Game Theory in Court, AI Arbitration Bias, and the T-Shaped Lawyer with Federico Ast

Welcome to Studying Law Around the World. I'm Claudio Claus. In each episode, I talk with lawyers, law students, and professors from different parts of the world to talk about legal education, careers, and what the profession looks like in real life. We talk about the hard parts, the surprises, and the decisions that shaped their paths. Whether you're planning to study abroad, thinking about working in another legal system, or just curious about how law works around the world, this podcast is for you. Today I'll be having the opportunity to interview Federico Ast. He has an incredible story and, and work, uh, working at the intersection of law, economics, and technology. building a really interesting institutional infrastructure for the digital age. Thank you so much for making the time to be in the podcast today, Federico. Hi, Claudio. Thank you for having me. Absolutely. And to start off our podcast, I'd love for you to introduce yourself to us and tell us a little bit about your background in economics, philosophy, and institutional design, and how that led you to create Kleros. I'm from Buenos Aires, born in Buenos Aires. My background in economics. I went to the Universidad de Buenos Aires. I studied economics first and a bit in parallel philosophy. I was interested in institutional economics where I studied my economics background and then political philosophy in the philosophical aspect. And then started my career working in the online media. I used to work in the business section of an important newspaper from Argentina called Clarín, and I was very much in contact with innovation startups. So I always start— worked on the internet, you know, business in some way. At some point I got tired of that. And then I went to do a PhD, back to school. It wasn't business, but what I studied was collective intelligence and how people make decisions together on the internet, like deliberative democracy, stuff like that. And back in that moment, I learned about Bitcoin because of Argentinians you know, were very much early adopters of the technology because of all of the problems that the country had with, well, you know, the currency, high inflation, and also bad governance, corruption. So since I was working with these innovation people, some startups, so this Bitcoin was already a bit important. People were starting to speak about that as this magical currency of the internet, but also In my view, what interests me more about that was not that much the financial aspects, but more the, this idea of using this technology called the blockchain to build institutions better adapted for the digital age, institutions that could be good for resolving governance, community coordination problems. Argentina had back then like a bunch of freelancers community trying to sell software design or like whatever service to customers from other countries. You know, with the internet, you could like just learn how to code or how to design with YouTube or with tutorials, whatever. And so it was really an enabler of opportunities for people who were in emerging economies. But then you had another problem that if people will pay your services through the banking system in Argentina, I mean, from another country, then the government had capital controls and they would like tax you like 50% of the money that you make. But you know, with Bitcoin now you could circumvent those problems and now you could learn how to code and, or to do web design or whatever, and then be paid. And then there was just one more thing that you needed was, okay, what if this guy from Canada wants to hire me to do like a video for his— but I mean, if there is a dispute with the guy and he's not going to come to Argentina to sue me in court, or he's not going to go to arbitration against me for like a $1,000 dispute. So you need like some governance piece of infrastructure for dispute resolution or arbitration in this like global settings. And the idea connected with my background in philosophy economics, so like this idea of collective intelligence. So what if we pick a jury of people from different parts of the world who are experts in like making videos and they can make a decision on the dispute that at hand and all of this system built on top of blockchain, because this is good for transparency, traceability, auditability. And this is how the main idea for Kleros was born. What an incredible story. I'm happy to hear about that. And in the last few weeks, as we were preparing for this recording, I was also, you know, trying to learn a little bit more about your company. I found it incredible how many educational pieces you have there on Kleros YouTube channel. And for our listeners who may be new to the idea, what is decentralized justice and what does Kleros do in that sense of using the blockchain, using the crowdsourcing, and even using game theory to resolve these disputes? Kleros is an app built on blockchain where you and I make this agreement where you hire me. You are in Canada, I'm in Argentina. I mean, you hire me to do this work for you, this video. And then, okay, you, we agree on what are the conditions, what is, what I'm supposed to deliver. And then after that, you send some payment in cryptocurrency into an escrow. And then we both agree that if there is a dispute, this is going to go to Kairós as the arbitrator for this dispute. If there is no dispute, I mean, you release the payment. I mean, you get your video, I get my money, it's happy. But if there is a dispute, the money is going to stay locked into the escrow and then the case is going to go to Kairós. Spheros is going to pick randomly a panel of arbitrators or jurors, and they will, each of them independently evaluate the case at hand. And they, then they will vote on who they think is right. And one important thing here is why is it built on the blockchain? So 3 main things. So the blockchain makes sure that nobody can tamper with the evidence. I mean, nobody, nobody can change a contract. Or change the, you know, the product that was delivered. Nobody can tamper with the jury selection process. Everything is going to be done as the jury is going to be picked as it was agreed by parties because everything is auditable on the blockchain. And nobody can tamper with the vote counting process. Okay. So these 3 things are stamped on the blockchain. So everyone can make sure that this is going to be done as was expected. And this is a bit like what in the off-chain offline world, traditional world, people call like the rule of law. I mean, in the sense of the system works as it's expected to work and nobody has any arbitrary power on it. So this enables both parties who are in this dispute to trust how the process is going to work. So the jurors evaluate the evidence, they vote, and let's say that you win. Claros makes a decision and then sends you the money. The escrow releases the money in your direction because you won. And that ends the case. And interesting thing is the way in which jurors are selected and the crowdsourcing mechanism is quite inspired in how the ancient courts used to work in Athens, in the ancient Athens, in the sense that in order to be drawn as a juror, you need to have a token, a cryptographic token that is like a lottery ticket that enables you to be selected into a court in Kleros to solve cases. Kleros has different courts. Some of them are like for insurance, e-commerce, others are for video or whatever. So you, as a potential juror, you take the token, you stake it into a court where you want to work, and then this random number generator is going to pick one of the 3 of all of those who stake their token into that court. And then those guys are going to be the ones who will solve the dispute. Okay. And the game theory in it is that you have— there's a system that we developed that is inspired by a game theorist called Thomas Schelling that It's about people who are jurors in the cases are rewarded based on them being coherent with the majority of the other jurors. And if they take their time and make a good decision and evaluate the evidence and all, they will tend to be in coherence with the rest and they will make money. If they just vote randomly or they just don't care and vote whatever, they are likely to be on the wrong side of the majority and then they will lose money. And this is a system that rewards people for honest behavior that we invented. And this is one of the key components of decentralized justice. So decentralized justice in a nutshell is like this idea of crowdsourced jurors that can come from anywhere. And this idea of economic incentives through this token that enables them to become honest in this system to produce very, very fast, very cheap and fair rulings for the cases they have to solve. So this is a bit of I really appreciate you bringing that up and explaining it a little more in detail. It's such an incredible concept and it just fits so well with the kind of world that we're living nowadays and with that expectation as well that results will be fair, but also that they will be timely. I see here that, you know, traditional courts and arbitration systems were designed for a world of national borders and human decision makers. As we evolve as a society, it seems like both of those get a little more and more blurred. So, you know, where do these systems struggle most when dealing with digital markets, when dealing with global online disputes? Well, these systems are typically better for dealing with digital disputes, like kind of situations where you have a freelancer that needs to deliver a digital whatever asset to a customer. Imagine like esports, imagine like gaming tournaments where you have like 20 people from 15 countries and they have a dispute. with other people. Think of like a kind of metaverse if you want disputes. Think of people you hire in a Telegram group, on a Discord server, people that maybe deliver a service to you. And then, but you don't even, you don't even know who they are because you don't know their real name. You only know their avatar. And you know, I sometimes tell people that I, every day I interact, I mean, during my workday with lots of people that I don't even know they are. I mean, I've been interacting with them for years, like 5, 6, 7 years, but I only know their avatar. I only know their internet name, but I never met them in person. I have no idea what is their actual name and where they are. So this is a bit of the reality. When people invented arbitration, the modern arbitration of the 1950s New York Convention and all that, nobody thought of this. I mean, This was for multinational corporations doing businesses with other multinationals or with governments. So this is, this was what people were thinking in that time. So the economy changed completely. So we need the governance structure to change with it. So, and it's so for everything digital, cross-border, and relatively low value, it could be like a few dozen, you know, thousand dollars, could be even more because $100,000, you're not going to go to arbitration for a contract. over $200,000 between, you know, 2 countries, it's too expensive. So those types of cases are typically better for Claros. This is not very well suited for, obviously criminal stuff, no. And high-value physical stuff, it's hard to implement. People already have solutions about that. And so like, you know, oil and gas type of typical arbitration. So high-value construction claims. So this is not really well suited for Claros. This is more or less, you know, a picture of where it works and where it doesn't. Kleros has now handled thousands of cases across different industries and countries, from what I understand. And what have those real disputes taught you about how people respond to the decentralized decision-making process and also a more incentive-based system? There are 2 main, 2 big value propositions. For example, when we work with municipal governments in Argentina and they use Kleros to resolve neighborhood disputes, or they, or with fintech companies, that they use Kleros to resolve disputes with their customers. Of the users are like very happy with Kleros being very fast and inexpensive method to reduce the court backlog in small claims courts. So Kleros is very fast, so it can give the judge in those situations like analysis of the case, discovery of evidence, which is done typically in just a few days. So one of the people we work with mentioned, for example, that they had a case that would take them 6 months to solve, would be solved in just a week by using Kleros. So this is super fast. And it is a great service for the citizen, you know, because if they want to go to consumer claims, small claims court, and it takes like 1 year, so that doesn't work very well. So one big service is make it very fast. And the other, It's just the idea of a decision that is made by, by your peers in the sense that you have a neighborhood dispute and the case is solved by other people, other neighbors. And this, in the eyes of the people who have the dispute, it's more legitimate. It's being resolved not by a lawyer who is like a, like on a kind of ivory tower, it's by other people who have usually the same problems. And one of the key concepts of arbitration in the— before the modern 1950s evolution of the New York Convention, the Middle Ages, you know, understanding of arbitration was that it's a system that is meant for other merchants to resolve merchant disputes. So it was a type of peer-to-peer resolution system where you want your peers to be those who solve the case because you want the people who resolve it understand what's going on. If you have like a fish business, you want other fishermen or fish business people to resolve them because they will know better than a random judge to make this decision. So these are the 2 things. So on the one hand, price and speed. On the other hand, legitimacy of having your peers resolve the claim. I also understand that, you know, Claros has been part of very, very interesting research and in a very recent research, it was found that different AI models can reach very different outcomes in the same disputes. So I'm also curious to ask you a little bit more about that, about that study in particular, on the sense of how should parties think about choosing an AI model when that choice may shape the results? Terof is already 10 years old almost, and back in the very, very early days already, there was an TV interview where I was saying, look, in the future we will have maybe a panel, I mean, a dispute between 2 bots, and this is going to be resolved by a panel of bots. But then it was like science fiction. But now this is something that is starting to become a reality. Now are trying to see how they can implement AI into this type of solution. And typically people think about it in the sense of, okay, you put an AI model to solve the case. And then people ask, okay, but what if there are biases? How do you know that this is not going to be in favor of one party over the other? How do you know it's not going to be biased toward the customer or toward the company? And so what we, what we did is to test a number of LLMs. First, we did a comparison. We took some cases, 99 cases from a company called Lemon, that it's a fintech company where claims resolution is put between the company and the customer and chargebacks, disputes, those sorts of things. And we put them through Claude Opus 4.7 and ChatGPT 5.5, same prompt, same cases, same evidence, same everything. And we found that Claude was 5 times as likely to rule in favor of the consumer, right? Against the company. And so this starts to bring a number of questions like, so if one model is going to be in favor of the company more than the other, Which one do you think the companies want to want for their dispute with their customer? So obviously, you know, the company will have an incentive to try to choose the model that she thinks is going to be in her favor. And customer advocacy groups are going to try to choose in favor of the customers, right? So this is one thing. So first finding depends on what model you use because of how it's trained, because of how it was developed. judges or arbitrators can to the next version of Claude 4.8. And what we saw is that 4.8 tended to rule more in favor of the company than 4.7, because it started to treat the burden of proof in a more strict way that was typically against the customer. If the customer could come up with like a more or less verosimile or narrative about what happened with the company. I made this payment and it didn't go through and then they charged me, but the payment hit the right— like some very basic narrative. The model would rule in favor of the customer. But when they started, we did a test with 4.8, it changed. It was the customer had to produce not just a narrative, but also some kind of documentary proof, like screenshots or whatever other thing that the model would see as proof that this actually happened. This was really interesting because this code was just updated. They improved the model, but this has an impact on how it evaluates evidence and this has an impact on how it decides disputes. So what is going to happen in the future, I think, based on these findings, is that first you will not have one model deciding the cases. You will have a panel, just like you have a jury or a tribunal. With more than one person because you want diversity, because different people appreciate facts differently and you want this type of diversity. So you will have panels of different models that are going to evaluate the cases. And then there's going to be like a need for more documentation on traceability on what is exactly inside the models. In the sense that let's say you and I make an agreement and we put as a clause for arbitration that we are going to have our dispute solved by Claude. Okay. But when the dispute happens, what Claude will you use? Because I mean, I will want to have to use 4.7 and you will want 4.8. But by that moment when we have a dispute, maybe it's like in 2 years, who knows what model of Claude will be out there? And maybe the models that we could agree on now are deprecated. Will we actually use AI for solving disputes? We will definitely use AI for solving disputes because it's so much faster than methods we— at Claros, what we are starting to develop now is like a mechanism where people submit their claim in the morning and they get their decision in the afternoon. So imagine justice delivered in under one day, and this can be done through AI. And so it's really a huge opportunity for access to justice, but also huge challenges because of all of these biases and situations that that are arising with the use of the models. That's so incredible. And it makes me so excited for what the future of dispute resolution might look like, right? I don't think we foresee a future necessarily without courts or without, you know, the due process that we have nowadays, very much jurisdictional bound and all of that. However, these new kinds of disputes that are really, you know, so intrinsic to a new world in the same way, they do require that new style to deal with them. I'm curious here now, Federico, as AI becomes more involved in arbitration, what safeguards are needed to protect transparency, party consent, even enforceability of this decision? So that enforceability under frameworks such as the New York Convention. Oh, this is a very big question, and I think nobody has a lot of clarity about this. So in general, between parties that are, one could argue, that are informed and they can be parties to an informed consent situation, like 2 companies agreeing on an arbitration clause and then agreeing to an arbitration done by AI. One could think of, okay, this is going to be just— consent of informed parties is enough for this to be like binding and all good. The question arises when you start using this for consumer claims, when you have a company and a consumer, and then they are in an asymmetric situation of knowledge, of power, and so In these cases, there is like this temptation or need or from both parties, you know, the company because it wants to save money and the consumer because it wants a fast decision to just use AI to make it, to do it fast. And then the question is going to be, yeah, this should be binding or not. What's probably going to happen always is that you will always have a human in the loop at some point. You may have a first opinion done by an AI that's going to take the case and evaluate what it thinks. Like it's a kind of fairly neutral assessment of the case that is not going to be binding. And then the customer can still appeal to go to a panel of human jurors to take, to look at the case. Kleros can do both parts. You can start by the early panel of AI jurors evaluating the evidence and then If the user or the customer still wants to go to a human panel, it can be appealed to humans. And then the human decision is going to overturn the previous panel done by the AIs. In the short term, at least, accept AI decisions are valid, binding, without any human oversight. In the medium term and long term, honestly, I don't know. We get more and more used to this type of decision-making by AI in different aspects of our lives. I'm not sure if people will not just prefer an AI that's more reliable than humans who, you know, are way more affected by their mood, by, you know, he had this problem with his family this morning and now in the afternoon he's going to make a decision about the case. There's like lots of studies of human judges changing their decisions based on like trivial matters, like if their team sports team won the previous weekend, or if they decide before or after lunch. So there is a lot of noise in human decision-making that maybe AI models will avoid. It's still a bit dystopian to have decisions made by AIs. It brings back memories of Minority Report kind of situations. But I don't know, when we get more and more used to this type of tools and we see them producing decisions reliable enough, maybe people will prefer them. So it's really one of the biggest questions in the future of law to try to understand what's going to happen. What will people see as legitimate in decision-making in the future? So we really don't know at the moment. Fantastic. And, you know, the main audience of our podcast here are young lawyers, early career professionals who are, you know, entering the profession. and a profession that is being shaped by AI at this moment. So, you know, new forms of private governance, new forms of blockchain and ways to interact with it. What advice, what kind of advice would you give to the next generation of lawyers who are preparing to enter into this changing market? I guess, I mean, try to learn AI as much as you can because, you know, the practice of law is based on the use of language and the LLMs are excellent at language and the LLMs that we will have in a few years are going to be so much better, so fantastic, you know, so a lot of the work that lawyers do today is going to be automated in the sense, but not, not, not, not in the sense of lawyer being like disintermediated and left out of job, but lawyers building like law firms based on like agents, you know, I've been experimenting a lot with agents and You can have them like, you can now build a solopreneur kind of a legal firm that is based mostly on agentic processes. You know, an agent finds clients and other agents, you know, receive the claims and then processes the claim and then studies the claim, makes a litigation strategy. So I would say if I was a lawyer in my early days, so learn a lot of agentic processes, learn a lot about AI, learn Maybe become a specialist in some particular industry, you know, and especially in an industry that is going to be growing in the future. I mean, imagine you can become an expert into commercial law, but commercial law applied to gaming or applied to metaverse or applied to IP in the, like tokenized IP. It's not, now it's not enough knowing the law because knowing the law was the usual thing of lawyers in the past, it's like going to be done by the AI. So you need to— there is this concept of like the T-shaped lawyer, you know, the T-shaped lawyer, you have the T, it has like 2 like sticks, the vertical and the horizontal. The vertical is knowing the law. You need to know the law if you're a lawyer for sure, but then you need to be a specialist in some topic that could be like a particular industry or a particular situation within an industry. You can become an expert in You have to choose one or a few of those specializations and also try to build your branding, personal, professional branding by producing content. You know, it's everything is a bit of a, like a show business in the age of social media. So getting clients as a lawyer is about people knowing you. So you need to produce this trust for the community to know who you are and then decide to choose you instead of choosing any other of the lawyers out there. So that's also a very important thing. One last thing, if you are into like, into legal innovation, I mean, try to also work maybe at a startup on legal tech, some startup that does something that aligns with your professional goals. I mean— If you want to work at Kleros, we have interns. I mean, if you are into this, feel free to reach out to us. I mean, we have a Positions announced on our LinkedIn or our social media. So try to intern at some company doing something that you would like to do in the future. Thank you so much for making the time to come on the show. I really appreciate you making the time to share these thoughts and advice. And thank you all for tuning into today's episode and we'll see you on the next one.