 back to ThinkTech. This is transitional justice and our guest today who joins us from Lisbon is Bruno Adele Papa. Welcome to the show Bruno. Thank you, thanks for the invitation. It's nice to see you here. Thanks for being with us. So Bruno is a physicist but he's also a computer guy and today we're going to talk about how AI and other data processing helps project expedite justice to its work in Europe and elsewhere. So let's first start with how AI is different than the data processing that we used to think about. So yeah, I think it's a good point to start. Well, so first by AI we already mean different things. So there's different kinds of models, different kinds of problems that AI tries to solve. Broadly from our AI combination of so machine learning deep learning models that have evolved in the last 20 years or so with the computer capacity that has also evolved. So we have ways of dealing with large volumes of data that are necessary for AI. But those sometimes people don't, when they mean AI algorithms, they don't include those. They mean more like the models that quote, think like humans, which they do not. That has always to be very clear. But AI is a broad term for that. So how it's differs from normal or let's say old statistics is just by the size. So statistics in the past has been able to deal with what are reasonable amount of data. But today we just increase the amount of data available by many, many orders of magnitude with that. So internet, computers, processing capacity and so on. And like AI algorithms were a way to deal with this. So it's a very fancy new word for applied statistics, very specific methods and algorithms that have developed in the last years or to deal with this data. Yes, that's probably a short answer. Okay, well, no, that's, that's fine. I wanted to get some distinction going. You know, when we, when we think of AI, we think of facial recognition for one thing, you know, like Xi Jinping has got, you know, the Western China and the Uyghurs all wired up and they can recognize any face and and he's doing that actually all over China. And for that matter, the UK has a lot of cameras around London and they can recognize, you know, using AI, they could recognize faces. So does this help in project expedite justice's work? If you can recognize faces using AI facial recognition? So at the current moment, that's not anything I'm working on, but it has the potential to help. Essentially, as you mentioned, there are many, many different models for facial recognition. I think it's maybe the one classic case that governments have been using. And you can use that if you have enough data to train your model. So one very important feature of AI is AI is only as good as the data you use to train your models. I think that's that is very important. So in the Chinese case, they do have a crazy amount of data that they use to train their models. So they can easily identify some kind of key facial points and facial features. I'm sure in the UK, the government has similar approaches. What organizations such as PHA could do is if we need to, we have videos that have been recorded by witnesses by media or that sometimes people post on social media. And in these videos, you have faces that could appear. Can we identify them? Potentially. We can identify the same person, the same face appears in different videos. And that's just as a comment, it's not only for faces. You can identify, for example, carbs, vessels, or anything essentially, anything that you can train the model to identify, which is something that is quite new, something that has been developed in the last maybe 10 to five years. Well, you know, that really speaks of the network analysis. What I mean by that, maybe that's an old term. But if I have, say, a stack of email, let's say I have 20 million emails, okay, and I want to see where you, Bruno, have appeared in an email with regular data processing, I can, you know, and we have been for the past 20 years after 9-11, the US government has used that to go through those emails and find every time Bruno's name appears. But now, with AI, we can do it much faster, and we can get more reliable results. Can you talk about how that works? Yes. So it's not only a matter of much faster, it's that you can also infer the context in which my name would appear in the emails. It's not only like, let's call it a brute force search that you go through the text very quickly. You also now have models that are able to understand some of the text and understand in which context this name has been used and to what's other names or what's other words or what's other companies this name has been linked to. And also, you mentioned only emails, but what is the case in realities that you have different kinds of data data, for example, emails, videos, registers of companies in some country, they are all in different formats. So we also need some kind of software that look at these different formats, which is one thing that was not possible before. I mean, now combining different kinds of data is one of the key points. Yeah, that's really interesting. And I'd like to jump off that track of thought now and say, well, just as Project Expedite Justice and say prosecutorial authorities in Europe looking for war criminals can use this technology to get a beat on where you have been, Bruno, what you have done, who you have consorted with, what points of life experience will help to define you. The government can do that too. And actually, the government has greater resources to do that. So it's the old story about how technology has two sides to it. What are you finding in that regard? To study AI means to study both sides. Where are governments, especially autocracies going? Well, it's always, well, I try to be optimistic about technology in general, but I tend to see AI as a very powerful tool that everybody can have access to. I mean, you mentioned governments, but it would also add these huge tech companies, for example. They can train models on people when what information people openly put online on their profiles, their social media profiles, show their friends they want sometimes to have, I don't know, some funny glasses on their face images and so on. And people just give this data for free and not only to governments, not only to autocracies, they give it to companies that own those piece of software, right? Where this is going? Well, I like to think that this will be eventually properly regulated and not abused, but, well, we are all humans. We kind of know that these will eventually be abused, that it has been examples in some cases, right? So autocracies tend to use that to, well, consolidate their power, right? I was thinking about this before, and let's say when the computers were invented, right? You could, of course, use computers to scientific research to, well, essentially to help people, or you could use computers to, well, use new technologies to control people. With AI, it's exactly the same. So I just think we have a much, much more powerful computer or comparable to the computer invention. You use the word train, and I guess the context was train an AI model. How do you do that? And can anyone do that? Can I do that? And also, does it require coding or just presenting data? So yes, I mentioned training on the context of training AI models. So more specifically, these models are, well, deep learning models, which is like a particular kind of model. So how does it work? You basically have a model with lots of different variables that you want to find the right value to process your data. You don't know. So we just showed the model a lot of data. For example, the simplest example is you want to train a model to detect, to differentiate cats and dogs, right? You show this model a thousand images of dogs, a thousand images of cats, and you tell the model, this is the dog, this is the cat, now you do it on your own. So in a way, the model is generalizing the data that you presented. So that's what I mean by training. Sorry, what was the second question again? Can anyone do it? Well, it has become easier. So at first, programming was a very, well, difficult task and you'd have to study computer science for years and understand how programming languages work. It has become easier and easier. So a little bit more high level in a way. A lot of the problem to train a new model is not on the knowledge, but on the hardware, because you need very good hardware to do it, which normally people do not have. You can train a simple cat-dog model in your laptop. Most people could by studying a little bit of cozy, but this is not enough, right? If you want to train a model for recognized spaces, you need a much more powerful hardware, much more powerful computer, and now people have GPUs and all that. Normally the way it works, that's they use GPUs from companies, that many companies rent them. So for example, I don't use, I have my laptop, but I would train my models remotely in some companies cluster of GPUs. So yeah, that was an answer. It's becoming easier. Here's a third part of it. Yeah, sorry. Please continue. Now, the last very recent test on how it has become easier and easier is this advent of large language, language involved is like chat tpt, right? I think this is like a topic everywhere now and it's interesting, even for me, how quickly it is developed. So just before chat tpt appears, I knew about some language models, but I was away for a few months away from from work. And then when I came back was just like two weeks after chat tpt had been released and a few other models quickly appeared. And you see how quickly that's changed because now if I want to, for example, if I want to implement this cat and dog model, I can go to chat tpt and say, please write me a piece of software that will implement using this or that model, an AI software that will differentiate cats and dogs. And if you give you the code for some few mistakes that you have to catch, but it essentially gives you what's in the past you spent a day or maybe even more writing. So I see AI is kind of making itself a bit easier as time goes by. But sorry, there was a third question that I think. No, I just wanted to know if you if you had to know any specific coding language in order to set up an AI program or whether it was agnostic? Well, it's, in principle, you can develop in any programming language, but most of the work has been done in Python, which is a specific programming language. You have frameworks developed by, well, big tech companies nowadays that are introduced. So you can easily learn how to, I mean, easily. I mean, if you are, if you have some little bit of background in math and coding, you can understand, read all the documentation, ask questions to chat tpt and kind of teach yourself and implement that using this free software that they provide. Now, which is a little like a little bit of a detour, but it's another thing that I appreciate about AI. A lot of models are open source. So if you want to write your own model to detect cats and dogs, the standard procedure wouldn't even be let me write my own model. It would be let me check at this database of free trained models, what people have have done. And it already like some obviously not commercial companies do not open their codes because, well, they are how they make their money, right? But a lot of people contribute to this open community of software, which I find really, really interesting. And you kind of build on the work of other people. It's like, you know, every, every other kind of popular software, the development of the popular software, if you go on the web now and you look for AI programs, you'll find hundreds, hundreds. And some of them are way better than others. Some of them have better data, better models, and so forth. So I, you know, I want to mention this, this movie I saw. And it's actually a play now in London. And it's called Operation Mincemeat. And it's about successful operation during the war, 1943, where the Brits were able to deceive the Germans about where exactly where the Allied forces would be attacking. And the question is whether it would be Italy or Greece. And they deceived them into believing it was Greece. And then they attacked, actually attacked in, you know, in Italy, as we all know. And that was a turning point. So, you know, what's interesting about that is that deception is part of AI. I mean, I could make, how do you know this is my voice? How do you know this is a picture of video of me? How do I know it's a video of you? Because I can use AI and deceive you in every which way, as to the voice and the video. And what you are saying, I could create a completely fictitious deception using AI. And it would be very hard to decipher that. Am I right? Yes, this has been a very recent development. It's called generative AI models, which basically are models that creates, as you mentioned, voices, images, videos, music, even bubbles. Well, they can kind of copy a voice of someone or a pattern, a musical style, and so on. There are people also developing models to detect them. But it's a little bit of a competition, right? You develop a model that creates a better fake, you develop another model better to detect fakes, and they kind of feed on each other and get better and better. But the problem is kind of always there, right? This is a little bit of this carry part, like it's also very, very difficult for people working with law, right? Because how can you say now 100% with 100% certainty that something is really a piece of evidence, or it hasn't been fabricated, right? And this is an area that is very new and very difficult. I don't think we have a final answer for how to deal with this problem, right? And yeah, I mean, it could ask me, I wouldn't know the solution, actually. Well, you know, we talked about, this is an interesting point you mentioned. We talked about how the white hats and the black hats may use AI for different things, for good or for nefarious. But there's another element too, and that is that the AI can be used to see if the other guy is sending you AI. Like, you know, for example, chat, CBT students at very important universities have written their term papers using AI, but then there are other students who have written AI programs to identify the programs that are false. And so you get two sides, two sides of that equation also, the AI to search out the AI. I have seen, recently I saw a very funny joke that one professor was talking and he said, yeah, almost students keep saying, I might write my papers with chat, CBT, and he said, no problem, I might grade their paper with chat, CBT as well. He was like, okay, I'm doing everything. Well, I think, you know, after a while it's going to be the normal, everybody is going to use AI to write papers and the AI that's supposed to spot them won't be as good as the AI that wrote them. But I think what, you know, what's interesting, and I'm coming to my case study with you, I want to have a case study with you. So if I want to plan something, I want to plan an operation. If I want to plan an operation in mis-meet, for example, why don't I just trot out, you know, chat, CBT and say, how can I deceive the Germans into believing the invasion will be in Greece? And it will immediately provide me with a plan, won't it? So therefore I could use AI, chat, CBT anyway, to start my planning. And I wonder if that is so when you are looking to hold war criminals accountable? Can you have AI write you a plan to find them and to, you know, deal with them and to hold them accountable? Is that a worthy case study for you? Well, I think, I mean, we have to think first about how humans do it, like how the human be able to hold someone accountable for war crimes. So you have today's, so you have a set of, so you have the law set of rules that this person has to break to be considered a war criminal. And you can, in principle, have, this is not the case, but let's assume you have the best data sets in the world that contain all the possible information about everything that everybody has ever done. So in that case, you could have like a very, an AI that looks at all the data and have a clear cut. Okay, criminal or not criminal. One thing that's a little bit dangerous in this case is, so AI is just as good as the data that's used to train the models, therefore biases are going to be there, right? And if we do go in a little bit of dangerous territory, especially in cases that involve data about, well, racial profiling and socioeconomic studies and all that, just imagine how a human would generalize some kind of statistics, the AI would do the same. So we still need someone to kind of check if the AI is actually doing what it is supposed to do. And if you can play, you can use an AI to check if the other AI is doing what it's supposed to do, but I wouldn't, personally, I wouldn't say that it's very, very easy. Like we still have, humans still take context in acting to account much more than AI. If that ever will change, that's, that's a question for the future. I guess it's more of a philosophical question about what's made this experience. You know, at these days, you know, when I practiced for a long time ago, there were very few exhibits, very few documentary evidence, very few. And now, you know, it's not uncommon to have exhibits in the millions, which have to be evaluated and which have to be authenticated and presented to the trier of fact. And the trier of fact could use AI to evaluate that. And the lawyers on both sides could use AI to sort through those millions of documents. So this obviously would be helpful in a trial of a, of any trial, but certainly in the trial of a war criminal to find him or her and to to line up the evidence and authenticate the evidence, present the evidence. And, you know, in Ukraine, for example, we, we know through videos and testimonies and the like that there's an awful lot of evidence about an awful lot of war criminals. And so, but it hasn't really come to pass where these criminals are being taken in front of tribunals and tried and punished. That hasn't happened yet. And maybe what I hear you saying you can correct me if I'm wrong, but I'd make a guess and say when that day comes, when the people you have identified through various sources as potential war criminals are in front of a court, AI is going to help you, not only identify them, but try them. Am I right? Yes. And, and help to identify them and to identify all the pieces of evidence that's helped to prove the all the, and the crimes, let's say, because as you mentioned, for example, let's say you have thousands, hundreds of thousands of hours of video about different events in Ukraine, for example, there's no lawyer in the world that's going to go through all those videos look one by one. I mean, you can do that, and you take 100 years to be able to actually identify everything. But you can use an AI tool to identify, for example, cases that appeared in every video. So in which seconds, in which frames of the video do we have this person, in which frames do you have the car of this person, and where, and this can be done much, much quicker than a human do. So I think this is, as of today, I think this is probably the main contribution of AI is to help sort through data and collect potential evidence to bring someone to trial. In the trial itself, I'd still say it's a little bit for the future and for lawyers to decide how much we want that to be the case. Then that's a little bit of the end of AI is a tool, which is what I've been kind of selling so far, and more to AI as a decision maker, which is it has complicated generalizations. If you use that in court, why not for the government? Why not for decision making in the government? This is a completely different territory. I think as of today, the best investment in AI and data analysis is how can it help us to sort through and identify evidence? Yeah, you're really, when you talk about AI as a decision maker, a little black box that determines guilt or innocence, that's really science fiction. But I agree, it's irresistible. It's irresistible and it will happen at some point, don't you think? Well, I think it depends a lot on us. AI will always be a black box in the sense that every model has billions, maybe two billions of parameters now. No human can go through all of them and understand what every number is doing. We hope it generalizes things in a useful way, which it does because, you know, there's just so many numbers, so much math. You are kind of just crunching your data in a big pile of math. But it's a black box, like we, as humans, as a society, we want to trust the black box. Last, I also think it's important to comment that you have what people call adversarial attacks, which are ways to break the AI. So the classic example, you have a picture of an elephant. You put a little bit of random noise, random pixel colors, and suddenly it's not an elephant anymore. It's, I don't know, a panda bear or some other thing. Why? We don't really know. Lots of studies go into that. And until you understand, or I don't think this is actually going to be possible, but until you are sure that these attacks are not possible, personally, I don't think we should use AI as a decision maker because it will make mistakes, even this kind of weird data or, let's say, not weird difference or very different data points. Well, the genie is out of the bottle. And as I said, you know, Jack Smith, who is prosecuting Trump, could probably use AI in organizing all those materials in various cases. And the prosecutors around the country here will think of that, and they will think maybe they should use AI too in examining the documentary evidence and so forth. And the judges will want to have the benefit of that. So what you have is a perfectly legitimate and constructive use of AI in enforcing the criminal law and maintaining our democracy, if you will. But at the same time, you have autocrats who do precisely the same thing in order to destroy democracy somewhere else. So it goes to something you've referred to a couple of times for enough, and that is regulation. And you've kind of skipped that stone across the leg. But what about regulation? How in the world do you regulate this? So for the moment, in our case study, I'm going to make you the American Congress. You are now designated as the American Congress. What in the world can you do to regulate AI? Well, there has been some tries. So for example, that every company or everybody that uses AI should state that this is an AI-generated piece of data, image or voice or whatever. But this is the easy part. Like you can regulate and say that companies cannot use data that has been acquired without consent to train their models. They can do that, for example. But that's regulating the easy part. Will it be done? And how can you check it? Like how can you check if the data used to train a model has been acquired with consent? How do you go through this millions and millions of pieces of information? It's a hard task. I don't know. I don't have an answer, actually. My answer as the Congress would be, well, the companies producing the AI have to kind of convince us that this is a tool that's a safe tool. Same as, for example, computers. Like we are convinced that using computers, that computers are tools that help us in everyday tasks. We know they can, every now and then, they break. There's a bug here and there. And sometimes things don't work as we expect. But mostly they're fine. How did we get convinced of that? By using it for many, many years and improving, like constantly improving, fixing mistakes. I believe a Congress should look at AI in a similar way. At the beginning, right now, there will be tons of cases of people breaking violations, of computers exploding here and there, computers being, I don't know, computers misbehaving and having accidents and even killing people as if it happens. But ideally, we should strive for incremental developments. And slowly, but surely, get a better computer in the future, as we do today. Well, I made you the Congress for a reason, because Congress presumably, although one wonders about it these days, has the power to regulate. But how can any legislature reach out and say, okay, we're all going to do this with pure heart? We're not going to be moral. We're going to be ethical. We're not going to take advantage of anybody. It's like you'd have to write a new constitution of human affairs in order to make sure that it was constructive and helpful to our global society. I myself, I don't think that's going to happen. I'm sorry. I completely agree. But, well, another thing that is important to mention is that I would be more comfortable with is knowing that, well, at least the people making the decisions in the Congress would know what they are making decisions about. Because you do have examples in many countries where clearly they have no clue how these things even work. And so those, as a Congress, are the first things of like, at least get advisers to explain you through a basic level how this works. You don't have to go into math. You don't have to go into coding. But some things are very, you clearly see that people come from a world in which this was not a thing. And this will change in the future. The current generations are growing up in a world in which AI already existed. So how this is going to affect their decision maker in the future is a question that I can even answer. Like, I think I wouldn't like to be in this responsible decision maker world. It's a very, maybe the hardest problem of our generation. I totally agree. And you really struck on something that is the people who make the policy or don't make the policy, people who are responsible to make the policy have to learn how it works. And right now, it's hard to find policymakers who really know how it works. So I think that's really an important point. And we have to do what we can to educate them or to cause them to be curious enough to learn. So we're almost out of time, Bruno. And I want to ask you one last question. It's my five year question. Okay, look forward, if you will, five years. Look forward in Europe. Look forward in the Ukraine war. I hope it's over. I hope it's over very, very soon. Matter of fact, look forward in government here in the US and around the world. How will AI change those things? How will they be affected for better or worse? Well, first of all, this is an impossible question to answer to 100% uncertainty. As an example, people who have taught self-driving cars would be already a thing. It could ask them 10 years ago, and we are not nearly there. How will I would change? So my, let's say, wild guess is this large language models like chat CPT will be more and more incorporated in our everyday lives. And we won't do like investigative analysis without them anymore, because a lot of people will just incorporate them into their work. In terms of work, we see that already. So like drones, for example, drone attacks, like they are all coordinated by algorithms that have been developed with AI. So this kinds of warfare are going to become more and more common. And they're going to, well, they are already changing like traditional, traditional warfare, like you've seen Ukraine was able to attack like Russia, I mean, or drones have been struck and deep behind the front lines. And this is just going to get better and better, like better drone controlling algorithms, better, well, better data collection tools, better recording, more and more data and more, well, let's say AI should deal with this volume of collected data. But my personal guess, large language models are going to be more affecting our lives and all questions in five years. Will it be a better world? Well, that's very hard to say. Depends what you mean by better. Okay. Good answer. I'd say as someone working with tech and AI, I try to make it, but well, if it works, it remains to be seen. Well, goes to show you're a very important person. And anyone in your shoes is an important person because this will have a huge effect on our society. Brutal Hotel Papa, thank you very much for joining us from from Lisbon. Thank you very much for discussing these things with us. Well, thank you very much for the invitation on the opportunity. Yeah, that was a very nice talk. I love her. Bye.