Keynote - Ten Key Questions that a Company Should Ask to have Responsible AI Keynote

Speaker 1 [00:05]

So thank you for coming. I have never talked to developers, so this is a new experience for me. Usually I talk in scientific conferences. And basically I want to talk today about responsible AI. So how you use well AI, and I will explain this through 10 different questions. Now, why responsible AI? Some people use the term ethical AI, but ethical AI doesn't exist. because ethics is a human characteristic. So if we start using words like justice, hallucinations, trust to machines, we will lose part of what we believe is human. So I don't like also to use trustworthy eyes. Sadly, both the US and the European Union have research programs in this. Why I don't like it? Because if 100 years ago a person would come and said, I want to sell you a ticket to go to Paris in a new transportation means. I said, okay, what is the name of that transportation mean? Oh, it's trustworthy aviation. I will not buy the ticket. If you need to put trustworthy on front, it means that something is wrong going on. And we know it doesn't work all the time. So while we are putting the burden on the user, we are saying, you need to trust us, but we know that doesn't work all the time. So that's not ethical, so that's why we use responsible. Responsible is also something human, but because of law, responsible has extended to institutions. And I use responsible in that way. Who is the institution responsible of, for example, a product in the market? For example, ChatDBT will be OpenAI, and we are doing the alpha testing. So, I will skip some comments, because I have too much here. So what are the values, the main values of ethics? This comes from bioethics, the Belmont Report in 1979. This is autonomy, respect of human dignity, do good and don't do bad or don't do evil, the motto of some companies, and I guess the motto of all religions. And then to be just, give more opportunity to people that have less opportunities. Here in the right, I hope you can see, are some characteristics that we can have in AI and in software in general that are associated to autonomy, like, for example, transparency or harm benefit, for example, safety and security, and, for example, justice, for example, protecting the vulnerable. and sometimes this conflict this is a real ethical conflict two of these principles conflict, for example I need to put Ricardo restrained because he wants to make some harm it's just an example I'm sure it's not the case but we do that as a society this is when we have some ethical issue that we need to solve where in 2017 the ACM which is the largest professional association in the world so who is member of ACM here ok a few good so so they published this in 2017 about algorithm transparency and accountability nobody noticed that transparency was not in the list right I noticed right away we have a conceptual problem here So, in 2022, I was involved on the new set of principles, and one of the two main authors, and these are the new principles of the ACM for responsible algorithmic systems. And I will come back to this. Algorithmic systems, not AI. Algorithmic systems. The first one is my principle that I care a lot, legitimacy and competence. We show that things are legitimate for society, and we have all the competences to do it well. And I will get back to this. We put minimizing harm to make sure that you are minimizing harm, but because some people think only about humans, we added number nine. We are not talking only about humans, we're talking about the planet. We need to also limit the environmental impact. I'm sure you know, for example, that every time you use ChatGPT, you use 15 times more energy than, for example, using a search in a search engine. So you are really wasting resources, especially because the answer will be wrong. Especially if the topic is very specific, probably the answer will be an invention or a mistake. So please don't use language models as search engines because they are not, and I'm an expert on that. Okay, so this is the governance. So these principles have some order. This is the way that you need to use them. So the first one is when you have the idea and you do an ethical risk assessment or a human rights assessment or whatever, impact assessment. But you need to know what are the dangers and what are the benefits. and if you cannot show that the benefits are much more than the dangers, you shouldn't do it, even if you will lose some money. So that's the ethical thing to do. These are like vaccines. Now, if you go to the design and deployment, basically you need to use these five here, and then you do validation and testing. You know how to do that. You can use sandboxing. You can use monitoring tools when it's already deployed to make sure that the model is working as intended, and, for example, there is no data drift or anything that is changing. And when it fails, you need to have contestability and auditability, and there you need to do an audit. Basically, something wrong happened, and a third party needs to audit you. And here is something very important. Never audit an illegitimate system. So, any audit should start by seeing if the system is legitimate. Why? Because if you do an audit, and that has happened in the past, on an illegitimate system, you are legitimizing the system. And there are many systems that are not legitimate. And finally, when you do something really bad, you need to go to court, and I will show some examples of that. So let's start with the questions, the ten questions, and these questions are based on these principles. So the first one is just when you have the idea. So what you need to ask, I already said it, is your proposed system beneficial for society or you are only thinking in getting rich? If you earn half a billion, it's enough, plenty of money. You don't need to earn a billion, right? Even a hundred million for most people is enough. This could be a waste of resources. Anyone here read the Stochastic Sparrows paper? The paper that led to basically the layoff of Timnit Gebru and Margaret Mitchell in Google. So in this paper, they show, for example, that a very small transformer model is equivalent in training time of 57 years of the carbon footprint of an average person on Earth, let's let's say, a person in India. You also use a lot of electricity, between one and three million dollars. But what really matters, and these several papers last year, is the usage of more than one million people of these systems every day, like a search engine. Now, I don't know why they look at ChatGPT only, because they should look also at TikTok, Instagram, and all these things that are, in some sense, wasting more resources than, say, a language model. But this can be really bad. So anyone from the Netherlands here? So this is the city. City with a Y in the second letter. System for Risk Indication. Someone had the bright idea of looking for fraud in poor people. That's not ethical. You shouldn't start with poor people. You should start with rich people. This was the tax office in the Netherlands. So they accused mostly, falsely, 26,000 families of fraud. They had to return a lot of money. That was targeting immigrants, because it's true that some immigrants abuse of the system, but not 42%, only 3%. So it was already a problem. And then in 2018, the civil society sued the government, and at the end, a court in 2020 said that this was illegal and they had to be compensated. The government decided not to appeal. You know governments usually appeal and go all the way to the Supreme Court. In this case they decided this is already too bad, let's not do it. And in January 14, 2021, the chief of the opposition in Netherlands, I don't know if you remember that, for the Dutch people, resigned. The opposition of the government, not the government. What happened? He was the minister when this started in 2013, and he said, I didn't know all this, maybe the person didn't have the competence to do it, but some engineer did it, and he decided to resign. So January 14, the chief of the position resigns. The next day, the whole government resigns. So this is the highest political impact of a bad system. Maybe it uses a little bit of AI, it's not much AI. We don't know because we don't have access to the code. This is an example of things that you shouldn't do. Now we have thousands of cases, and I will show an example. The OECD has started a monitor tool where using AI, they are classifying news, and now they have almost 10,000 cases, and I will show a graph later. So what is the real ethical dilemma? Just to make sure you understand the idea. So how many of you know the trolley problem? Yeah, please delete it from your mind. You will never have this problem on your life. Never, okay? It's theory of philosophers. But this is a real problem we have today. People killed by cars. You don't need to convince me automated cars will kill less people because they don't do many things we do. So the problem is that the people that will be dying from automated cars is not a subset of this set. It's basically like this. So some people that were not in danger of being killed now will be killed. So I will use some gender bias on purpose because it's true. Most of the people we will save are irresponsible men playing with the Tesla autopilot. This is true. You can go and check. Women are not so stupid. So, what happens? We will basically kill some people that are vulnerable. And I will get back to this later. Maybe a kid, an old person. These people are vulnerable people. So the question is, where is the right balance between these two things? For example, in COVID vaccines, how many old people can die so the rest can survive? I don't have the answer. It's very hard. But let me tell you that not even the smart people have the answer. And I find this amazing example from the 18th century. So this is Voltaire. You know Voltaire, right? One of the philosophers of the French Revolution. He said that something that we all agree. better to have a criminal free than an innocent in prison right well a few years later we got what is called the Blackstone Rule this was an English judge that said it's 10 people not one ok I got the same answer from a doctor I asked the doctor how many people do you want to see that is not ill to basically do not miss an ill person. So basically, how many false positives do you want to see not to miss a false negative? He said 10. So, you know, doctors have a lot of ethics code. And finally, in 1785, Franklin, one of the fathers of the American independence, said that. 100 people. And also he said it in a way that you cannot, basically argue with him because this is like nobody has argued with him and win. So, okay. So, this is why ethics is hard. Okay, second question. And really, I think this is important. Does your solution rely in a proven scientific fact or at least a reasonable hypothesis? Sometimes we use any data and remember, my Murphy's law of data is the data that you have is exactly the data that you don't need. This happens. So, do you use really science? Let me give you an example. Physiognomy. So, physiognomy is using the shape of your face to, for example, predict your personality. This was done in China for criminality. Later, for example, it was done in Stanford for sexual orientation. Again, for criminality during the pandemic. and finally came back again the same person at Stanford detecting your political orientation in the US you can say it's easier because you have only two parties but I'm sure in other countries it's much more difficult and please you cannot do this so these are spurious correlations he got 70% but maybe if I use a bird I'm democratic in the US if I have a hearing also maybe I'm a democrat So these are all things that have nothing to do with your face. So this is phrenology. But this can be worse. This was a published paper in CVPR, one of the most important computer vision conferences. They claim, this is MIT, so a very important university, they claim that your face depends on your voice. I can believe that anything below my nose may depend on my voice, but the color of my eyes my ears my hair come on, this is not science so here I have my my master algorithm of phrenology phrenology was something invented in Germany so who knows what is phrenology ok, this was invented at the end of the 18th century by a German doctor that postulated that criminals had different convolutions in the brain very hard to prove You need to open brains. And also we have cognitive biases. For example, we invent categories that don't exist. We did it with race. The skin color is a continuous variable. And I will get back to this because the key question is here, why data from other people can be used to predict your behavior? So today, if you go to a bank, after the university most probably the bank will tell you you cannot get a credit. Why? Because you are too risky. Why? Because you are young. So this in average may be true but of course it's not true for all people. So you will favor some people and you will discriminate some people and this is happening every day. We are averaging people. There is no science that says that we can average people. So for all the people doing here behavioral predictions please think about this, especially if you don't have data about me. Three, this is very important. Are you authorized to solve the problem at hand? Probably this is what happened in the Netherlands, in the city. Basically, the engineer didn't ask anyone, maybe only the manager, if the system to detect fraud in poor people was a good idea. Of course, the minister didn't know. At least, that's what he said. I don't know if that's true. But this is already something that happened. At the end of 2019, two high schools in France decided to put cameras for security, basically surveillance. And you know this is a complicated topic now with the new regulation of the use of AI. But some parents went to court, and at the end the court said that this was illegal. And it's very interesting the reasons of why it was illegal. The first one is they didn't have the competence. The directors of the high school couldn't decide that. It had to be the mayor of the city. They never asked. Second, if you are using in a public space a camera, because of GDPR, you need informed consent. It's impossible to have informed consent. You need to sign a paper saying that you approve the use of a camera. In the U.S., you don't need to do that. you just say a sign and have you seen you are being recorded, smile this is like consent and finally this is very interesting the solution was not proportional to the problem so if you want to have security put a guard in the door, you don't need to do surveillance you don't need to become Singapore's airport for example and here is an example of what I said before probably this minister social affairs didn't know what was happening in the Netherlands. Question number four. Do you have the right experts for the problem you are trying to solve? Not only computing experts, the domain of the problem that you are trying to solve. Many times computer scientists are answering problems that are not computer science. And this is bad. For example, here is some work with two of my PhD students. One is a German. She's a professor now in north of Germany. And for example, do you see anything special here? Here you have the data size and the accuracy. So with 4,000 people, you get 81% accuracy in Spanish. And with 1,500 people, you get 90% accuracy in English. If you see this, you think that something is wrong. But if you're an expert on language, there's nothing wrong with it. Spanish is basically a language that you, now I forgot the word, is a language where you pronounce exactly what you read. In English it's not. So it's not that kind of language. Okay, let's go to the second set of questions, technical questions. Now this is the most important one for you. You are developers. and I will start with the same thing that the French court said. Is your solution proportional to the problem being solved? You don't need to use AI for everything, right? I like AI, but many times binary search is fine, right? So be careful. Here's an example from a paper that was published. This was done in Spain. This was used by the police in Spain. So basically, they were looking for fraud in reports. So basically, you say that someone has told you something, and it's false. This paper had a lot of problems because they didn't even use the text of the person reporting. They used the text of the police that did the report. So, of course, this has problems. And it's interesting because here you see the difference between cultures because in 2021, they tried to do the same in Germany. And the civil society immediately reacted, and this was forgotten. It was forgotten. So no light detector in Germany. So remember, minimal data collection, minimal time stored. These are good data protection rules. A group of questions. Number six, have you checked for bias in the data? In the objective function of the model, in the feedback loop between the systems and its users, are you at least not amplifying bias? Because many times you can amplify bias. So the system can detect the bias and put more. And this is the main problem today in AI. So basically all these type of biases from the data, from the optimization function, and from the feedback loop between users and the system. and here is what we are doing all the time we are receiving biased data all the time information is biased if you get a random noise nothing can happen so the question is should be neutral or fair with respect to negative bias so this is the first bias we have when we hear the word bias we think it's negative no it's neutral, in principle, it depends on the context. But what happens? Most of the time, we don't do these questions and we get the same bias. And worse, sometimes we get amplified bias. Now, if we get amplified bias, you cannot blame the data, right? Something else is happening. And this is what I recommend you to download this paper. It's my most downloaded paper now. The vicious cycle of bias in the web, I published in 2018. So bias is not only in data, it's in the interaction is in the models that we are doing. Even the recent paper, like two years ago, proved that the biases of the programmers get transferred to the code. So what experiment I would like to do is to do a recommender system only by women, and probably the recommender system would be more fair and empathic than current recommender systems. So that's an idea. well during the pandemic they did this and again this is a question about using data from other people they said ok we cannot have all these students in the same room so let's predict the scores and they found something that we know on average students of public schools get worse grades than students in private schools on average but there are good students and bad students everywhere So, of course, they couldn't use it because they were discriminating poor people. So these are brilliant ideas that some people have. So not always we need to answer this question, but if you may harm people, we need to answer this question. If it's for advertising in the web, okay, we have the brightest minds doing advertising in the web, but at least if you get offensive ad, nothing will happen. But if you harm people, you need to worry about it. And here you need to ask two questions. Is a secret algorithm ethical? Because of transparency. For example, especially if it's a public, it's a government algorithm. And then is a public algorithm safe? Because then if it's completely transparent, then you know how to game the system. And this is also a problem. So of course the solution will be in between and you need to find the right answer. It can be the optimization goal. So how many people from Italy here? Some Italians? Only one? Okay, so maybe you know this case in Bologna, the delivery case. Delivery is like Uber Eats, like basically food delivery platform. And there was a group of riders that felt that they were discriminated, but they couldn't find any common demographic characteristic. They were not all women, they were not all dark color, they were not all some type of hair. No hair, for example. But they sue the company, the company found the problem right away, and the solution was very easy. After you know it. There's a book called like that. Everything is obvious after you know it, yeah, right? So what happened? Some people couldn't deliver at night, women, people with kids, people with dependents, or people that really want to work at night. That's legal in Italy. So the solution was easy, give more work to the people that only can deliver during the day, and then you give more work at night to the people that only can deliver at night. Just change your optimization function. So very easy. I will skip this, this is about the feedback loop, this is what is happening in recommender system today, this is exposure bias. Other systems are generating the data of the system. Because you get three recommendations, you can only click on one of these three recommendations. So you get popularity bias, the rich get richer, and the poor get poorer. So exposure bias is the thing that you can see. And it's not the tip of the iceberg, it's the ice cube on the tip of the iceberg. Imagine that you have, you ask for some product, and there are 1,000 copies, sorry, 1,000 possibilities and you get only three. Really, you are seeing nothing. But the clicks will be only on those three. So these systems are basically generating the data that they want to see. And here we have a dilemma. What is better? A biased, fair algorithm, fair in the sense that AI does always the same if you are using deterministic AI, if you are not using pseudo-random generators, or a tired person? Why a tired person? because a tired person doesn't do the same every time. We are valuable. For example, judges do that. There's a famous paper a long time ago that shows that if you go to a court after lunch in the U.S., that's the worst time you can go to a court because if the judge didn't have lunch, uh-oh, you will get a big fine or maybe more time in jail. So this is the three cases, Exact, noisy, biased, unbiased, and noisy. Now, this question is a fallacy because basically algorithms are like C, are biased, but people is like D, are biased and noisy. I was using a very interesting paper by Daniel Kahneman and co-authors. Sadly, he died recently. That was published in Harvard Business Review. But two years ago, he did a whole book on noise, the problem of, basically, inconsistent decisions. So noise can be worse than bias. Question number seven. Are you tailoring the operation of your model to your problem? Are you using arbitrary thresholds? How many are you using? 0.5, right? 0.7, 0.8? Oh, you have a point that is 0.799, and it's not bad luck. I'm using categories that are hard to justify. I will get back to this. All errors have the same impact. Every time you use accuracy, you're assuming that all errors have the same impact. This is a problem. This is like going here to an elevator and the elevator says it works 99% of the time. I will not take the elevator, right? But if the elevator says it doesn't work 1% of the time and when it doesn't work, it stops, I will take it because I know I'm safe. This is the difference between alchemy and engineering. We're doing too much alchemy in AI. We're doing too many black boxes without understanding what is happening. Have I used a threshold in my talk? Let's see if you understood the class. Have I used a threshold, arbitrary threshold in my talk? Very good, 10 questions. Who said that there are 10? Only because of the fingers, the same mistake of the first person that they just are changing. Let's say, let's show a page with 10 answers, independently of the size of the screen. Crazy. So, yes. And I will get back to this. So, here I want to remind you the limitations of data and machine learning. Humans are very good at filtering and abstracting data. We need to see a cat once, and we know a cat forever. Maybe you can do zero-shot, but it's not trivial. Second, you cannot learn what is not in the data. Also, data does not capture everything. So don't do thatification. In many cases, data will not capture qualitative things, contextual things that are part of the problem. Data is a proxy of reality. And sometimes it's a very bad proxy. So this is a question you need to check. Is my data really a good proxy for what is happening? And many times you are optimizing not what you want to optimize, you are optimizing a proxy. Why? Because you cannot measure what you want to optimize. And there's a very interesting paper that was published early this year. I recommend it. It's called Predictive Optimization. And this shows seven problems that we have in machine learning for eight different systems. And basically all the systems have these seven problems. and some of them are can you measure what you want to optimize? This is what happened in 2018 in Arizona when a woman crossed in a bicycle at night a road and was killed by an Uber automated car. There was a backup driver another woman inside but she was watching a video. I will get back to this. So, this is a kind of error that I call a non-human error. These are errors that people don't do. All of you, if you're not drunk, will recognize a woman crossing in a bicycle at night, and the car was not going very fast, because they cannot go very fast. So sometimes AI does errors that are really dangerous, because we are not accustomed to these kinds of errors. So accuracy is not relevant, as I said. The impact of error is, maybe I want to go in a Buddhist car that goes very slow, doesn't kill even a cow, but maybe if I take a faster car, I may kill someone, is our choice. So, how many of you have done a classifier where the answer is, I don't know? One, great. Do that, all of you need to do that. If the threshold is not 0.8 or above confidence, you have to say, I don't know. That's what smart people do. The best teachers say, I don't know, in front of the class, when they don't know. Okay? So you have to do the same. Today we have a system with no confidence that says whatever. Let's go to the last set of questions. These are the responsibility questions. Number eight. Is your solution secure, safe, private, and it does comply with all legal regulations. This is something that your lawyers will force on you, so it should be easy to fulfill. But this is about regulating the use of AI. Very important, I put there, the use of AI. We are not regulating the technology, we are regulating the use of the technology. And they are taking too many years. I don't know if you know, it was voted second time in the parliament in March, and they found out that they voted the wrong version of their final agreement. So they need to fix that. I don't know when this will become public. Of course, what happened between April 21 and today was generative AI, and now it includes generative AI. But there are problems. The first one is that we may have an AI loophole. This is not for AI. This should be for any software. You don't need to use AI to harm people. and the best example finally appeared this year. And it's an old example, has been going on for 20 years, but people really didn't know in the United Kingdom about it until a private TV channel did a series about Mr. Bates against the post office. You can see it in Amazon Prime now. And I'm sure you know this actor. He has been in many movies. And what happened? Well, an accounting system, just an accounting system, accused 900 postmasters in very small towns, but basically they were representing the post office because the town was too small to have a post office. So it was a store, and the person had a machine and had to manage the machine, and basically accused 900 people, people that all the community knew, because these were small towns, of fraud, that they had stolen money from the post office. Four of them committed suicide. Many went to prison because they couldn't pay the money. But Mr. Bates went to court, and basically, in 21, he won the case with other 39, 38 people that went with him. They were innocent. The accounting system had a problem. Amazingly, right? An accounting system that doesn't know how to count money. It was a memory problem. It was a corrupt memory problem. So no one expected that. So what are the issues of regulating the use of AI? So the example says here, this should be for any software, not only for AI. Otherwise, what will happen? I will say, oh, I use very sophisticated statistics. I don't use AI. And I have a very nice model. What are my problems with the current regulation, proposed regulation should be technology independent. This is an experiment. We are the guinea pigs of a very complicated experiment. This is like saying you shouldn't hit a person with a hammer. No, you shouldn't hit a person with anything. Doesn't matter what you use. It should be technology independent. So we should regulate sectors. And this is what happens for example in medicine. And we should not create fictitious categories. So today we have four fictitious categories. Forbidden, that's okay. Not fictitious, we are forbidden that. High risk. Low risk. Why not medium risk? I don't know. And then no risk is implicit. But these categories don't exist. Risk is very hard to measure. If you do a self-evaluation, you basically will fool yourself and you will say, oh, my system is no risk. Because you have a conflict of interest with yourself. So this is a problem, and we will see what happens. But I believe it's better to have some regulation than not, because people are doing unethical things all the time. Question number nine. Are your users fully aware of the impact of your system on their lives? Do you need interpretability or explanations in your system? Well, GDPR already has this. So GDPR is already in place. and Article 22 is about automated decision-making. And don't read this. This is what is in the Article 22, but this is the important part. I have the right to contest the decision. I need the right to talk to a person to contest the decision. Today, most digital transformation is making things harder for people to talk to people. Why we cannot talk to people? We need to talk to people when we have a problem. So what it means for us, it means that if you want information of the processing, you need to have interpretability. If you want to challenge a decision, you need to have explanations. Now very careful, in medicine, explanations can be complicated. I don't know how many people watched House in the past, this doctor that will try to find what is the real cause of an illness. The same happens if you explain something that's not true in medicine. If you explain the wrong illness, maybe a person will die anyway. And finally, if you want to make sure that the system is working as intended, you need to do continuous validation, testing, and maintenance. And the last question. Do you have in place systems to allow contestability and auditability? Do you log everything that the system does, from the design to the deployment? same for accountability because human incompetence is everywhere right so in 1976 George Box a famous decision said all morals are wrong but some are useful this is still true for AI all morals are wrong because data is a proxy of reality but some are useful but I guess it's a minority our system can be really stupid Let me give you some example. In end of 2020, I think it was end of 2020 or end of 2019. Yeah, I don't remember. This was in January of, I think it was January 2020 or 2021. Elon Musk, you know the guy, right? Too bad. Elon Musk said, use signal. so what he's saying if he said use signal what it means use the chat app signal that's the most private one right well this system that uses data from influencers in the stock market so that he was saying buy stock from signal advance inc and this company went 400% up so the company was very happy all the people following the trend were very unhappy and if you have seen the recent movies on fighting the stock market you know where I've been or for example this one, it looks funny but it's not really funny in 21 a Facebook engineer decided to use an English train model in France and decided the town of Biche was forbidden I know the French people maybe know this town and they had to wait three weeks to fix this because there was no human in the loop right very hard to talk to human the beach is okay in France maybe not okay in the US so this is funny but for example imagine that this town is using the page in Facebook for COVID announcements it's not funny any longer or for example this case last year and this has happened like three, four times now and there are many courts that now have rules on how to use generative AI in US courts is a lawyer that decided to prepare the case with Chad GPT Chad GPT invented two cases because he will never say I don't know and of course the judge did his work find out that the case didn't exist and this lawyer almost lost his license I don't know if it's true or not but said that the judge asked the lawyer, but did you check this? Well, I asked Chargé-Petit if they were correct, and Chargé-Petit said yes. Stupidity goes a long way. So this has happened many times, and it's getting problems. And let me give you an example of accountability, a very sad example, because why Uber was not guilty of the case of the woman in Arizona. Well, Uber reached a settlement with the family the same week of the accident. They paid millions of dollars. Nobody knows how much because it was a private settlement. And basically, the family didn't sue Uber. By the way, Uber sold the self-automated car unit after this because they realized how difficult it is. Cruise is not working now, and only maybe Waymo and maybe Tesla are the ones left trying to solve the problem. So what happened? Well, only two years ago, finally, the woman that was inside the car decided to give an interview in Wired. I recommend it. I think it was March 22. And what happened with this woman that was inside? She was doing a very boring job. She was a transgender immigrant from Mexico. She was basically earning the minimum salary and was the only work she could get because she was basically discriminated because of being a transgender. So she was accountable for this accident and now she is in her house with these rings in the ankle without being able to go out. It took like two years, the case, but at the end she took responsibility and Uber is not responsible for this. So both have some responsibility, but Uber basically didn't. But now, genetic AI adds more problems, more misinformation, more bias and hate, implicit unknown censorship, all the things that you don't see in the answers that are deleted because someone thinks that you shouldn't learn about it. We don't know. but I'm sure I know like two or three topics like GANs that you shouldn't know. Decreased diversity. Chagibiti, for example, and Lama and all these things support less than 100 languages. There are 7,000 languages alive. 15% of the people on Earth talk one of these languages that are not supported. So please don't say we are democratizing AI. We are not democratizing AI. We are increasing the digital gap between the people without internet and without one popular language as a native language. So we are increasing the digital gap in an increasingly amazing speed. So this is what we are doing today. And we are doing cultural colonization. So how many continents there are in Germany? How many continents? A German guy or a woman. How many continents? Seven. Seven? You're American. Or you already believe that America is two continents. Okay, if you see the Olympic flag, the Olympic flag has five circles. These are all the continents that can participate in the Olympics. One is missing, Antarctic, because there is no country there. So the tradition in Europe is six continents. But if you ask Charter BTG, you will get seven. This is colonization. And we are getting a lot of colonization in any topic that basically is using data mainly from the U.S. And we are losing skills. So how many people can find a place without the GPS? We took hundreds of thousands of years to learn this, and we are losing it. I saw a meme that said, in five years, you can put in your CV that you can write by hand. You know how to write by hand. But the problem is not writing by hand. The problem is thinking. Because to write, you need to think. You need to put what you are thinking in words that another person will understand. So this is not evolution, this is involution. So please be careful, try to do things by yourself. Mediocrity is going high. And then we get this. You saw this last year, or this. And really today we cannot believe anything we see. So everything we gain by having people sending a video from the street is gone Because that video can be invented And the best example already happened in Hong Kong in February In February, there was an online meeting between an employee in Hong Kong The chief financial officer and other people At the end of the meeting, the person in Hong Kong had to transfer $25 million to an account of another company He did, because that's what he was ordered to do by the chief financial officer. What he never realized is that all the people in the meeting were faked. So they faked the videos, the voice, everything to do this. So be careful, be well. And then we get to the last problem. Insanity issues. Maybe you saw this about Jessica, or in China that one person basically recreated his grandmother, and in March Jaron Lanier, one of the fathers of virtual reality, said the danger isn't that AI destroys us so please don't believe in these existential threats. So how many people believe in AGI here? Okay, good. Or they are afraid to say it. Okay. But it will drive us insane. I said, I thought when I read this, yes, This is exactly the problem if you saw the movie Hair or if you saw this news about this Google engineer that thought that the system was sentient and had to do a new religion. He was a bit crazy. But sadly, and I know there's one person from Belgium here, sadly, less than one week, he was proven right. So one person in Belgium, a PhD student with wife and two kids, after six, sorry, this is wrong, six weeks talking to a chatbot, not chat DVD. Chad Yee, basically decided to commit suicide. And I will let you read the last conversation because it looks like coming from a science fiction movie. Can you read it? So it's implicit that they will meet in the afterlife. And that's why the person killed himself. So, do not humanize technology. Keep it human. Be lucid. Do things. What's wrong with this picture? It's imitating that it's thinking. But the robot, of course, is not thinking. They don't think. They don't even understand what is a cat. They can recognize an image of a cat, but they still don't know what is a cat. So, LLMs do not hallucinate. They make errors. they don't think, they mimic conversation they don't have opinions they reflect those they don't have intentions, they have outcomes and they don't intend to manipulate and harm but they do cause manipulation, disinformation and harm, so this is you need to remember and to finish, I want to show the graph of the OECD this is growing exponentially this is basically like more than one problem per day in the last three months So this is the problem. And these are the ones that get to the news. So there's a selection bias here. So there are many, many more. And this is where we're working on the institutes for experiential AI. All people read experimental, but it's experiential. It's because we want to make it practical. And we have an AI ethics board, a worldwide AI ethics board, if you need help. And then to finish, we need to empower people instead of replacing people. If your system is empowering people, increasing productivity, that's the best system you can do. We need to work together with AI, not AI replacing us. Of course, AI can replace in everything that we don't want to do. This is the perfect world in the future. AI is doing what we don't want to do, but we are doing the things we want to do. There was a very good tweet from a Polish woman. It says, I don't want AI to write for me so I get more time to do my laundry. no I want AI to do the laundry so I get more time for writing this is the right way to think about it and it was a very good tweet so you can read my concerns here

Speaker 2 [51:31]

Is the mic on? Okay, now it is. Thanks, Ricardo, for the presentation. We're going to be taking questions from the audience here, so you can raise your hand, but also going to read some questions that are coming on Slido. First question from the audience, anonymous. It seems a lot of bias fixes could just be introducing a new bias. What's a pragmatic way of removing bias that's politically viable?

Speaker 1 [52:00]

So there's a problem in the question because the person says how we can remove bias. You cannot remove bias. You can mitigate bias. In many cases, we don't know what is the right reference value for bias. And in many cases, the bias will be unknown until you have a problem. For example, you have intersectional bias. Maybe your system is discriminating women from some region of Turkey. And you will not find that until a person from that region gets affected and complains. And if the person doesn't decide to complain, you will never know. Bias is a hidden problem. So you can also always check for standard bias, like mitigate gender bias, ethnic bias, geographical bias, but you cannot remove it. But you need to check for the basic biases, at least not to have a big problem, but then you may have a problem in specific cases because of intersectional bias in the demographic case. And if you saw the delivery case or the case of e-commerce, there's always other kinds of bias that you need to make sure. In e-commerce, so how many people is doing e-commerce here? I'm sure there are many working on websites, yeah. So, basically, the long tail is discriminated today. The long tail of offers is discriminated because people don't do enough exploration. they do mostly exploitation because exploration is losing money but it's the only way to know your real world so you need to do more exploration in spite of losing money

Speaker 2 [53:37]

The next question shouldn't loss of us of certain skills be expected in an evolving society Many traditional important skills have been lost nowadays due to tech advancements

Speaker 1 [53:54]

I should apply what I say. I shouldn't answer questions that are not for computer scientists. These are questions for sociology or something like that. My personal opinion, yes, we should learn some things that really today can be done by technology, but we should learn those things because they are human, like writing. Because writing is about putting your thinking in words. The same with finding a place without a GPS. I do that all the time. Because I don't want to depend on GPS. You don't have internet and you're gone. Right? So try to keep exercising your brain.

Speaker 2 [54:38]

You made now several times references to women. What could be done to bring more women into higher positions in tech to balance out bias? And would this solve the problem?

Speaker 1 [54:53]

Another question for a sociologist. I think all of us have the responsibility of doing something. If all of us do something, we can change. So, for example, my contribution to this problem is that I'm maybe the only one in computer science that I have graduated 32 PhD students. Half of them are women. Most of the time it's 20%. so we can do this thing we can do this small affirmative actions to change things but you need to do it and many men don't do it so this is part of the problem and I have one of my female PhDs here Joanna Edwards from Amazon here in Berlin any questions from the audience? There's one question here No, no, there was a question behind you Okay, thanks so much for the talk I spoke recently about the security and the privacy in LLMs and I just want to have some more insight from you on the data and also on the prompt because to reduce the output of LLMs for instance aids and things like that then these companies need to have what they call system prompt so should system prompt be managed by the companies or by or will manage the system prompt, because the system prompt is very important in LLMs. I'm very pessimistic here. Let me do a metaphor. Testing LLMs is like going to the Atlantic Ocean, taking a sample and saying, oh, there's no bacteria here. We are talking about exponentially many ways to ask things. It's impossible to test that. So basically we're testing, all the tests we're doing are completely anecdotal. You change the test, you will get different results. Many companies do a test that favors them, not other LLMs. So this is like testing almost infinitely open domain. So I don't think we can do something very interesting on that. So with the traditional ways, I don't think we will do something interesting. We need to do it in a completely different way. I believe that LLMs need to have knowledge. So they need to have real knowledge. So when they say something is true and they know it's true. Second, they need to reason. They need to do planning like Jan LeCun says. And the last one is they need to understand the context. Because if you don't understand the context, the answer will be wrong. So in many cases, the context gives you the answer. So if we don't put these three things, so basically common sense, knowledge and planning, we are not going through something more intelligent. Can I ask, do you think AI will increase inequality and is that an inevitability of AI?

Speaker 2 [58:21]

flexibility of AI.

Speaker 1 [58:22]

Well, yeah, I think, yes. LLMs are increasing inequality between all the people without Internet and without being able to use one of these languages supported by LLMs. So we are increasing the digital gap in a very fast way. Yes, I believe that. That's why I don't like when people say we are democratizing AI. It's like Internet. We say democratizing information, but it's still 35% of the people on Earth don't have Internet. But maybe it's okay. They have one thing that we don't have. They have privacy. I wish I had privacy. We live in a situation where we always have internet and the internet forgets never. So that means that the data which is now biased and it's also wrong maybe, stays forever and gets then always into new models. So maybe we need some situation where this need we have to clean up or something or how could be this solved that they always have old informations and yeah this will be the bias will increase I think it might it might yes it might so I think I think part of the problem is that we are trying to use LLMs for everything and we shouldn't for example we shouldn't use LLMs to generate information We just use LLMs to, for example, write a simple letter. I have a list of things that I believe are good for people to use, but for example, not to replace a search engine or not to generate something important for the company. We need to be very careful because they will invent everything. So a smile.

Speaker 2 [60:15]

We have time just for one more question, but if you want to ask your questions that I haven't had the chance to read, please reach out to Ricardo during the break.

Speaker 1 [60:26]

break. That will be here, yes. But the last question

Speaker 2 [60:27]

But the last question.

Speaker 1 [60:29]

hi Ricardo I'm Danny I want to challenge you if I may and don't you think that AI can also increase inclusivity by example asking things to chat GPT can be very very more easy for by example people on the autism spectrum or writing some essay for people on dyslexia so don't you think it can also increase and then And especially inclusivity of people with different kinds of disabilities. And why not give them the advantage? Yes, I believe technology, if used well, can help all these people. But after that last year, I met a person that was blind and deaf. I don't believe that they are disabled people. I think if they have the opportunities, she said to me, I'm not disabled, I'm just different. so I think we need to recognize that there were just different people and she's a bright lawyer in Harvard from Ethiopia but she can speak perfectly and she's deaf and blind so you can do these things first you need to have opportunity so what we need is more affirmative action so people can have these opportunities for example can go to a good university and study in most cases that's not

Speaker 2 [62:05]

Thanks a lot, Ricardo, once again for joining us. Thanks all for joining the session and enjoy the rest of the conference.