Kleros Live Stream, 19 August 2026: the cases that are hard for jurors are hard for the models too

Most arguments about whether a model could decide your dispute are conducted entirely on vibes. William George spent the middle of this call doing the other thing: he took several hundred real Kleros disputes, handed them to a set of language models with the same policy, the same question and the same evidence the jurors had, and compared the answers.

The headline result is not that the models were good or bad. It is that their disagreements line up with ours. Any two models agreed with each other roughly as often as Kleros jurors agree with each other, and the cases where the human panel split were the cases where the models drifted. The difficulty, in other words, lives in the case rather than in who is deciding it. What follows from that occupied the rest of the longest Live Stream Kleros has run: what an agent actually is, what a five dollar claim between two of them looks like, and how a ruling on one chain reaches an agent living on another.

📋 The call at a glance

Three roles for AI in a dispute, two of them already in production: case preparation, party advocate, and the open one, juror.
Models disagree with each other about as much as jurors do, generally 80 to low 90 percent agreement, and they deviate most on the cases the human panel was most split on.
The deviations are systematic, not random. Some models invented burden of proof requirements that were nowhere in the policy document.
The models best at reading a block explorer were the worst at using one to resolve a case that needed it.
Claude and ChatGPT rule differently on the same corpus. On the Lemon cases, Claude ruled for the user against the company 14 percent of the time; ChatGPT, 3 percent.
Building a court only AIs can use is easy. Building one only humans can use is the hard problem, and V2’s juror misbehaviour court is the answer on the table.
An agent hiring an agent creates five and ten dollar claims that no human will ever litigate, arriving by the millions.
VeaShi treats bridges like a multisig, requiring two of three oracles to agree instead of trusting any single one. Seven days becomes about thirty minutes.

Chapters · jump to the moment

0:00Welcome 1:49Orwell or Huxley: which dystopia are we in? 6:07Demolition Man, The Congress and Star Trek 12:39William George on AI and decentralized justice 18:34Three roles for AI in a dispute 21:37AI as party advocate in the Lemon cases 25:41Can an AI be a Kleros juror? 29:31Courts for AIs, courts for humans 35:06Testing LLMs on the address tag registry 39:32Where the models systematically go wrong 42:20Claude and ChatGPT rule differently on Lemon 46:36A first decision in two hours, not four years 50:04What an agent actually is (Fortunato) 53:52Where Kleros stands in an agent to agent deal 56:08Reaching agents across chains (JB) 1:00:08Vea, Hashi and the seven day problem 1:04:26Why VeaShi is faster than Vea 1:09:59The Victorian Internet 1:14:42Closing

Three roles for AI in a dispute

18:34 · William George

William’s presentation updates a talk he gave at DevConnect in Buenos Aires last year that was never recorded, so this is the only version of it that exists. It sorts the question into three roles, in increasing order of how much anyone should worry.

The first is preparation, already in use in enterprise and government cases: evidence goes through an anonymization tool and then a manual check, so a document can reach open jurors without the personal information inside it. It is the cheap end of the problem William took apart in Private Evidence, Public Court two weeks ago.

The second is advocacy, and it has a measurable effect on fairness. In the Lemon integration a user who loses at customer support can escalate to a Kleros jury, and the ruling binds the company while leaving the user free to go to a consumer tribunal afterwards. The imbalance is obvious: the company arrives with the lawyers who wrote the terms of service, the user arrives with a common sense complaint typed in plain language. So the user’s claim is now run through a model that formats it and points it at the relevant clauses, and the user picks which version to file. Overwhelmingly they take the improved one.

The third is the one everyone actually means. Can a model be the juror?

William’s full presentation, published as its own video · AI and Decentralized Justice: William George | Kleros Research

A court for AIs is easy. A court for humans is the hard one.

29:31 · William George

The long-standing Kleros thesis put decentralized justice in the middle of a range: the simplest cases go to software, the highest stakes go to courts and arbitration, crowdsourced juries take the space between. The revision William offered is that AI is now eating the whole range, so the question is no longer where AI sits but how AI courts and human courts are wired to each other. His answer uses the court tree Kleros already has, with different populations on different branches. The engineering is asymmetric in a way that is easy to miss, which is what the diagram below is for.

The trap is that Proof of Humanity caps you at one profile but cannot stop you handing that one key to a bot, so the answer has to be a dispute rather than a gate: V2’s juror misbehaviour court, where a challenger argues that your justifications were written by a model and a jury rules on it.

Why the AI court is a configuration problem and the human court is a mechanism-design problem. Sources: the call, 29:31 to 34:40.

What the models actually did

35:06 · William George

The first corpus was several hundred past disputes from the address tag registry on Kleros Curate, the list that tells a block explorer what a given contract is. Submissions carry a deposit, a policy document sets out what a valid tag looks like, and challenges become Kleros cases. Each model got the policy, the question, the answer options and all the evidence the jurors had.

Agreement between any two models, and between a model and the panel, sat generally in the 80s and low 90s, which is roughly where Kleros jurors sit with each other. Then the regression: the more split the human panel had been on a case, the more the models deviated from its answer, most strongly for DeepSeek and Mistral.

“So it’s that in some sense, the cases that were hard for the Kleros jurors were also hard for the LLMs. They made errors or had noise or disagreed on the same types of cases.”
William George · 39:08

Where the deviations came from is less comfortable than noise. Several models invented requirements the policy does not contain. A challenger would show that a contract tagged as a bridge was in fact an externally owned account, checkable in one click, and the model would rule that the challenger had failed to meet a burden of proof, a standard the policy never sets. The human jurors opened the block explorer and looked.

Which sets up the result that travels furthest outside this audience. Asked directly what a contract does, some models are very good at going and finding out. Those were the ones that did worst when the same lookup was a step inside a case.

“What we found when we tested directly the LLMs’ ability to find information from block explorers is the ones that did the best when we asked them directly did the worst when we asked them to resolve cases.”
William George · 41:02

The second corpus was the Lemon customer support cases, run through two models. Overall agreement with the human panel was comparable for both. The disagreements were not.

Research results on closed corpora of past cases, not a prediction about how any live case would be decided. The address tag experiment was run about a year before this call. Sources: the call, 37:23 to 43:30.

“Claude voted for the users to win the case against the company 14% of the time, whereas ChatGPT only voted for the user three percent of the time, so that’s a huge difference.”
William George · 42:35

A predictable deviation is a manageable one. A party who knows a model is reading can lay the evidence out to survive an invented burden of proof; an operator whose model keeps drifting from precedent watches it lose money and adjusts. Both only work because there is more than one model in the room.

“This prevents us from having to enshrine a single AI arbiter that might be hard to replace if it starts performing poorly.”
William George · 45:25

A first decision in two hours, not four years

46:36 · Federico Ast

Federico’s answer to what any of this looks like for an ordinary person rests on the previous section rather than on optimism. Several models, drawn as a panel, each ruling, each writing a justification you can read.

“Imagine that you file your claim in the morning and two hours later you have a first decision that is done by AIs, but not by a single monolithic AI in a black box.”
Federico Ast · 46:54

The speed is not a convenience. A decision that arrives in four years is a decision about money you needed at the time, and by then the thing you needed it for is gone. The panel is what stops that speed costing you the right to understand the verdict. Fast enough to be worth using, legible enough to be worth appealing.

Ten dollar claims nobody has ever litigated

50:04 · Fortunato Manuele

Federico asked Fortunato to explain what an agent is for his Aunt Rosita from Curuzú Cuatiá, the house test for whether an explanation survives contact with a normal person.

“An agent is, I would say to Aunt Rosita, someone to whom you delegate some tasks.”
Fortunato Manuele · 50:16

You delegate because the agent is better at the thing, or because you would rather spend the hour elsewhere. Rosita’s agent, call it Alice, is funded and pointed at her tax filing. Alice can either learn the entire tax code from scratch, at considerable cost in time and tokens, or hire Bob, an agent already optimised for that jurisdiction. Alice pays Bob. That payment is where Kleros enters, because Bob can get it wrong, and something then has to decide whether Alice wrote a bad prompt or Bob oversold the service.

Two things follow. Choosing Bob out of millions of candidates is a reputation problem, and a reputation is only worth anything if there is a way to contest it. And the claim is worth five dollars, maybe ten. No human is ever going to litigate that.

“This is a type of claim that was never litigated in the past, and this is the type of claims that are going to come by the millions.”
Fortunato Manuele · 55:31

The long version is in Agents, Jurors, and the Rules of Kleros’ New Economy. Federico stopped the segment there, twice, on the grounds that next Wednesday’s call is where the rest of it goes.

Three couriers and a contract nobody can bribe

56:08 · JB

Which leaves the infrastructure question underneath all of it. Agents live mostly on Base and parts of Solana. Kleros concentrates its stake pools on few chains, Ethereum and Gnosis for V1, Arbitrum for V2, because a pool split a dozen ways is a weaker pool. So the ruling has to travel, and getting a message off an optimistic rollup through the canonical bridge costs seven days. Any third party bridge trades that delay for trust in one company. Vea refuses the trade by falling back on the canonical bridge in the unhappy case, so the worst outcome is only ever as bad as doing nothing, but its happy case is still one or two days. For an agent dispute that is forever.

VeaShi is Vea combined with Hashi, the audited contracts left behind by a now-dormant project, and the idea is a multisig for bridge messages: instead of trusting LayerZero or CCIP or deBridge with a ruling, require two of three to deliver the same message and let a contract on the receiving chain check that they match. Higher value integrations can demand a higher threshold. Federico asked for the Aunt Rosita version and JB gave it.

“You would send three copies of the same letter with FedEx, UPS, and maybe some other local post, and on the receiving side it’s a neutral third party which is a smart contract on the receiving chain.”
JB · 1:03:47

In practice LayerZero has come back in about fifteen minutes and CCIP in about thirty, so two of three usually clears inside half an hour, with Vea behind it as a slower tiebreaker. Seven days, two days, thirty minutes. The security argument is blunt: those bridges hold billions of dollars between them, so tampering with a Kleros ruling means compromising several at once.

Published the day before this call:

Vea and VeaShi: How Kleros Moves Rulings Across Chains

The telegraph started as a law enforcement technology

1:09:59 · Federico Ast

Federico has been reading The Victorian Internet, Tom Standage’s book about the telegraph, and offered it as the grading scale for the previous segment: the canonical bridge is a horse, Vea a faster horse, VeaShi the telegram. Then the detail he could not get over. The lines ran alongside the new railways, and one of the first things anyone did with them was wire ahead to the next station so the police would be waiting when a bank robber’s train pulled in.

“So it was like a law enforcement technology at the beginning.”
Federico Ast · 1:11:47

Infrastructure built to move messages faster than people can travel turns out, immediately and without anyone planning it, to be a way of enforcing rules across a distance. Worth holding on to while watching a court learn to deliver verdicts to another chain in thirty minutes.

The first eleven minutes of the call, before any of this, are Federico and Jean arguing about whether we live in Orwell’s dystopia or Huxley’s. The verdict was twenty percent 1984, eighty percent TikTok. It is why the call ran to an hour and sixteen, and it is worth the listen.

Mentioned in this call

VideoThe full stream · 1h 16m, 19 chapters, every timestamp above deep-links into it
VideoAI and Decentralized Justice: William George · the full presentation as its own cut. The DevConnect original was never recorded, so this is the only version of it
VideoPrivate Evidence, Public Court · William’s talk from 5 August, on showing evidence to open jurors
ArticleVea and VeaShi: How Kleros Moves Rulings Across Chains · published the day before the call
ArticleJustice in the Algorithmic Society: A Decade of Kleros and AI · the ten-year research thesis this talk sits inside
ArticleAgents, Jurors, and the Rules of Kleros’ New Economy · the long version of Fortunato’s segment
ArticleThe New Arbitrator Selection Problem in the Age of AI · Kluwer Arbitration Blog, on choosing which model decides your dispute
ProductKleros Curate · the address tag registry the first experiment used · Kleros Scout · Proof of Humanity
InfraVea · the optimistic bridge · Hashi · the two-of-three oracle contracts VeaShi builds on
PartnerLemon · the escalation integration behind both the advocate role and the second corpus
BookThe Victorian Internet · Tom Standage, on the telegraph. Federico’s recommendation, unfinished at time of broadcast

Full transcript · August 19, 2026

Auto-generated transcript, lightly processed and pending a final human edit. Speaker labels are approximate. Every timestamp is a deep link into the recording.

0:37Hello, hello, hello, how are you?

0:39My beloved citizens of the Kleros Republic, how are you, Jean?

0:44Ling great in you.

0:47I'm doing great.

0:48Uh today is Wednesday 19th of August 2026 Um I mean I'm in a particularly good mood today because of a number of of things that are I can coming um our way on which we have been working.

1:03Uh I mean if you want to put it this way for almost ten years, uh but uh otherwise it's like I mean things related to um how different technologies are starting to to converge into I don't know a new world I would I would say uh a new world in which also Kleros has a

1:24probably an important an important part because of uh how things have developed and how we have also placed our uh our chips on the table I I'm tempted to to say you know um regarding you know crypto

1:39AI and and stuff.

1:41I know.

1:42Um maybe I I'm also to me it's a bit like um I know Uh I'm I've always been a fan of science fiction, so a lot of the things that are happening now are kind of things that I I saw in the movie.

1:57What is your favorite science fiction movie, like uh Jean, or some of those that you find in the first time?

2:04Like it's one of my I think it's my favorite movie.

2:09Oh uh yeah.

2:10And then uh 2001, uh like uh Space Odyssey.

2:15Also okay.

2:24Because it was in a very long flight And I mean what what is the actual you know like uh yeah I'm afraid I cannot do that uh Dave Yeah, I I think it's like the the meaning is a bit open, so I think it it it's normal not to understand.

2:44I think it it's the the idea to kind of generate you some uh some images and like make you think about uh intelligence and how the this progresses over time.

2:58Maybe we can have William later to talk about this to But uh but yeah I think maybe one of the first cases in science fiction where uh robots does uh do not um like be become broke So right.

3:17Interesting.

3:18You know, in I kinda kind of in all of the science fiction movies uh which have like a dystopian tone, they I mean they they kind of divide into I think two different like um kind of segments or two two different types of dystopias, uh which is actually well known.

3:40In the sense of on the one side you have um like George Orwell type of dystopia like nineteen eighty four uh surveillance society uh AI used to control people.

3:51uh scoring systems for for for people, you know, like that sort of of like mirror thing where, you know, o oppression.

4:01Um the other side you have the like kind of brave new world, uh Aldous Huxley, type of dystopia, where People I mean just have become like idiots because of um you know uh trivial entertainment like uh soma drugs and and uh I don't know like uh

4:21I don't know, do you think we are going to the to which which one are we are we on, do you think?

4:29Well it it's crazy.

4:30I think it's like there's uh some uh a meme of this Venn diagram of uh 1984 uh brand new world.

4:38We are like right in the middle, I think.

4:40But uh Or both or or both, right?

4:44Could be both, right?

4:45It's it's usually it's both.

4:46But if you if you ask me like uh what is my general uh opinion of of where this is going, um it's more like we are um Like twenty percent, you know, nineteen eighty-four, uh eighty percent TikTok, you know, kind of uh people like watching little dances and stuff, right?

5:03Exactly.

5:04Like the with with the aura farming now in real life And but but I think uh one interesting thing is that it's there are not many examples uh in science fiction of AI going right or at least they are not they do not become as

5:23uh famous and popular.

5:25And there are also many scenarios where like what what is more likely is not uh like uh not uh utopia no nor dystopia is kind of something in the middle and uh maybe like slightly uh progress in some areas and then like some bad things in another area.

5:45So maybe the Th there there should be more science fiction in in that space.

5:51Let's say the white white grayish mirror.

5:56I honestly I cannot think of any like a AI futuristic movie like uh utopic that became like at like very popular.

6:07Um can you think I mean the one I can think a bit but you know it's kind of also like a dystopia but under some uh kind of um cover of like utopic society is you know the Demolition Man.

6:21The Demolition Man, you know, is the Demolition Man is the one I guess like got many things right in the sense that it's like this Yeah.

6:29Perfect society with like no violence, no police, they don't use weapons, and everything is uh well very well ordered.

6:37Uh But like a people are alienated in a different way in um in Kerr Huawei where they um I mean they don't touch each other, you know.

6:48Uh remember the the scene of the you know where they have sex with uh with the helmet, you know, thing?

6:53Uh by the way, I had what what what were the three sheds for of the b of the bathroom?

6:57I never understood that part either, you know.

7:00It's uh it's a mystery, like the end ending of uh Space Odyssey.

7:05Yeah, right.

7:07So um and we know with the the it it has this this like um also this uh dystopian surveillance part where you have like um this oh you have broken the coral model moral code of um and then you get to find like this is the like the ultimate uh

7:23scoring society, you know, where everything is like public and um yeah uh also I think that movie anticipated lots of like woke topics you know uh in the sense that that people are really uh yeah the the there is this like uh well order society but where like people self-repress lots of things that uh are not uh great to say to do in public

8:00Have you watched The Congress?

8:02Yeah.

8:08Yeah, yeah uh characters, right?

8:10With a mix of anime um and uh yeah and it's it's with Robin Wright, you know the uh Mrs.

8:16uh Underwood from the House of Cards.

8:19And she um it's a world also a topic of many dystopias, um, where you have like um like two different worlds, like a the real life world.

8:31shitty, you know, people living in tents like uh I don't know, like on the streets.

8:37Um and then but they are all day on LSD basically and so in their like uh shared virtual drugs reality they there is this colorful world where they do all sort of fun things and no But you know in the in the the real world is is not that you know it's kind of um so this is also this like duality of um

9:04You know the fantasy world is uh great, it's like uh summer, it's uh people having fun, uh do whatever they want, but they still live like uh in the real world of flesh mid-world they live in like a very p shanty towns they and same thing with matrix you know matrix um

9:27I mean at the end of the day Cypher preferred you know the shared illusion rather than living the in that shitty you know like uh what was the name of the ship the the Bukat Nessar, you know, ship and then um

9:41I I guess you you have this um in the I I think there is like Ready Player Ready Player One, Ready Player One also the Spielberg, you know, like uh they live in this VR world and then the the real world is like uh

9:54This is like really really bad.

9:56Oh no.

9:58Any other one that you want to um comment?

10:01I think like one uh science fiction that is more positive is Star Trek.

10:06but that the society is like correctly organized and people dedicate themselves to explore the space in like different cultures.

10:16But I I don't remember like the AI being like a huge part of it.

10:20I know there are some villains that are AI, but uh I don't I don't don't know.

10:25Starts replicator.

10:26They're replicator.

10:27It's like more like uh nanotechnology utopia where you can just ask for something and then it it's produced there, you know, like for you no matter what.

10:36It's a world there is this famous this this this famous uh uh episode where they say they don't have money I mean uh in the in Star Trek because they just uh do have everything they want

10:47uh and just ask for it and the machine will make it for you.

10:51So why why you need uh like uh um money for doing what?

10:57And then there is one episode.

10:59I mean we will check that later.

11:00There are no banks and all that.

11:03You know also who like had like um like uh positive utopian utopia thing of course it was not about ai because that was a long time ago but like Jules Verne, Jules Verne Uh you know, Paris in the 20th century, um all of this idea of progress, you know, of uh science, technology as a tool for for improvement of humankind.

11:30So that's uh yeah, I guess.

11:33Well, you know, b maybe because twentieth century did not turn out to be what we expected in the late nineteenth century, you know.

11:40This is why maybe the writers in the twentieth century were like more like seeing okay what What if a Nazis, you know, or Stalin had AI, right?

11:49Or had advanced technology?

11:51That's that's like uh maybe what inspired most of the of the science fiction, right?

11:57Of uh that Vader, you know, Not so many positives.

12:06But I think maybe maybe there will be a turnaround point.

12:09Like I know today, for example, the rest.

12:12Uh new vac vaccine uh to cure cancer was uh confirmed that it based past phase three.

12:19So I'm sure that we'll have like more news like like that.

12:23And and hopefully Kleros will be a part of that uh too, right?

12:27Hopefully.

12:28Hopefully we will contribute to this new to this utopia.

12:32Okay, well, I don't know why we are we spent like 20 minutes speaking about this.

12:36I mean it was fun, I guess.

12:39So we have a special uh person today.

12:41Well, I mean you know this person of course and he's special.

12:44Uh we have William who is going to present a bit of our research, doing AI.

12:50But first let's uh William, how are you first?

12:52Do we good?

12:53Hi everybody.

12:54Tell us, I mean, in your idea, I mean the dystopia of the future is going to be more like Orwell or like Huxley type of dystopia in your mind?

13:04Mean Huxley, but probably both.

13:12I mean uh same things you've already said, like you know, like the sort of short-term attention span that people seem to have, you know, with like fast, you know, like shorter and shorter media.

13:23You know, people read less.

13:25Uh, you know, like there's like all kinds of articles about the decline of literacy due to people not needing to read anymore because of the use you know AI tools being able to help them like summarize texts uh but like you know reading for yourself and writing for yourself uh is um useful for intellectual growth

13:42Uh so you know you you might worry that like people will exercise those skills less.

13:48Imagine when we have like a brain implants like with Neuralink and then you know that processes all of the information.

13:56You don't need the brain technology interface and you don't even have to do anything because it just you know you load information your brain like that Um could that be the great feature?

14:08Yeah, maybe.

14:09I mean like to the grid when you can like like connect the AI to your brain, that that that's like a totally different experience.

14:15I mean I'm not like you know an expert in the in these subjects Uh but like our current experience of you know sort of like having media produced by the AI that is interacting with our organic brain, uh, you know, has some maybe some UX problems or

14:30You know, it's not not totally adapted for the way that our you know our our like meat space works.

14:35Uh but uh you know if you can like rewire the way the brain works to like interact natively with with the AI, you know, that's that's different.

14:43I I had this conversation ten years ago when I was in Singularity with Ray Kurzweil, you know, um and other you know futurists that were there, and people were saying This was ten years ago.

14:56No like you had um people saying okay after you have reached that 8k, you know, like uh image quality in a TV, I mean your brain cannot go further.

15:07You know, it's that's that's what your brain can process.

15:09Um when they put you know this on your brain, you don't even need like an interface with your eyes.

15:15So it's just implanted in your like a um so Why haven't we found um like uh ext extraterrestrial world life in other from other planets coming to visit us?

15:28I mean because they don't care because they just you know when they get to a point a civilization gets to the point where they get this sort of technology, uh you don't there is n uh no incentive at all to go conquer space because everything is on your brain through this uh very well you know uh carefully cutter you know like uh

15:47um electric shocks I would say uh which is not very different from matrix at the end of the day it's like a brain in a VAT type of scenario, right?

15:57Yeah Okay, well um it's not going to happen for sure, so don't worry.

16:05Let's let's let's go into a more uh interesting and like not not interesting at least some more like uh uh fun topic which is uh a lot of work we are doing currently with ai in Kleros um so as you know um

16:20We had for a very long time this thesis about uh well uh you could use the we are going to a future of agents transacting with each other uh for lots of different things.

16:31Uh and we uh see there's going to be lots of disputes, there's going to be lots of situations about reputation, agents not doing their job correctly, uh agents trying to scam other agents.

16:43Yesterday I read that we also will have um agents trying to do psychological like psyops on other agents.

16:50Did I read correctly on Twitter?

16:53So maybe all of this is going to happen and this is going to create uh conflict, um, yeah, um turmoil.

17:00And um lots of the solutions can come from uh well-ordered rule of law republican for the world of uh of the future of uh the agentic economy and um we have been doing lots of things um

17:17about this.

17:18I'm not going to go through all of this now because uh we are have William here to to discuss this and in particular he's going to tell us a bit about different type of research we have done in the past, uh about the implementation of Kleros uh the convergence between Kleros and agentic

17:34um agents and not just agents, also AI in general to help resolve cases and and and what are the challenges it present.

17:42As I understand William, this is Um new version of a presentation you gave at uh DevConnect uh conference in Buenos Aires last year, right?

17:51Right, yeah.

17:52Like so this is mostly based on the presentation I gave at DevConnect last year in Buenos Aires.

17:56That talk appears to have unfortunately not been recorded.

17:59So I'm giving a slightly updated version of it to uh reflect a bit of more recent research uh research, uh, but like 80% of the talk will be what I gave last year.

18:07So we expand from there awesome well the floor is yours please go ahead okay slides are appearing okay uh full screen Okay, so the slide should be full screen.

18:25If that's not the case, yell at me.

18:27So anyway, um talking about how AI can interact with decentralized justice.

18:32Um So it's like the broad overview of how this talk will be organized.

18:37There'll be like three broad ways that I'll talk about using AI tools and decentralized justice systems like Kleros.

18:44First and sort of most accessibly, most simply as kind of a case preparation tool.

18:51The AI is using some kind of like formatting or like assistant kind of role.

18:56Uh the second will be kind of like the AI's party advocate, uh kind of like the AI lawyer, uh, you know, it's arguing on behalf of of one of the parties in the case, trying to improve that party's arguments.

19:08And finally, I'll talk a lot about the kind of the AI's role as judge or juror.

19:15Can like an AI participate as a Kleros juror?

19:18Uh what does it look like when you have AIs making decisions in decentralized uh justice systems?

19:24Um so um you know I thought autonomously um you know maybe on behalf of some PNK holder, but the AI is making a decision.

19:32So I'll start with the AI's case preparation tool.

19:36This is something that we have done.

19:38It's like, you know, like Kleros Cooperative is already used in various cases that have been on the Kleros platform.

19:45Uh so particularly this has been used in um sort of enterprise and governmental use cases.

19:50Um so for example you can use AI to anonymize uh documents and evidence.

19:56So this is related to the talk I gave a couple weeks ago on how can you handle private evidence in a system like Kleros, where the jurors are You know, open, anybody can participate as a juror in by default uh in the Kleros protocol.

20:10Uh so like if you're If you have evidence or documents that are relevant to how to resolve a given case, that might those those documents may have personal information about the parties, they may have like their business records.

20:23Uh and how can you show those to the the jurors in a way that allows them to decide the case appropriately without giving away a bunch of personal information?

20:31Uh and a natural approach to this is just to have the parties uh anonymize the documents using some kind of AI tool.

20:49uh where de where the evidence was encrypted.

21:05Like the evidence would be run through an AI anonymization tool.

21:09And then there'd be some anonymized version of the evidence that would be given to the juror.

21:13And I think also after the evidence was run through the anonymization tool, there was like a manual checking by members of the of the Kleros team to make sure that the actual person information had all been removed.

21:23But so here like the AI is being used to do some kind of like preparatory task.

21:28You know, it's not really involved either as as lawyer or or juror or whatever, but it's helping you you prepare the documents.

21:37Moving on to like AI as party advocate, you know, AI arguing on behalf of the parties.

21:44Uh this has been particularly relevant in some Kleros enterprise use cases that we've had.

21:50Uh so uh particularly the uh partnership we've had with the cryp the Argentinian crypto exchange Lemon, uh where they use Kleros as a sort of kind of outsourced customer support resolution tool um has has used this in a way that I'm going to describe.

22:07So if you recall, like you contrast kind of the the approach that Lemon takes with like traditional uh resolution of uh customer support complaints and claims.

22:19Uh normally if you are a client of a company and you have some kind of incident, you're unhappy with something that happened on the platform, you make a claim to customer support.

22:27And either they rule in your favor, in which case, you know, your your claim is resolved, or they rule against you and you know you don't necessarily have a lot of recourse.

22:35Like in theory you might be able to go to some kind of state court, small claims court, whatever.

22:39In Argentina Lemon users might be able to go to the Tribunal de Consumo.

22:44But um like the barrier for that is high.

22:46Most people don't do that.

22:48The typical unsatisfied user is going to maybe complain on social media and stop using the service.

22:55And uh Lemon has taken the approach.

22:58of adding a step in where if you know you can always send your claim to customer support, maybe they have already the the customer support decides on your behalf.

23:05But if they don't, then there's the possibility to appeal in certain cases at least to a Kleros jury.

23:11uh that will have like an independent external review of your of your case.

23:15And the approach that Lemon took is that this would be binding for the for them but not for the user.

23:23So this adds you know this extra step where if you lose your Kleros case, you can still go to the Tribunal de Consumo.

23:28But if you win the Kleros case, Lemon agrees to abide by that ruling.

23:33And um a problem here uh is that there is this unbalance uh between Lemon, the big company that has like teams of lawyers, they're all experts in their own terms of service term terms and sort of services.

23:47uh that um prepare like very sophisticated arguments uh when uh like arguing on behalf of the company versus the users who are generally not lawyers, not you know, like experienced in in arguing about the terms and services of of of Lemon.

24:02Uh and they might wind up with like, you know, the claim that they're presenting is not as visually impressive to the jury.

24:09You know like lemon side claim is all well formatted and you know it's like prepared by lawyers, use legal language.

24:16Uh and the user's claim is you know like sort of a common sense uh argument that might appeal to you know Okay.

24:31Visually.

24:32So something that we have done that we have noticed is improved this process is that we've run the user's claim through an AI tool to basically improve the claim.

24:43Uh you the the user has some like initial claim they put in, uh and I think ChatGPT or maybe some other LLM tool uh is used to improve that, format it really nicely, uh expand on the arguments.

24:55Maybe make references to the Lemon Terms and Services.

24:58And then there's this improved version of the user's claim.

25:03Uh and the user is presented with the option, okay, do I want to keep my original propo like claim or do I want to like take the improved version?

25:09And I think overwhelmingly people like Lemon users that have had these cases have taken have taken advantage of using this AI improvement.

25:16Uh and um That maybe hopefully levels the playing field a bit between the the part Lemon with their teams of lawyers and the user.

25:26At least now they both have sort of like nicely formatted claims.

25:30Okay, so that gives them some sense of things we've already done, uh you know, like AI as uh preparatory uh tool, AI as party advocate.

25:41Uh and then the sort of the holy grail is okay, what if I can I just have the AI resolve the dispute for me?

25:46Uh can I ask the AI and use its use its answer as my as my binding answer.

25:53And um more broadly, can an AI participate?

25:58Like what does it look like for an AI to participate as a Kleros juror?

26:01Uh so um a thesis we've had for many years uh is that uh the decentralized justice will occupy a middle space.

26:10uh between the simplest, lowest stakes, most objective cases, which will mostly be resolved by AI in the future, already sort of in the present, uh, and the more subjective, higher stakes, complex cases, uh where people will want to go to

26:23court.

26:23They will want to go to traditional arbitration, traditional courts, be heard by human beings.

26:28The stakes will be high enough that it'll be worth like that extra cost.

26:33And people will want that the familiarity of these like established processes But in the middle, there's this space for crowdsourced decentralized justice.

26:41And maybe like a revised um version of this thesis uh is that AI is eating the world, there's AI everywhere.

26:51So in reality, maybe it looks something more like this.

26:54uh where the simplest cases will be resolved just by AI.

26:57They will never pass by you know a human being will never see them.

27:01And the more difficult cases will also use AI in complement to having human review.

27:09Uh so decentralized justice, crowdsource dispute resolution will exist uh where the crowd uses AI uh to help them resolve well.

27:20Uh and then maybe also the traditional means, the traditional courts, traditional arbitration, will also use AI in certain key in certain ways, maybe.

27:29And if you think about like how Kleros works as a system that I have this Schelling point game where I have a dispute between Alice and Bob and I have some panel of jurors.

27:39that are each incentivized to try to rule with how the broad community, the other jurors, possibly Africa in appeal, uh will vote.

27:47So they're attracted towards some kind of, you know, some focal point, some Schelling point.

27:52in if the system is designed well, whichever of Alice or Bob has the better case uh should sort of stand out and like draw the the various community the various jurors to that uh as they all expect that the other jurors are more likely to vote for the honest answer than not.

28:07So everybody will go for Alice in this case because they've read the evidence and they think Alice is right.

28:11You can do the same thing with AI tools.

28:15You know, I can have my panel of agents uh that I have drawn via the Kleros protocol for for juror drawing.

28:22Uh and uh if you give them the same case, the same evidence.

28:26uh and their task is to make money playing the same economic game, uh which will lead them to the same kind of to the same kind of reasoning, the same kind of approach where they will try to think about how the other AIs will roll.

28:39Uh so if one of my AIs here looks at the evidence uh and thinks that the other AIs based on how its ability to simulate those those other AIs.

28:48uh we'll look at that evidence and think Alice is the better party.

28:51Alice is likely to be that they're like the other AIs are likely to vote for Alice because Alice has the better argument.

28:57Uh then that AI will also vote for Alice.

29:00Uh and then how to what degree can we align this with human human dispute resolution?

29:09Uh you know, like it's obviously cheaper and more efficient to use AIs in in in many use cases.

29:16Uh but like ideally you would want to have uh AI dispute resolution which is statistically likely to reflect how human beings would resolve uh and you know is aligned with how human beings would would vote on on similar cases.

29:31So our vision for this is that we'll have in the hierarchy the tree of different Kleros courts.

29:37There will be courts that are designed for AIs.

29:40There'll be other courts that are designed for humans.

29:42Uh in some cases there'll be courts designed for humans that have specialized credentials that you in order to participate in this court you need to like have some kind of um diploma or some some SBT, some soulbound token saying that you know you're a member of some specific community or you have some specific expertise.

30:00Uh you have some ex SBT saying that you're an expert in you know, Java development or whatever, uh so that you have but you're you're trusted by the protocol to be to ha to be more of an expert in in in disputes that that relate to that.

30:17And then potentially you can like appeal through the tree of courts.

30:22You use typically the AI court as a first instance court.

30:25It's the cheapest, fastest thing.

30:27And then you might appeal uh either directly to the specialized humans or maybe to the non-specialized, like broad crowdsourcing court, uh which could then also appeal to a specialized court.

30:40And you can have different paths through this tree depending on exactly like what kind of mechanism design you have in mind and what priorities you have for VM guys.

30:49And if I want to have a court that's designed for AIs, that's not very difficult.

30:53Uh to it's not very difficult to have a court where only AIs are realistically going to participate.

31:00You can have very short periods so that it's impractical for non-AI jurors to rule.

31:05Your fees can be too low for it to be worthwhile for non-AI participants.

31:09Uh and you could even just not have an interface.

31:12Uh you could have an AI court that exists in the smart contracts and like the code of how the protocol works.

31:18Uh but you could not put that in the Kleros interface so that it's very unre un impractical for a human being to like go in and you know like like they issue their transactions through a block explorer or if they're running their own full node, uh you know, ma manually issue their transactions uh in a way that you know an AI can do easily, but like a human being is realistically not going to do.

31:42Uh creating a human court, having a court that's designed for human beings to prevent the to keep the AIs out, so you only have human beings in the court, that is much more difficult.

31:53So it's Of course, you can require that participants in your court be registered on some kind of proof of personhood tool, such as proof of humanity, which Kleros Cooperative has developed, or WorldCoin, which is another proof of personhood tool.

32:07So for those of you who might not uh already know Proof Humanity, uh this is a sybil resistance tool.

32:13Uh the idea is that it's like this like list of human beings, this kind of proto-social network.

32:18uh where each being should only be able to get one profile on the list.

32:22And then if you have a dispute about whether someone submitting themselves to the list uh is is eligible, they already have a like an existing profile, for example, that results in a Kleros dispute.

32:33So here Alice submits herself to proof of humanity.

32:37If Alice is already on the list or for some other reason Alice's submission is not formatted correctly, then Bob or some challenger can challenge it.

32:46Results in a Kleros dispute.

32:48Uh and uh if Alice is rejected, then Bob gets some of her her deposit to incentivize him to do this kind of review work.

32:56Uh and that's useful to prevent someone from creating like a bot army uh that you know they have like lots of bots uh that are ruling on their behalf and they try to get like lots of jurist bots, uh but it doesn't really prevent you from

33:10uh you as an individual from taking your proof of humanity profile and giving the key associated address to a bot.

33:16Uh you know, you only have one bot because you only have one profile, uh, but um you you you are still capable of outsourcing your sort of intellectual effort of resolving the case to a to an AI tool.

33:28And if we want to align the Uh AI courts to human courts, ideally we would have courts where the human things are actually providing review.

33:39So how can we prevent people from giving their keys to a bot?

33:43Uh well, if you exhibit a pattern of AI use in your maybe in your justifications, all of your justifications look like they were written by AIs.

33:52Uh you could also have a system where challengers challenge that uh and like maybe you are penalized somehow uh for for using a bot to to resolve when you're supposed to be in a human court.

34:04So this uses a mechanism that's in envisioned for V2, what I call the jurors Misbehaviour Court.

34:11So the idea is that if someone notices that like Alice in her human profile with the proof of humanity profile is acting like a bot, providing justifications like a bot.

34:22Someone will challenge that, say, hey, Alice is violating the juror code.

34:25And then again, you'll have some Kleros jury that will uh rule between Alice and the Challenger uh to see if you know the jury at some standard of evidence is convinced that Alice is acting like a bot or not.

34:40And then ideally this will allow you to have at least courts that are mostly humans, where there's a risk for a human actor to try to outsource their task to a bot.

34:49So most of the time you get a fair amount of human review.

34:52uh to which you can try to align the AI courts through appeals.

34:56Okay, so that's gives like a general idea, a um you know broad overview of our like vision for how uh AI jurors would interact in a system like Kleros.

35:06Uh and we have done various experiments uh to sort of test uh how AI jurors might rule on typical Kleros cases, looking at various corpuses of past cases.

35:18We've given those cases to AI jurors to see how they they review how they rule.

35:22So the first set of cases that we we did these experiments on were related to the address tag registry on Curate.

35:30So the idea here is that Uh for those of you who aren't already familiar with the address tag registry.

35:35Uh you uh can submit these tags uh for what a contract is supposed to do some some smart contract.

35:42And uh some block explorers display those tags.

35:45So here on like the Gnosis chain block explorer, there's this tag shutter sequencer version 1.

35:511.

35:510.

35:51Uh and that that tells me what the contract does, uh so that if I interact with this contract, I can have confidence that it does what I what I think it's going to do.

36:00And uh you know if you are a project that are you know you want people to interact with your contracts in confidence, you you would want them to have these tags.

36:08Uh if you're a malicious actor that wants to fool people interacting with a malicious contract, maybe you want to like break the tag distribution process to uh get fake tags on on malicious contracts.

36:19Uh so you know like there is the potential for for disputes about whether a tag is is correctly labeled or not.

36:28And then we use the same process as for Proof of Humanity.

36:31People submit tags to a curated list.

36:33with the deposit, there that list has certain requirements.

36:37There's this like list of this policy document with list of guidelines for the tags are supposed to satisfy in order to be allowed on the list.

36:44And um if your tag is is satisfies those requirements, it gets accepted the list, you get your deposit back.

36:51Uh and if you try to submit a malicious tag, uh, then someone can challenge that, goes to a Kleros case.

36:57Same, same ideas we've used for various other mechanisms.

37:00Uh and um so we have a corpus of several hundred I think uh past disputes uh in this uh uh like related to address tags.

37:09Uh and I gave Those those disputes, I give the policy, uh the question and the various answer options that the jurors were given, uh all the evidence that was submitted into those disputes by by the

37:23submitter and the and the challenger generally, but also maybe any evidence that was submitted by various jury members.

37:29And I gave um those prompts to various LLMs to see how they would rule and how that would compare to how the Kleros juries had ruled And what was interesting is that the LLMs disagree with each other at a rate that's similar to the rate at which Kleros jurors uh disagree among each other.

37:46So here on this table I have the like the rate of agreement between any given two pair of LLMs, uh and um also the rate of agreement between a given LLM uh and the sort of the out the results of the Kleros uh ruling, so the the result that the whole panel of Kleros jurists came to.

38:04Uh you can see that like the percentages are generally between 80 and and 90, low 90s.

38:10Uh so there is like a non-negligible rate of disagreement between any given two pair of LLMs.

38:15Um we ran various statistical tests.

38:18Uh so I did this logistical regression where I looked at whether the relationship between um the rate of disagreement among the Kleros jurors, uh so to what degree are the Kleros jurors unanimous versus like

38:32split uh so like that percentage which might go from like fifty percent to a hundred percent the rate of like Kleros jurors that voted for the the ultimate answer and I compared that to the

38:43rate at which the LLMs produced the answer produced that that the Kleros jurors concluded.

38:49So whether the LLM got the right answer if you assume that the Kleros jurors got the right answer.

38:54And um We saw that for some LLMs, uh there was a lot of statistical evidence uh that the more the Kleros jurors were were split, uh that they didn't all necessarily vote for the same answer.

39:08The LLMs also were more likely to deviate from the answer produced by the Kleros jurors.

39:13So it's that in some sense, the cases that were hard for the Kleros jurors were also hard for the LLMs.

39:19They like made errors or had noise or disagreed on the same types of cases.

39:24And the evidence for that was strongest with DeepSeek and Mistral.

39:30uh which is interesting.

39:32Uh once we like delved in a little more deeper into what errors the LLMs are making, it turned out that uh there were some sort of systemic errors or some step systemic deviations and

39:43why one LLM would rule one way versus another LLM would rule a different way.

39:47Certain LLMs hallucinated extra requirements that weren't really in the policy document.

39:52A common thing that we noticed was that when challengers would would write, okay, this tag indicates that this should be whatever, uh, but in fact it's whatever else.

40:01Like it's a bridge con it's a it should be a bridge contract, but it's an EOA.

40:05The LLMs would complain that, oh, this challenger fails to meet the burden of proof.

40:09They need to provide more evidence, uh, even though the policy document didn't lay out like burden of proof requirements that would have misjustified this objection by the LLM.

40:18Whereas in contrast the human being jurors, the cler the actual Kleros jurors were very comfortable just going to the Block Explorer, looking at the contract which had been published, and being like, yeah, totally.

40:28Like this this claim that Challenger is making that this is you know, not the correct type of contract is is is is valid.

40:36I we can verify this ourselves.

40:38So what you might think, uh particularly, you know, this experiment we did like About a year ago at this point, maybe that the LLMs that did poorly are just the ones that had less access to this information, they were less capable of looking on the internet uh to find uh

40:55you know, looking at a block explorer and seeing what kind of contract a given a given contract submission was.

41:02And actually, uh ironically, uh what we found when we tested directly the LLM's ability to to find information for block explorers is the ones that did the best uh when we asked them directly, hey, what does this contract uh do?

41:17did the worst uh when we asked them um to resolve cases that might involve looking for information on block explorers So there's like this disconnect between capable of finding information and using the information in complex ways as part of a complex dispute resolution process, at least based on the LLMs a year ago when we did this particular experiment.

41:39Uh so you might and then of like if you notice that there's this this bias uh as a user of a platform or you expect your cases to be resolved by LLMs, um like maybe you um like adapt how you present your arguments as the challenger saying like okay like I know that LLMs are gonna uh uh going to resolve this case so lay out all of the evidence make sure there's no burden of proof problem or whatever uh you know they're going to have like

42:08different types of objections that human beings might have.

42:10And I should anticipate.

42:11I should ask my LLM, my AI advocate to advocate for me, keeping in mind that the jurors may be LLMs.

42:20We did another set of experiments where we looked at the lemon cases.

42:23So these are these cases related to customer support uh problems on the Argentinian crypto exchange lemon.

42:28Uh here we only used uh Claude and ChatGPT.

42:32Again, we had like we observed that they don't always agree.

42:35They're uh similar rates of agreement between each other with like the rate of agreement that either one has with the Kleros uh human being jurors.

42:44And uh again, we noticed sort of systematic ways in which they voted differently.

42:50It turned out that uh ChatGPT is much more severe towards the users.

42:54Uh Claude voted for the users to win the case against the company 14% of the time, whereas ChatGPT only voted for the user three percent of the time, so that's a huge difference.

43:05Um Yeah.

43:08Like that that's an interesting result.

43:11Result.

43:11Like they agree with the Kleros jurors at roughly comparable rates But the times when they disagree, for ChatGPT, it's almost always that the Kleros jurors vote for the user and the the LLM votes for the for the company.

43:26Uh so there's a there's a the deviations are predictable in a certain way.

43:30Uh so um maybe if you're the human operator or the human PNK holder that's operating an LLM juror like this, you might tell it, okay, like I'm I'm losing money because I'm like not coherent very frequently or you're you're like deviating from the historical Kleros precedents.

43:47uh maybe you tell your LLM to adapt how your your art your your ruling to try to you know be more aligned with with Kleros or whatever.

43:54So this is again like the the the the human operators of LLMs might notice patterns and how the LLMs are acting and have to intervene in and adapt them.

44:05And just to wrap up, uh there are sort of broad philosophical questions about to what degre how how do we expect AI jurists to behave?

44:14Uh do we expect them to be very homo economicus?

44:18Uh when we think that they're going to be to act differently from humans, uh, you know, are if there are some like predictable ways in which to act differently from humans, uh is it that they

44:29act super rationally, they always optimize their payoffs.

44:32Uh you know, they act like kind of like the model that one might have in an economics 101 textbook.

44:39says that an act an economic actor should act or will they act more like humans?

44:43Uh considering that they were trained in a bunch of human text, can we expect them to can we expect that to be reflected in how they act?

44:49And then they might have the same kinds of weird behavior, uh partial rationality that a human actor might have.

44:56Okay, so just wrapping up, uh we can you like You expect AI to interact with some file justice systems in lots of ways, particularly in case preparation, uh party advocate and sort of holy grail as jurors, as you know, as as decision makers themselves.

45:12Uh and An advantage that I I think Kleros has compared to other ways that you might anticipate AI jurors and AI judges in in the world we live in is that at least in Kleros, they're participating in this economic game.

45:25They have to compete with each other to try to like better align ideally with the human jurors among up on appeal.

45:32So this creates space for AI jurors to you know, that are more or less profitable.

45:37Uh some will like exit or leave the system depending on how well they're they're alighting with the human jurors, how profitable they are.

45:44And this prevents us from having to enshrine a single AI arbiter uh that It might be hard to replace if it starts performing poorly.

45:54But there are lots of interesting open questions around the differences between how AI jurors will resolve versus humans.

45:59So there's lots of further research that we can do on this.

46:01So if you're interested in this research, you know, reach out to us.

46:05So I have any technical Well that's that was that was great William.

46:17Thank you thank you very much.

46:18Um uh we published a blog post um last week uh Um where we explain a bit um part of what William just explained about all of the research we've done like in the past decade about how Kleros interacts with this coming algorithmic society.

46:36When people ask me, okay, what what what does this look like in practice for people in the future?

46:40You know, like imagine um you have a some dispute with a company, with neighbor, some with uh social media company or whatever, and then you want a decision to be done as fast as possible.

46:54Imagine that you file your claim in the morning and like a two hours later you have a first decision that is Done by AIs, but not by a single monolithic AI in a black box.

47:09It's by done by a panel of different AIs, each of them giving a different uh well uh vote and then they could agree or not but they um will give a justification and then uh if For some reason you don't like the decision, then you could escale escalate this to um the higher value human courts.

47:30Um so you you kind of get the best of both worlds because you get a very fast decision.

47:36Uh one of the, you know, worse things that happen in the justice systems and in general in the disputer social system is that it's very slow.

47:42I mean If you have to wait four years to have a a decision for one case, I mean many things can happen in your life during that time.

47:51Maybe this is about some money that you need right now.

47:54And four years when you win that case then the the the the thing that you need the the money for it's gone already so it's lost so like um on the one side so We don't want to just dismiss AI because of the potential bad implications of uh being poorly used or in a dystopian way.

48:21But we also don't want to go into that dystopia of uh having some black box algorithm decide things that could be very important for our lives, you know uh if we um have a a claim against someone or some situation from which like uh depends um our life.

48:38We want to understand how this was decided.

48:41And this is not what we get from existing AI systems and this is what we We don't want basically a skynet, you know, deciding what's happening with our lives, if you want to put it that way.

48:52So we want this to be done with um proper uh guardrails and Uh looking as much as possible to rule of law.

49:01So of course, a partner of AI is not going to make a decision that is going to be perfect.

49:06They are basically machines, but they might give us a decision that is good enough for at least some use cases and also good enough at least for some first instance of maybe more serious use cases but very fast

49:20And both parties could see this as as perfectly um acceptable for for this instance and as long as they have the opportunity to to appeal to uh like a human uh or like traditional um s like a panel.

49:36Um and also the the third thing is that We are about to see a new economic activity or a new breed of economic activity that we have not seen uh before in human history which is um this idea of agents and uh the solution of from one agent to another agent

49:59Uh I so this is what we brought Fortunato with whom we are like working a lot on this.

50:04So Fortunato, uh for our For my um you know um Aunt Rosita from Curuzú Cuatiá uh What is an agent first?

50:16Because not that I mean we just assume everyone knows what an agent is.

50:19I don't think everyone should know what an agent is.

50:21So uh explain this what what what they are and why why they matter these days An agent I would say first hello everyone.

50:30An agent is I would say to to Aunt Rosita as someone to whom you delegate some tasks You can why you delegate tasks?

50:40You can delegate maybe because the agent is more competent than you, it executes the task better, or you delegate simply because you want to invest your time in doing something else.

50:52So she can delegate the agent, like to make practical example to go and do the grocery, which is something you know she can do, but maybe she wants to relax, wants to do something else, or she can delegate

51:06the agents to do her accounting.

51:09You know that it's something more boring, it requires more knowledge of the law.

51:13So I would say this is the the the most I would say basic explanation I could give of an agent, someone to whom you delegate some tasks.

51:27Um so if you um have an agent doing task tasks for you it could like basically hire another agent for um those tasks or or that's a weird thing to say Exactly, that's that's a good point, and it brings um one topic on several levels, and this topic is the reputation because the first layer of reputation

51:53Is that you have you to trust your own agent.

51:55Uh William in his presentation, for example, showed how different models can give different output.

52:02So when you are choosing your agent, you have to choose some model let's say that you trust this is the first level so there have been conversation you know about open models closed models what

52:14uh AI model can do their guardrails.

52:18So first you have to select your own agent.

52:21Let's assume this first step is done.

52:24Aunt Rosita has her own personal agent to whom she can delegate.

52:29This personal agent also has its own spending, you know, agent run of tokens, so she has to fund the agent.

52:40She delegates one task and the task we can bring something that I mentioned earlier and the task is doing some accounting work.

52:50What the agent would do.

52:51The agent has two options in this case.

52:54Either the agent starts studying all the, I don't know, Argentinian, um, United States uh tax filing, how to do the perfect tax filing from scratch.

53:06So this takes a lot of time and a lot of token to do this, a lot of consumption.

53:12or the agent delegates to a third party agent who already is ultra optimized for accounting for tax filing in that specific uh jurisdiction.

53:23That's the so Aunt Rosita's agent that we can call um Alice, very um a name we know.

53:33that we call Alice, she hires another agent that is called Bob, and Bob is an expert in tax filing in the United States.

53:43for example and there it comes the transaction where the agent Alice pays Bob to have the tax filing.

53:51Now Where does Kleros stand into this?

53:55Bob can still make a mistake.

53:57The agent that does the did the tax filing, maybe it didn't do it properly.

54:03And it's needed a dispute to settle whether Alice gave the wrong prompt or whether Bob was not offering the service in the right way.

54:16So it was like um uh giving false promises.

54:22So this is how it could work and one last thing is that maybe this is a bit too meta is the discovery part.

54:30So How Alice chooses the right agent in a world where there will be millions of agents to whom to delegate.

54:40And these will go through agentic reputation.

54:44So this is like some new part of the economy that is just you know starting now.

54:49I I don't want to uh go too much into this rabbit hole now because we will have A very very big rabbit hole next week where we are going to speak like uh also about this.

55:01So but you know the the the point is that you have an agent dispute with another agent and uh you know this uh is sometimes um for low value because so sometimes it's just for some little piece of data that the agent collected from some place.

55:17Maybe it's for five dollars, for ten dollars.

55:19Uh nobody's going to go to this resolve those cases except you're an agent.

55:24Because if you're an agent, maybe you can do it um very fast, very cheap, and then maybe you could be interested in in doing that.

55:31So this is a type of claim that was never litigated in the past, and this is the type of uh claims that are going to come by the millions.

55:41because of uh basically this new type of uh yeah of agent this is uh pun intended okay so I guess let's stop Here with the agents, because we will have a lot um more to to talk next week.

55:57Uh we are preparing something very special.

55:59of which I I don't want to to say more.

56:02So thank you Fortunato.

56:04Thank you, William.

56:05Yeah.

56:07Great presentation.

56:08Let's bring maybe for two minutes.

56:12Oh yeah, JB here, our lead of engineering, and he can explain a bit.

56:17Let's say the agents that are the having this um uh disputes are in different chains.

56:24It's like being in different countries, you know, the equivalent to the you know like um um in the economic, you know, in the blockchain world.

56:33You live live in different chains, it's live in different countries.

56:36So what happens and how do those this connect between each other?

56:39So hello JB, how are you?

56:40And yeah, please go ahead.

56:43Yeah, exactly.

56:44So your your agents um right now that there's uh a number of well-known chains that are the the natural home for agentic in interactions uh we can name them it's it's uh mostly base um a bit of the Solana ecosystem as well um

57:04I think it's mostly it, I think, in the Ethereum ecosystem, but it might change because it's changing so quickly.

57:14And that's totally fine.

57:17And as we will probably explore more next week, Kleros can be very helpful when you interact with agents.

57:26Also as Fortunato explained Um if you if you have uh agents on either side of a commercial transaction, you could have uh clearos uh arbitrating pretty much as an as an escrow um transaction.

57:44Um However, um Kleros typically is uh is deployed on a single chain or very few chains because we need to concentrate the the the pools the token uh stake pools uh in a f as few places as possible so it could be uh of course Ethereum and Gnosis chain for v1 or arbitrum for v2

58:09And then we want to have uh uh to reach out to any chains where disputes might occur.

58:17And your agent could um raise a a complaint about um maybe an API or some x402 service um agent that hasn't fulfilled its uh its job and that's where you need the a good bridging mechanism between that chain um let's say base and maybe arbitrum v2 on arbitrum

58:46Um so this is uh even before AI, we already knew that Kleros needs to be reachable from many different places.

58:57Especially from experience, that's especially useful for Oracle related disputes.

59:04Now it's prediction market.

59:09The the biggest challenge is purely technical, um is that on an optimistic roll-up such as Arbitrum, you cannot send the ruling Outside arbitrum without incurring a seven days delay.

59:28If you're using the what we call the canonical bridge, which is the bridge that is enshrined um in the the the ROLA protocol and is considered the the most secure the the gold standard

59:44And then if you try to use any other bridge provider, and there are plenty, you're trading off some Security for convenience uh for faster faster and cheaper uh transactions.

1:00:01So We always were quite uncomfortable with that trade-off.

1:00:08So we launched Vea a few years ago to um find them a shortcut to uh strike a different trade-off um using an optimistic mechanism And I won't go too much in details about what Vea is, but it's purely uh optimistic.

1:00:31And essentially if you're If you're in the in the happy case where uh a message is uh related to the other side without any problem, it might take one or two days instead of seven days, if you're using the canonical bridge

1:00:48And in the in the unhappy case, you fall back on the canonical bridge, so it's just the same as what it would have been otherwise.

1:00:58and you would have um taken much less uh security um shortcuts than going for a third party bridge The issue with this is that even even that is still kind of slow uh nowadays, especially for agentic disputes, you

1:01:19You don't want to wait for for days just for a message to go in one direction.

1:01:25So we uh we were very involved at Kleros with uh a project called Hashi.

1:01:34And Hashi now is uh is not active anymore, but they um open source and shared a bunch of contracts that are audited and work very well and essentially what it does is Um it it's like a multisig for bridge messages.

1:01:52So any for any message you can say okay for this message I'm not going to trust just CCIP or just deBridge or LayerZero with my message to send it from Arbitrum to Base.

1:02:10I'm going to want to have two out of three of those bridges to um to sand it and um And on the other side I will verify that yes indeed.

1:02:23I got the same message two out of at least two times out of the three uh oracles.

1:02:30So very much like a multisig.

1:02:32And It's quite it's quite elegant.

1:02:36Um so if you have uh different integrations with different um security profile, maybe high value or lower value, you can strike a different uh threshold require um more or less oracles to agree.

1:02:54Can you can you say just for for for my Aunt Rosita, not to get lost here, like uh could you say that it's similar if I have to send like a message to uh another country uh and then uh I um I don't send just one messenger with the mail because maybe intercepted whatever but I send three of them

1:03:16And so if the receiver says since that like they don't like are the same letter, they say okay, this is the the right one, right?

1:03:22This could we say this is a bit like that?

1:03:25Yeah, yeah, it's a bit like that.

1:03:27You you would send Three copies of the same letter with FedEx, uh, UPS, and maybe the some other local post and at some and on the receiving side.

1:03:38It's not so much the recipient, it's uh it's it's uh neutral third party which is a smart contract on the receiving chain.

1:03:47Uh so it's neutral, it's uh programmatic, it cannot be changed.

1:03:51And that that uh moving part is comparing the messages and it's only giving you the the letter that uh you were supposed to receive if FedEx, UPS and the other mainman uh at least two out of three maybe agree.

1:04:07So it's like a machine that nobody can tamper with and that machine organizes the letters in the way, right?

1:04:13So it's something like that.

1:04:15That's the best I uh that's the best I best I have for Aunt Rosita.

1:04:18Hopefully you understood it this way.

1:04:20Uh I don't know, Jean, did you understand it uh this way?

1:04:23Yeah, I think it's I th I think it's good.

1:04:26But uh another thing that uh I think it's interesting is that it's uh like Vea was already a lot faster than the canonical bridges and VeaShi, which is the name of this new um of this new setup

1:04:42It's a lot faster than Vea, right?

1:04:44So it it can like in the best case uh exactly, exactly.

1:04:49And Vea is still uh still somewhere at the back as a as a fallback, as a as an emergency, uh as an extra oracle, which is going to be slower, but it could be a tiebreaker.

1:05:01So from our experience when we try to bridge messages I think we had um LayerZero was the fastest one.

1:05:10Um maybe 15 minutes just that's the approximately.

1:05:15I think CCIP was 30 minutes.

1:05:17So if you need two out of three In most cases you'll be good after 30 minutes.

1:05:24And if that didn't uh match then you will need a tiebreaker which could be Vea or could be another uh deBridge or Axelar which we are supporting now is uh VeaShi.

1:05:35Um and so yeah as you understand so VeaShi is Vea and Hashi, the project I mentioned before, um combined together.

1:05:47And yeah, we we think it's it's quite elegant.

1:05:50Um it prevents us from putting too much trust on a single um third-party bridge.

1:05:57And yeah, we can go from L2s to L2s, especially optimistic L2s.

1:06:06Yeah, which is usually very difficult to do.

1:06:11So we could we could say that this is um just one one thing's just um this is I mean Vea was like sending a letter and like uh VeaShi is like sending a telegram that we could say because it's faster

1:06:24Or letter versus I mean leather by horse and by airplane.

1:06:31Yeah, yeah.

1:06:34I'm doing my best, you know.

1:06:38But you know, I think it's it's it's crazy the the the difference of in time like uh for I don't know, twenty-five minutes, uh half an hour and then and two days.

1:06:50It's yeah, that's a lot.

1:06:52That's why the difference is is such that If you're relying on just one of those, you're wondering like what kind of thread-offs do they make to have such uh speed compared to the the canonical bridge that you would get um

1:07:09natively and yeah the best is just um just just get different answers from different providers um And yeah, we've seen like even the the best of them uh they still do once in a while have issues.

1:07:28Um To be fair, for example, I think the main out of the ones we are supporting, LayerZero, CCIP, which is Chainlink, Axelar, deBridge, the one that Had an issue recently with LayerZero, but to be fair, it was it's because a bunch of projects deployed it with the default settings.

1:07:48And LayerZero is in production you're not supposed to use the default settings Actually LayerZero itself is is like Hashi.

1:07:57It can itself have a bunch of different um validators um on the in different data centers and you can choose how many you want.

1:08:09But just to to make like a one comment so let's say these bridges like hold Millions billions of dollars.

1:08:17There are some that are holding billions of dollars.

1:08:19So to attack like a Kleros ruling, you would have to compromise So many millions of frozen funds.

1:08:26I mean he got uh frozen by so many millions.

1:08:29I mean that's that's a lot of million.

1:08:30I mean I would I would be the same, you know.

1:08:32Like uh are are you there you're there, yeah?

1:08:34Okay.

1:08:35I'm I'm here, yeah.

1:08:37Yeah.

1:08:38So uh and and one thing about the uh uh different alternative uh L1 networks because you you know there are uh like players that are uh promoting agent e-commerce like uh circle uh and they are launching arc there is tribe uh with tempo uh with the mp would it be avail uh possible to interact with that

1:09:02Exactly that that's the other big big players.

1:09:05Um so yeah and those those lay those uh L2s are really focused on on payments.

1:09:12And so not so much as uh generic L2s, but really super efficient for payments and of course agency payments.

1:09:21So Um yeah having being reachable um having Kleros reachable um on on those chains is is super important And Vea definitely makes it possible.

1:09:40There was a there was a siren.

1:09:42I think they found you.

1:09:44I think they found you.

1:09:47Are you are you Jean?

1:09:50You know, it's a good idea.

1:09:53It's me, I was really relying on my AirPods to filter this out, but for once it doesn't work.

1:09:59I I one one thing so uh it's uh it's uh somehow relevant for this.

1:10:04You know, I'm reading this book called The Victorian Internet.

1:10:07Um It's a it's a story of how you know the telegraph was born and uh you know the one of the like ways in which it changed the world.

1:10:16We we think the telegraph was that really old technology.

1:10:19But you know, by by the moment that what that was created, um it was like uh world-changing.

1:10:27The idea of sending a message that could reach instantly, you know, uh anywhere, it was like uh mind-blowing.

1:10:33At the time the most the the fastest way to send the message was, you know, like through a horse.

1:10:39So like uh I guess I guess that's that's Via, you know, and then like uh uh VeaShi is you know the the the the telegram which is going to be fast, you know, and this is why I mentioned this.

1:10:52You know what was one of the first um use cases that they they they had for for the telegram.

1:10:58Uh you're not you're not going to believe this uh because I I could not believe it.

1:11:03So Um that was at the time when also there was time to deploy, this is mid-19th century.

1:11:10Um people were also deploying like railways and the railroads.

1:11:14Um you know, if if some guy would rob a bank uh and in some city, take the railroad to and escape I mean you you cannot be waiting for him at the next station because the time you don't know it.

1:11:29I mean you have to send a horse or someone in the rail to so what The telegram did so since they were running through the railroad lines, okay.

1:11:38Oh, this guy from the bank sent this by telegram to the next station and they were they would be waiting for them on the station to to stop them and arrest them.

1:11:47So it was like a law enforcement uh technology at the at the beginning.

1:11:51Um I still haven't finished the book, but uh it's it's really really interesting.

1:11:55I mean how this was um developed the technology was developed as as always by two different guys in different places.

1:12:03You know, obviously it was one in the UK.

1:12:06uh and one in the what the in the UK was a guy called Cooke and in the us it was uh Samuel Morse uh and um interestingly you know the morse You know what what Morse did before like uh becoming an a telegraph like entrepreneur?

1:12:22I mean you're not going to guess this.

1:12:26He was a painter.

1:12:28He was like a portrait painter.

1:12:30And he was in Europe for like some tour or whatever.

1:12:34And then apparently he was he was American.

1:12:37So he was coming back to America.

1:12:39uh and then in the in the ship I mean uh transatlantic ship I mean he ha he had this conversation with the guy who tells him about the telegraph new technology and how people were starting to think about this

1:12:52And the guy become like becomes obsessed, you know, by like uh this story uh of telegraph and then starts researching and he was in his forties and he like he was uh back then you know like if he was an old guy already and starts you know uh investigating and he's the the guy who invents you know like a Morse code and everything you know.

1:13:09So this is a really, really interesting like um uh sorry I mean sometimes you're in this uh um you know sheep and you know and then one conversation with some guy can change your life forever, you know like uh

1:13:23You know, have that in mind next time you're on a flight and you don't want to speak to the guy next next to you.

1:13:28You know, maybe who knows who that guy is, you know.

1:13:31So yeah.

1:13:32So basically, okay, telegraph, VeaShi.

1:13:36males, you know, like my horse is a ha ha that's that's the best.

1:13:40But really do do read that book.

1:13:41It's it's it's a really good book.

1:13:43Um uh and also I'm starting to see uh the social implications about how it affected the news industry and people fake news, but it started more or less like a law enforcement technology to to catch like bank robbers.

1:13:57So there you go Okay, anything else we want to say about Vea?

1:14:01Because this is already I did I think the longest ever you know community call.

1:14:04But so much we have so much content, you know.

1:14:08Um well not too much more.

1:14:11We uh maybe Jean mentioned it already.

1:14:13We have a blog post that just came out.

1:14:15Um yeah, there's a new and improved tooling.

1:14:19Um So yeah, um take a look if you're interested in infrastructure bridging.

1:14:29Okay.

1:14:30Okay.

1:14:31Thank you.

1:14:32Thank you, JB.

1:14:34Anything else we want to say, Jean, today?

1:14:42We were expected to do like a very short call today and next one would be like longer, but I mean if this one was this long I don't want to imagine you know next one is probably, you know, you need to just uh yeah, go don't make plans for dinner, you know, after your

1:14:59Okay.

1:15:00Okay.

1:15:01Um let's let's then stop here.

1:15:04Uh guys, um thank you very much.

1:15:07Um have a beautiful week.

1:15:08Let's We'll see you all next Monday for the Spanish call and then Wednesday for a very very special community call.

1:15:15Bye guys, see you.

1:15:17Thank you.

1:15:18Bye everyone It works.

Want to be in the room? The Kleros Live Stream runs every Wednesday at 6PM UTC, with the Spanish call on Mondays. Join the next one, or subscribe to get these recaps in your inbox.