different decisions in the same case. So, you even if you have AI, you don't want to have like one AI, you want to have a panel of AIs doing the decisions. So, it's it's long. Thanks a lot and if you have any questions, you can ask it uh outside. Thank you. Thank you.