Building Ethical AI: How Decentralized Justice Aligns Artificial Intelligence with Human Values

Transcript

[Music] Hello, how are you? I'm very happy to be here again. This is like I think my my seventh talk at HCC. Uh the first one was 2018. So, you know, bull market, bare market, the COVID, you know, World War II. I mean, you can count on me still being, you know, here on stage. So, Oh, thank you. Oh, you don't need um Okay. So, I'm a founder and co at Claros. is a decentralized justice protocol that is used for resolving disputes. The case people typically know is Alice. She's a designer and she makes a contract with a u like a freelancer called Bob, right? And for uh she sends the payment into a smart contract and

then they agree if there's a dispute they're going to use to resolve to resolve it. So if there is a dispute uh the case goes to Claros money stays locked into this escrow and then a pan of jurors is going to see the evidence see the case and vote on who they think is right and so let's say the majority here thinks Alice is right so Claros makes a ruling and the smart contract sends the money to to Alice okay so this solves the problem and Claros uses a token a token called Pakium for coordination of jurors it's based on shelling points and well there's lots of talks of me um presenting this so you can watch it.

So um everyone is talking about AI now but so we have been u experimenting with AI since 2018 uh before way before it was hype and in 2018 we launched the first version of the protocol and we made an experiment called doson trial that was an experiment to test how robust claros was and it was uh it was like this. So users could um submit an image and jurors would decide whether it was a Dodge image to be accepted into a verified certified Dodge list. And if you sent an image and your image was accepted into the list by the jewelers, then you earn a small reward. But the

jackpot here was if you could send an image and full jewelers that it was a dodge, but it was actually a cut and then you could make 50. It was $12,000. So, um, this was a very big incentive for people to try to to break it. So, this is how it worked. Alice is a submitter. She submits an image of a dodge a dodge. People can check it. And here nobody has any objection. Uh, and then the she gets back the deposit that she puts when she makes the submission. Um, and now she submits a cat image to see if it can go through. And Bob thinks this is not a dodge. This is actually a cut. So he can challenge the submission with another deposit. And then what

happens? There is a dispute between Alice and Bob. And guess who is going to solve this? Of course, Claros is going to solve this. And a panel of jewelers is going to be drawn and they're going to analyze the case. And of course, this is a pushing cut and not a dodge. So it's going to be rejected and Bob gets the deposit made by Alice. Okay, this is how the mechanism worked. And here you can see a bit of the images that people sent. Uh these were accepted as does that these were valid submissions. And this here you can see some attempt of sending you know cats and people were doing Photoshop to try to make a dodge look like a cat. Um a cat look like a like a do. So well people tried lots of

different strategies. And after we finished, we ran some AI algorithms through this to see if they could recognize the dodges, right? And we found that AI had 97% precision in actually recognizing the right images as as dodges. But it's not sufficient because 97% is not uh 100%. There was this case which was very famous called the cut in the snow case. Someone submits this image and uh they tell me look uh this is accepted into the list and they tell me okay you need to pay me because I got a cat image into the list and I tell him hm this could be a cat. Uh and this is

what he sent me. Uh this is the other side of the picture. And I told him yeah this could be this cat but could also be this dodge. So I'm not convinced that this clearly shows uh a cut. So what did we do? There was a trial done in Claros for the 50 that were promised to the submitter that broke the system and jurors had to define whether the image clearly showed a cut or not. So this is the image. Who thinks this is clearly a cut? Who thinks this is not clearly a cut?

Okay, this the big majority of course and in the Claros case the no option one two to one. So this was the first time we tested how AI and CLOS could interact and this is summarized in this article. This is collective versus artificial intelligence and this is where we developed our first hypothesis of how we imagine this solution to work in the future through a combination of technologies where AI is going to be used for the simpler cases. Uh the centralized justice is going to be used for like a bit more complex cases where you have like crowdsource jurors and then for even more complex cases that require legal knowledge or more like sophisticated reasoning then more

traditional methods of solution may be may be used. So then came the AI agents right um but this is not that new. I had an interview with Max Kaiser on TV in 2016 and here I was actually a bit speaking about um well I check for what we're going to do in the future will be in virtual reality world. My avatar could sue your avatar on this kingdom of virtual reality and there would be a jury to decide who is right. The jury we didn't define if it is human or not human could be perfectly robots and the defendant could also be robots. Maybe one robot could sue

another robot because he didn't do some job well and maybe they should go to trial and other algorithm made of robots could settle the case. So robots and VR are within the limits of possibility from here to 10 or 20 years. So I mean that was quite precient. I mean I was very very right about um the time frame and what would happen. I was very very wrong about calling these uh VR worlds and you know bots because if I had called them you know metaverse and AI agents I mean VCs would be throwing money at me but you know I got I got the marketing wrong about this. Uh but I mean the time is happening now I said 10 years and so it's from 10 10 to 20 and

this is starting. So now we are starting to experiment with agents and I'm Argentinian so of course uh the first thing we think uh is football. So we started doing some tests with using Claros jewelers as a kind of var a var version that is crowdsource. Um so we started testing it with um actual football um you know situations. A very famous situation was uh in the World Cup of South Africa 2010 there was this goal by Frank Lampard that was disallowed and this is what happened. Um I don't know who thinks uh this is a valid goal. This is Yeah. Okay. This is

a valid goal. Okay. So we put this on uh Claros and we got three jurors each of them with a different agent like one was Gemini, another was another was Claude I think and they all agreed that okay yeah this was a valid goal. So this is all all good and jurors gave justifications on why they thought this was well. So juror one gave this justification. Um juror two gave this one and juror three gave this one. Remember these are like bots. These are not humans uh doing this work. I mean humans are actually just receiving the case turning around loading the case into GBT and then copy and pasting the the decision. Okay. Um

okay so first case everybody agrees. Let's go to another one. Uh this one is uh collision between no and in the 2014 final. Is there any German here? I mean no you cannot participate in this one. So um uh so well was this a foul or was not this a foul? Okay let's see. Okay. Who thinks this was a foul?

Yeah. Okay. Who think this was not a foul? I I mean that is a 50/50 right around and uh we put this um AIS and the models don't agree on the decision. So you can very well have AIS where the decision made by different models are different um seeing the same evidence in the same the same case and there is a justification. So the the AI one said there's no foul no foul and justifying the second one is foul because it's a dangerous and excessive force. I agree with this one and the third one said uh it was a no no foul. Okay. So it's can you see that you have different situations and different decisions made

by different agents and this same logic we are starting to apply for example copyright law you can have lots of IP claims that you can resolve through agents for example think this one there is there a concerning similarity between image one and the copy uh well this is super uh relevant for lots of IP claims IP decisions that need to be made in this case. Um the decision was made that two votes against one that there is actually a concerning similarity. If you want to learn more about this type of IP use case, I recommend that you take a look at the harmony mediator B. This is a B we develop that looks into this kind of features and looking for percentages of

similarity between images. So this is where we're going. So AI judges are cool because they can deliver very fast and very affordable justice. They can resolve cases very very fast in matter of minutes. But of course you know there is a dark side to all this um because AI can be misaligned with our goals and values and this was very clear in the world of social media. So I recommend this book that explains how Facebook in the early days developed this algorithm with the goal of um curating the feed of people.

Uh and this was supposed the the goal was to maximize engagement of people with the platform. And what happened is that the AI discovered that the way to maximize engagement was to maximize moral outrage showing you the kind of information that would like trigger you against the other side. So um this is a way where the AI strategy was misaligned with the goals that were intended to to serve and well you know a lot of the things that happening in our societies these days might be I don't want to say everything but you know it might be linked to this misalignment uh situation and also there is a problem of bias um there was this tool was used in the UK

in Durham for making parole decisions so the AI tool would predict uh the percentage of like risk of some individual that would get parole of reoffending or not reoffending. If the like percentage was low, the individual wouldn't go free. If it was high, it would be kept in custody. And their model had different variables and at some point they started discovering that it punished unfairly people who were from poor areas. Everything else equal if you were from a poor area. I mean you would get a higher value of uh risk and you would remain uh in jail unfairly.

So lots of people say okay yeah but cool no problem because you can still have humans that will verify this and make a final decision in case you don't agree. But it's not so simple you know uh empirically because let's say you are the parole officer in the jail and you have to decide whether to free this guy or not. you have a meeting with the guy and um then you think he's decided to reform, you think he's okay, he's not going to reoffend, but you know the the algorithm tells you that this is a high-risk individual. Um I mean like you have to think now whether you will obey the algorithm or you will like overturn it. You have the power to overturn it but I mean what

would you do guys? I mean who would like overturn the decision of the algorithm? Okay, maybe one out of 100, I don't know, who would not overturn? I mean, who would accept the decision? I'm seeing like lots of people fearing to raise their hand. Of course, I mean, what if you overturn the decision, the guy goes free and like two months later kills someone, right? Uh people will come to you to ask you, hey, this algorithm said it was high risk. Why did you free him? I mean there is no incentive at all for humans to overturn decisions made by AI. Um this is why there is this famous quote you

know no one was fired for following the algorithm as before no one was hire fired for buying from IBM now everyone will just obey the algorithm and this will result probably in a situation where our societies might turn into like what we call an algoy it's a government where decisions are made by AI or very much influenced by what AI says right if you want to learn more about this there is this article I wrote of the American bar association uh where I explain all of this situation and and how this can be solved a bit and um the main question that we have in this this context is how can we align AI with our human goals and um there are like two main strategies. Um one

strategy is about certification is to prevent misaligned AI from even happening. This is something you do before you deploy the AI. And the other one is the kill switch strategy. So it's more like okay, the AI is deployed, it's working, and now you need to uh disabled it. And the thing is that I mean AIS that are deployed, they really don't like to be disabled. I mean, they just prefer to just, you know, um I'm afraid I cannot do that Dave situation. Uh so I don't have time to go through both. I mean I will keep the kill switch uh strategy and they're complimentary but I will kill do the kill switch situation and then future talk I will speak a bit

of how the certification approach might work. So imagine we want to create a safe AI certification that y list models that are acceptable for being deployed and as I explained to you before the versions on trial um pilot this can be used also for AI's on trial list. You need to define uh well a list of criteria that you would uh deem safe for an AI model to have. I mean from the science fiction world you can have isak asimov laws of robotics that is mostly you know about a robots cannot injure you know humans but more seriously now we have for example the ailomar

principles asar was a conference held in California 2017 where a bunch of AI experts get together and they define a number of conditions that they deem are important for AI models to comply with so they are safe for us to to use. Um so based on all this uh of this information you can create a list of um conditions that you deem important for an AI model to have. This could be about transparency in development could be about uh testing for for bias could be about robustness, safety, ethical compliance. you we can think what are the criteria we want to use and then you have a model a mechanisms looking like

the do trial so you have Alice who wants her model to be whitel listed to be used in the different applications so she makes a submission of her model with a deposit people can verify it and see if it's safe or not if it complies with the list of of rules that we defined if the model complies then it gets the certification and it's whisted to be used. Now let's say she submits a model that is biased or malicious in some way or non-compliant. So Bob thinks this model doesn't work. I'm going to challenge it. So now there is a dispute between the submitter and the challenger and the panel will decide whether this model is or is not acceptable. Let's say

the model is deemed not acceptable. So what happens now is that uh well the model is removed from the white list so it cannot be used and Bob gets the deposit. So this is a mechanism that helps create a list of certified models that will comply with a number of characteristics that we uh deem desirable for for use. So I'm getting to the end. So let me just recommend some um readings if you want to learn more about all this. This is very fascinating. As you have seen, I've been following these topics for a very long time. So, first this vitalic post about the promise and challenges of crypto and AI applications. Uh, this is from 2024. And then two very good books.

The first one is quite old, weapons of m destruction about well how um bias and human compatible is by one of the I mean godfathers of AI Steuart Russell, a professor at Berkeley. So strongly strongly recommended to understand these relationships between AIS and and humans. So conclusions AI judges are here and they bring many benefits in terms of access to justice. So they are great. We should not fear them but we need to put them into the right mechanism because misalignment um and bias can be very big problems uh for human rights basically. So we have different ways to address these problems and one of them is certification uh and

other one is skill switch mechanisms which I will uh discuss in a future talk. Uh so yeah uh well thank you very much. Hopefully you learned something from this and yeah thank you. [Music]