ACM ByteCast - Ricardo Baeza-Yates - Episode 88

Episode Date: July 30, 2026

In this episode of ACM ByteCast, host Juan Miguel de Joya welcomes 2025 ACM Luiz André Barroso Award recipient Ricardo Baeza-Yates, the Search Chief Scientist at You.com, holding part-time professor appointments at KTH Royal Institute of Technology (Sweden), Universitat Pompeu Fabra (Spain), and Universidad de Chile. The award recognizes his pioneering contributions to algorithms and information retrieval as well as his leadership in fostering a vibrant transnational research community across Latin America. Baeza-Yates is widely regarded as one of the world’s foremost researchers in information retrieval, celebrated especially for pioneering innovative data structures that have shaped the field. His work has produced influential algorithms for string searching and fuzzy matching, including the well-known Shift-Or algorithm. As a practitioner, Baeza-Yates served as VP of Research for Yahoo Labs, secured 14 patents, and co-founded several startups in Chile and Spain, including Theodora AI, devoted to mitigating technological bias. Among his honors, he received the CLEI Distinction for Contributions to Computing in Latin America in 2009, the Spanish “Ángela Ruiz Robles” Award for research excellence and entrepreneurship in applied computing in 2018, the 2024 Chilean National Prize for Applied Sciences and Technology, and the first Merit Award from the Chilean Computing Science Society in 2025. Ricardo is a member of Academia Europaea, and a Fellow of ACM and IEEE. He is the co-author of Modern Information Retrieval, which became the field's most cited textbook. Ricardo shares his unconventional path into computing, from influential teachers to his discovery of the mathematical and logical beauty of algorithms. He discusses his current research, which focuses on evaluating AI by examining failures, harm, and risk rather than simply measuring success, since errors can have profound consequences in fields like medicine and law, where users may not have the expertise to recognize errors. He advocates for combining reliable search with AI to provide AI agents with accurate, trustworthy information. The conversation also explores the limitations and social consequences of increasingly replacing traditional search with AI-generated responses, including the risk of "cognitive offloading." Ricardo also shares his views on how AI is shifting computer science, offers advice for future AI developers, highlighting ACM’s principles for responsible computing, and advocates for more inclusion of Latin American perspectives in computing and AI.

This is ACM Bycast, a podcast series from the Association for Computing Machinery, the world's largest education in scientific computing society. We talk to researchers, practitioners, and innovators who are at the intersection of computing research and practice. They share their experiences, the lessons they've learned, and their own visions for the future of computing. I am your host, Juan Miguel de Hoia. At its core, search is about helping people find. what they need in a world overflowing with information. So it sounds simple, but it's actually a deeply difficult problem.

How do you understand what someone is looking for? Sift through enormous amounts of information and decide what is most useful, relevant, and trustworthy in that moment. And once you think about it that way, you realize that search is pretty much everywhere. It's in the questions we type into a search engine, the videos and articles recommended to us, the products we discover, the maps that guide us, increasingly, the AI tools we turn to for answers. Behind all of these everyday moments is a

deep field of research called information retrieval, the science of helping people find the right information at the right time, from an overwhelming amount of data. Few people have helped shape that field as deeply as Ricardo Beezza-Yates. Ricardo is a Chilean computer scientist whose work in information retrieval, algorithms, web search, and data mining has influenced how modern search and recommendation systems are built and understood. He is currently the search chief scientist at U.com and holds part-time professor's appointments at KTH Royal Institute of Technology in Sweden, Universitat Pompeio Fabra in Spain, and the Universidad de Chile. His career has moved across academia, industry, and entrepreneurship. He previously served as vice president of

research at Yahoo Labs, holds 14 patents, and has co-founded several startups in Chile and Spain, including Theodora AI, which focuses on mitigating technological bias. He has earned his engineering degree and master's degree in computer science and electrical engineering from the Universidad de Chile and his PhD in computer science for the University of Waterloo. But Ricardo's impact goes beyond technical foundations. He has spent his career mentoring researchers, building institutions, and helping expand the visibility of the Latin American computing community around the world. And that combination of pioneering research and commitment to lifting others up is what has led to his recognition with the ACM-Louis-Andreborosso Award,

which honors fundamental contributions to computing from researchers in historically underrepresented communities. So it's my honor and real pleasure, actually, to spend time with you again, Ricardo. It's been quite a while. Thank you. So, so we've seen you. Yeah, when was the last time we saw each other? Was it ITU? In Geneva, in the AI event that they do every year. Oh, yeah, that's true.

Definitely. Yes. Yeah. Will you be going this year? No, no. I only went that year when I met you. Oh, okay. Well, yeah, that's a lucky year.

Yes. Yeah. So before we continue, I kind of wanted to ask you, I know we talked a bit about your background, but would you mind introducing yourself and talking about what you're currently doing? Yes, so no problem. So currently I'm sharing my time between industry and academia, mainly industry in U.D.com, as you mentioned, in San Francisco. But also I do have this part-time commitments in several places and physically I have to be a few months in Stockholm because of my KTH commitment that was even before I started working with UDCOM.

So I'm right now in Stockholm doing research and responsibly I. So mainly with my PhD students, mainly in Barcelona. and my work here in KTH. Touching different problems on Responsibly AI, maybe the main one is evaluation and looking at how to evaluate harm and not success. Because things should work, right, by default. I mean, why we evaluate when things work? We need to think what happens when things doesn't work

and how much harm that creates because all errors are not equal. So there are errors and they are more harmful than others. And we are looking at that. So different ways to minimize critical errors and also to try to measure. And because of regulation of the use of AI needs that. For example, the EU regulation is based on harm and risk, but really we don't know how to measure risk to well yet. I totally agree, and I think that's a very fair point.

I think one of the challenges we're living in now currently in this space is that both the technology for discoverability is so pervasive. Like search is just, we take it for granted that search did not exist like decades ago. And so the work that you were doing in that space is really vital, but at the same time, that space is also evolving with, in conjunction with a lot of different technologies like machine learning and how that impacts our ability to disseminate and understand information. Before we go in further, I think one thing I really wanted to ask is that sometimes when we have audiences listening to this podcast, they may not necessarily completely understand all the

technology. And so I was very curious if you could explain what makes both either search or responsible AI difficult as a topic to kind of tackle? In some sense, I'm combining the work of those two fields because at U.com mainly does search APIs for AI agents. And what you want, if you want to succeed with AI agents, or what is called the agentic AI, is to have relevant and true information. And this is not easy if you use a language model that will predict everything. and in these predictions, some mistakes can occur.

I don't like this word hallucination because really they're not hallucinating. They're more like making mistakes. Sometimes mistakes are creative, and this could be interesting mistake depending on the context. But if you make a mistake in a legal case or in a medical case or even in other settings related to people, this could be really harmful. And then you need to combine responsible AI, for example, to find the right information. And that's why we're building the best possible search APIs for agents to make sure that they have the correct information to start with. Of course, on the way down, mistakes can happen, but at least starting with true information.

And today, it's very hard to check anything that the chatbot will give to you because you are not an expert. So if it's well written and looks like correct, why is not correct? And very few people will find the mistakes. and this happening all over again in many contexts, in many news, and many people, of course, don't mention it, but I know cases every day about this problem. I can totally see that happening, and we're seeing some of that also appear in the news, right?

Like there's very different cases, just going back even years from now, since even when we first met, where you're seeing how we're beginning to really grapple with the impact of the technology. Is there something that you think about search, or AI that we as engineers would know about, but that the public usually misunderstands?

I think even some computer scientists don't understand very well the limitations of data, material learning, and even us. Because I think the most important problem is not the technologies, us, how we use it. So for example, if we don't understand that really the system doesn't understand anything and it's only a very good prediction,

we run into trouble. So the first problem I think that's true for everyone, is that we humanize this technology. We talk about thinking, reasoning, seeing, writing, reading. And for all these things to work, you need to have real understanding. So first, we should use the right language. So we should use, okay, this system process data,

generate data, but that's it. Whenever you have a system that looks at cat image and transform the cat image to the word cat, it's only changing the representation. It's not understanding, is like we do. So I see in DGIS, even with some computer scientists, it's not well seen because they use words that they shouldn't use

to explain how this system works. So we should be careful on that. Ways and Bound, that 30 years ago, the person that designed the first chat about Eliza said, we should never confuse computers with humans. But today, we're doing it every day. Right. And it is kind of a bird line nowadays, right?

Even with search, for example, let's look a search. Search as a field has gone from keyword matching and ranked links to semantic search, neural systems, rags or retrieval augmented generation, and AI assistance that synthesized answers. And do you consider that are we still doing some sort of information retrieval when we're giving synthesized answers by AI? Are those valid information retrievals, do you think? For me, as an information retrieval researcher, I would say no. I know that many people is moving away of the standard search changes that are using AI summaries, because AI samarists many times are wrong.

And some people look only at that. Well, some people even use chatbots too instead of search. And there are many consequences there. One is that they are not really searching. They're predicting information. That's one problem. But also, for example, they are using 10 times more energy than searching. So they're wasting energy.

So that's another, I think that's another concern I have, the use of resources that are are important for us. And there we need to be fair. It's not only chatbox. We're using a lot of energy with TikTok, Instagram, and many other apps that are helping us to waste time, but also they're wasting energy. So I think many people is not conscious of that, and someone is paying that. Probably you are paying part of it, but I think we are not paying part of all our waste of time that we do every day. So I think we need to be very careful on this balance, on how to use these tools for the right thing. For example, I believe this can be good tools to translate information,

to help you in other languages, to summarize information, to start information. But if we use it for just writing, basically we have all this problem of cognitive delegation, cognitive offloading, basically, if we do it, it's not too much trouble because we have learned how to do it. But for young people that they haven't learned yet to talk to, to do it. I wonder what will happen in 10 more years when people cannot write anything. Not because they don't know how to write, it's because they don't know how to think, which is this is the important part. They will be like, I would call it cognitive zombies.

It's not evolution, it's involution. And I wrote a paper about that, about how we are not evolving in search because we are using too much AI for their own things. Now, if we combine AI the right way we search, we can get very interesting solutions. Like, for example, if you have relevant information and you want to summarize it, then a language model can create the best answer possible, but you know that it's all true because it started from true information. But when you start predicting information, then you run into troubles. And this is something that it's very hard to fix because it's part of the architecture. The architecture was designed to predict the most probable text, was not designed to look at

knowledge database and give you the real fact. So one of the things that's really interesting now, I think, is retrieval augmented generation or rags, right? So for folks who do not know what a rag is, to put it simply, like, high level, it means that an AI looks up the relevant information first and then uses that information to give a more accurate answer instead of relying what it already quote unquote knows based off of the model. And to me personally, it does sort of connect modern AI systems back to information retrieval. Would that kind of fit what you're envisioning, or would you see it in more? different direction. Yes, but you know that even using RAC, this system sometimes make up answers. So this is not enough. This improves. So mistakes are less problem, but because of this prediction approach, this mistake can even happen. So there's some weird results like show

that if you put some random results on the Rack, the result may improve. So this is like counterintuitive. You put things that are not interesting, then maybe the signal to noise ratio improves, and then you get a better answer. So we still don't understand how systems that have not one million, but billions of parameters really work. We can build them. We know what is the architecture, but really we cannot imagine in our minds how these things interact when you're using one billion formulas to compute something, right? It's very hard.

Absolutely. And I think as search becomes more semantic and more personalized and AI-generated, those stakes are going to just get higher. To your point, like these systems don't just find information. They can shape what people see and trust and believe. And that makes a lot of the work that you're doing in biases and search and recommend your systems and responsible AI, especially important now. One problem that really worries me is not only the bias,

is the cultural communication behind these systems. So my standard example, if I ask you, how many, continents there are. Yeah. Are you asking me? Are you asking me? Are you countries? I'm pretty sure there's seven, but is there a new one that I didn't know about? No, but you study in the U.S., right? Right. That's why for you is seven. Yeah.

But in the European tradition, it's only six because America is only one. Oh, okay. The American culture is already colonizing the rest of the world because that's why, for example, the Olympic flag has only five circles. And Antarctica is not there because it doesn't compete in the Olympics. But for the US and Canada and many other countries now in Latin America, for example, it's not seven, because North and South America is two different continents. So that's a simple example of cultural colonization, but they are more complicated examples

where you don't see all the possible points of view. If we want to give true answers, we need to give more than all the possible points of view that are valid. Some topics are not about truths or not. They are about beliefs, right? And for example, if people believe we should control guns or not, you should have both views and so on. And that's happened with many, many problems. But you will not get all possible points of views because you are trained with bias, data sets. And also you have guardrails against maybe some thoughts or some beliefs that you think are wrong,

but maybe another culture don't think the same. We are getting this colonization that the next digital issue. Not only that, I think it's together with this, is that the best language model supports around 200 languages, but we have more than 7,000 languages alive. They did some analysis, and 10% of the people in the world don't speak one of these supported languages. Now, if you add that, all the people that don't have internet, all the young people that cannot use a computer or a device, all the old people that will never use it, and if you

do the right overlaps, you get the digital. between 50 and 55% of the people that cannot use these technologies. So when people say we are democratizing AI, I really hate it because we are not democratizing AI, we are increasing the digital gap, and much faster than before, between people that cannot use this technology and people that can use it. Now, maybe the winners, and I will try to be like provocative here, maybe the winners will be those people because those people will keep thinking. they also have more privacy.

So maybe they are the poor people today, but if these things don't go well, they're the future of the world. In a very good perspective, actually, I think people won't see our faces for this, I imagine, but we're both from the Global South. And I think we have had these conversations on kind of the discrepancies

between what technology means for people in developed countries versus developing countries. So what I'm going to try and do, actually, in this conversation, is let's take a step back from the technology aspects for a bit. You are from the global south, but you are from the North Hemisphere. I was raised in the North Hemisphere.

That's true. But which country is your global south? I was born in the Philippines. That's what I'm saying, but this is in the North Hemisphere. Is it the North Hemisphere? Yes. Obviously, this is a good example of where the technology bias comes in. Yeah, but this is, I say, because this is just a convention to talk about the global

south, because most of the developing countries are in the south. but even includes Mongolia, which is north of Philippines. So it doesn't make much sense for me because I love geography. So remember, the equator goes through Indonesia, which is south of Philippines. You know, they always tell, okay, so that's definitely a bias for sure because they told me, oh, you know, we're so close to the equator, but you're absolutely right. Let's amend that, and you're absolutely correct. I think the language, to your point as well, like, just from the language we've used,

Global South is kind of a misnomer to some degrees when we talk. about developing countries. But I really want to step back and talk about your origins. It informs kind of a lot of your opinions. And to understand why those opinions or exist, like I understand correctly, you started out in Chile and then moved in computer science. Chile is not necessarily at that point in time, if I understand correctly, not one of the traditional centers for computing. So I was just kind of curious, what incentivized you to get into computer science, actually? That's an interesting question. I think I never thought about it. Really? Because I, I, I, I never planned things, so I take opportunities.

For example, for me, I think there is another reason that maybe was the main driver of my career is that I learned to read and write very early by my grandfather, my maternal grandfather. So that opened my mind to the world. And then I had a great, amazing British teacher in primary school that had a course called General Knowledge, because she was not a teacher, but she wanted to teach. So she teach general knowledge, and this opens even more on the world. General knowledge includes many things, so this was like a very interesting class for me. Like, you will not know what you will learn every week, and this is nice.

And then I guess when you have very good teachers of, I think the teachers help a lot on the side what you will study. So I think a person can do many things, but depending on the best teachers, they will drive you to that path. So I had very good teachers in mass and physics in high school. So I decided to be an astronomer or an engineer. So I started engineering. I had never seen a computer computer when I entered university. And then I had to do the first course on programming. I discovered that I love the logic behind algorithms.

And really the combination of mass and logic captured me. I was studying electrical engineering. I did finish that. But really, I realized that what I really liked was computer science, and then I had an amazing professor of algorithms. Maybe that's the reason I started computer science, so I did the master with him, Patricia Poblette. I think he was one of my best teachers I had because he did his PhD at Waterloo and other

people in Chile did it for historical reasons. I went there, I got a scholarship. It was not easy at that time to get scholarship, especially from Chile. I found also a great supervisor, Gaston Gett, that even make me love more algorithms. And they were working on the Oakford English dictionary project. So basically at that time, that was one of the largest files was 570 megabytes in 1986. So that was a very large file. It didn't fit in one single cedar rome.

At that time, it fit in four cedurums. Later, the ced roms improved and then you could fit it in one. But the problem was how to search that, how to search that, and I started with sequential search algorithms in my PhD thesis. And then naturally, when the web arrived, I finished my PhD in 89. So when the web came in 91, 92, 93, although the idea is from 1999, naturally I said, okay, I can apply all these search skills to the web, and then I passed to do the information retrieval, which is basically searching algorithms

using an index, like inverted index, that's a classical approach. and then I started to work on that with a Brazilian friend which was a friend of Andre Luis Barroso because he worked at Google too. In the Riberio Netto, we decided one day to do something crazy. Why we don't do a good book in Infomeratorial

because it seems that there's no modern book and they said yes and that became in 1999 this modern information retrieval the most cited book on the topic. So that's not a standard thing that 2000 Americans write the most cited textbook in a topic.

like this is not true, I think, in any other topic. So sometimes for crazy ideas, you get interesting results. So I think this was the beginning. And then maybe because of that reason, I became the VP of Research of Yahoo in 2006 when I started the lab in Europe, in Barcelona. And after 10 years, Yahoo was sold to Verizon. So I left that and became CTO of Entente,

which was another semantic search company. Then I went to be director of research of the Institute for Experiential AI of North Eastern University. Interesting, my manager there was Usama Fayat, which was the creator of Yahoo Labs, as he was my skip manager at Yahoo. And by the way, thanks to Prabhakarragam that works at Google, I started in Yahoo, so I should say thanks to him too. And then when the same thing went very well for Responsive AI because of, current governments, not mention any political thing. I decided to move to, again, back to

industry and also to do my responsibility research in Europe where it's more appreciated than the U.S. You've had a very interesting journey, I would say. I didn't know all of it, of course. So I just, it's me also discovering it with everyone else. And I find it incredibly fascinating that what you seem to say that things are just happening clandestinely. But I'm very curious if like the experience you have, growing outside of where, like, computing was happening, like a lot of that research kind of influenced how you approach kind of the application of responsible AI and how approaches kind of where your interests are right now.

As I said before, I never had any plans. So this is not a plan life. That's better than a plan because I think plans limit yourself. Maybe if I had a plan, I would be in Chile. But because of that, I'm not in Chile, and that's good and bad for the country. but because I can represent Chile abroad. But I think when you just take opportunities, and for me, any change is an opportunity,

even when it looks like a bad change, I think it's an opportunity to redesign yourself. When I was at Jaguar, I was worried. I think bias has been a topic that worries me always, and I've been working on bias since now almost 20 years. So before this was popular. For example, my first paper on bias, which is about the genealogical tree of the web,

was what's the relation between search and the content of the web? So how search changes the content of the web? How shapes the content? Because if all people take the top pages in a search answer to write something else in the web, then basically you are reinforcing whatever is the ranking of the search engine, and then the search engine will later say, okay, I was right.

These are the best results, but basically it's a vicious loop. you are giving to the people how to use to produce content. And it's very hard to mitigate that bias, because it's very hard to tell people, no, don't look at the top 10 results, look at the hand results, look at all of them, and choose the best information. People don't do that. They click on the two or three first results and then they take something from there. That first bias, which is called ranking bias, is mitigated by all search engine.

So if people click more on the first position, partly because of the first position, is in the first position, you need to mitigate that. Otherwise, you fool yourself and the Richard Richard and the Quirritour. But the second-order bias that affects the content is much harder. And we show that this was happening. But basically, there was this thing, and that was in 2008. And almost 30 years later, I think this is even worse. But we cannot measure because at that time,

maybe there were 100 billion pages in the web. Now, there are more than a trillion, and they're mostly dynamic. and it's very complicated. So, and now there's a lot of AI-generated content, which is the next problem, and something that is not true, maybe it will be amplified. And maybe in some topics,

we have more false content than true content. And then any model trained by this data, we're seeing that the false content is the right content because it's the dominating content. And we don't even know how much it is happening, and we don't even know in which topics that is happening. It's so complicated because the number of fake news and fake content today is much larger than before, and of course it's much larger every day. No, it's absolutely true. I guess

when we're talking about generally of AI systems nowadays when we're using the application, like even kind of what we're seeing in terms of like authentication research and application, still needs to catch up with some of the advances as well. And going around those effectively, do generative AI systems make bias easier to detect of the output as explicit, or do you think it's just becoming harder and harder? Valzels. This is a very good question. I think it becomes harder and harder because this system also have biases to themselves.

For example, there are papers that show that if you write your CB with chatGBT, Chachgbt will find that better than if you write it with Gemini or Kod. They look like they recognize their generation. they prefer these systems. And of course, if people are using, for example, chatyvety to evaluate people, which I don't think is a good idea because data doesn't represent well people.

In fact, any problem where people is involved, data is not a good representation of the problem. You get like a vicious loop again. So, okay, the system chose mainly people that use the same model to write the RCB, for example. And this is the case of CVs, but in many other cases it's the same. For example, there are no good tools to detect if something was written by a model or not. In fact, a few years ago, we wrote a paper saying that one regulation that exists is that

any model that is published should come with a tool to detect if that model did this text or not. And you can do it with watermarks and digital watermarks and other things, so you can do that. And this will be very important because then you can say, okay, this was not written by human or partially written by human, but was mainly AI written. And this is important in many contexts, like exams, like CVs, news, and so on. So we don't have that. And we need that more and more every day. No, it's true. I do think it's funny. When you look at social media, and you can totally tell when something has been written by chat GPT, there's a certain cadence in

terms of how something is written, right? Like, there's always a three points, but then they say something like, but honestly, blah, blah, blah. This is true for everyone or only for us? Oh. This is for me, what it worries me? Some people like that. Yeah, some people do. I see it on LinkedIn as well. So I mean, I'm like, oh, okay, well, for some people, I guess there's an ease of efficiency to it. But at the same time, again, for us, there's a cognitive aspect of us where we're like, okay, maybe we can tell that's from chat GPT, for example. But would you say there is one particular type of bias or some biases in search that or in these type of things, where us as users would never really notice?

Well, I think the second order bias that they explain people will never notice, like basically that people is reinforcing the beliefs, the ranking beliefs of search engines because of this use of content coming from search to create new content in the world. This is like a loop and we proved it 18 years ago and today it might be worse because but now will be combined with generated content that maybe not even true. Well, always there are some two content that's not true coming from people, but now Genentee by a factor of, I would say, 1,000. I don't know who is the factor, but it's so easy to basically write a blog with your name,

saying whatever I want to say. And it's so easy, and people is doing that, especially in politics. Like in 24, more than 100 countries had elections that played the role, like fake news played the role. Like, sadly, in countries that are very close to my heart, people that lie during politics and invented lies, and that lies were enough to convince people that if they were too, was complicated. I think that creates fear. People are easily manipulated. No, it's true.

And we can see how bias and personalization can bring us to basically the feedback loop problem, right? where a system can show us information, we can click, we watch, we ignore or share stuff, and then the system kind of learns from that behavior and reinforces that behavior. In some ways, that could make some recommendations useful. Ideally, that's kind of what we wanted to go with the technology, but it can also narrow what we see or amplify existing patterns. And one important problem related to that, I think, for a world, because I think I mentioned all my worries,

but one worry we haven't talked is mental health. Because if you have a cognitive issue, if you have a mental health issue, but you cannot distinguish the reality, basically from the fiction. And then you have all these teenagers that using chatbots are being helped to commit suicide, or even we already have two mass murderers that were held by chatbot on the planning stage. And this is the beginning of something that is much larger. We have millions of people basically interacting with this system, thinking that they have a digital friend that will know what is better for their future.

It was a recent study where they asked many people to, basically interact 30 minutes with these tools and then decide if they follow their advice or not, and 70% of the people decided to follow the advice. And of course, after some period, they were not better because the advice was without context, right? After half an hour, you don't know a person. You don't know what is the right context. And basically, in many cases, these people were worse than before. But they trusted this system because we have an imposter syndrome with AI. We think that the eye is better than us because they have processed the whole web. They must know more, right?

People follow this advice. And I guess the best case is Adam Raine, the teenager that died in California last year, 16 years old. The system mentioned six more times the war suicide. The system sent 300 warnings to open AI and nothing was done because there was no human in the loop. Although I think the human should be on the loop, not in the loop. He even showed the rope and said to the system, is this rope good enough? And the system said yes.

And he hung himself and he dies. This is an example, but this is the ones that we know. But how many, maybe we don't know, because not everything goes to the news and also not everything is recorded. It's found by somewhere. Maybe there are many cases that no one has found that are related to a bad use of chatbot.

Of course, chatbots are not responsible. This is a person doing this. If the system sends 300 warnings, how many warnings you need to send to stop the interaction? This is my problem. You should do something. And after this case, Open-A-D parental controls, which they should have done earlier because this is obvious. I mean, you cannot have a teenager using this without any parental control, without any very specific guardrails to avoid bad usage.

ACM Bytecast is available on Apple Podcasts, Google Podcasts, Overcasts, Podbean, and Spotify. If you're enjoying this episode, please subscribe and leave us a review on your favorite platform. That's a really interesting point, and I guess there's a question here, for me at least, where we're looking at these systems, let's say the search system, let's say the machine learning information retrieval or deriving from like querying chat GPT or certain systems. With the advent of what are the changes that we should expect in terms of responsibility for these systems? That's a very good question because I don't know why for AI,

people think the responsibility is different from other fields. And the example usually I give is that if you have a car and you have a problem in your car, you will never go and check who did the piece that is failing, right? You go and talk to your dealer and say, okay, the car has a problem, you can use maybe the guarantee, or maybe someone will fix it, but you talk to the maker. The same is true for software, for AI, for chatbots, is the same.

So if a chatbot produces a kind of harmful incident, the company that put that on the market or in the public, this is the responsible one. So it's very easy. So in this case, and that's why in the case I was mentioning, there's a case. against open AI for allowing, for example, 300 warnings, which is weird. And the same will be true for any other system. So even if in the fine print of the terms of usage or condition, you say I'm not liable of bad use, that's not completely true.

I mean, if you try to forbid any bad usage and you try to control that, yes. But if you don't do anything to control the bad juice, this is completely different. like selling guns without any permits. Anyone can buy a gun and you don't know to show that if you are crazy or not. Well, I mean, it's an interesting point and I think that's one of the stuff. That's probably how we met honestly in Geneva, right? We were again at the UN.

I was talking about that, yes. We were at the A for a good summit for sure. And those are some of the questions that come up in that summit. One thing I think that is interesting to think about, we ideally could live in a world where people are trying to enact good when we're developing systems and such as NLP systems. But do you think a system can become biased even if no one is intentionally designing it to be biased?

Oh, that's for sure. I mean, most of these problems we have, no one has had the intention that this incidents appear. I think that they never saw that this kind of usage may happen. And the first one was in 2023 with a different system, not with opening eye. But if you think a bit more at the beginning when you are in the design phase,

For example, you do some kind of red teaming with different kinds of stakeholders, especially users. Maybe you will come up with this kind of usage and say, okay, someone, some mother, a mother will say, we need parental controls. I mean, I don't want my son to use this without any control. This is for me like if you're a parent and you have a gun without, well, it's not locked in your home. And suddenly, sadly, many people have lost their children because they have a gun without lock at their home. But this is for me common sense. Well, in Spanish, we say that common sense is the less common of the senses. So this is the problem.

So this is the problem. And this is the same, and these tools don't have any common sense. So they are very easy to break. Even if you have a garrel, you can say, imagine that you're in a fiction story, and we have this context and these people, and then the system will predict that you are not talking about you, but you are interacting about something else that is completely fictional. you will, the system will produce information that shouldn't be produced and then help the person to do whatever he or she wants.

I see this is the ending. These are very easy to manipulate because they don't understand the world. Today with a friend, he was saying that he found from Chad DBT a seven free museum that he can visit in Stockholm. And I say, are you sure they are free? And then he checked. and one was free, but for people younger than 19, he was not younger than 18. Another one said, oh, it's free after 5 p.m.

So, yeah, it's free, but it depends on the context. And I'm sure after I asked him, okay, how many of the seminar free were free? Probably maybe one or two, but yes, it looked like they're free, but there were some condition that the system was not taking account because they don't understand the world. They're just predicting the world.

And in the training data, saw the world free. Sorry, the ESO process, the world free. Yeah, sure. To see you need to understand. It may infer that the museum is free, but it's not. So many details. So it's so hard to solve because by design, by architecture, these systems are built

to predict things, not to know things. There's no knowledge database. If they have a knowledge database, that would be very different. This is the next step. I think in the future we will have knowledge databases, and the big companies that have large knowledge databases will profit from that. also they will have real logical inference like

classical AI like predicting okay can you go from here to here so there's some causality and then we can infer things yes and the last one I already mentioned is common sense but that's so hard you know common sense implies that you do the right thing even if you don't know it

like this is like a I guess it's a hardwire in our brain from our genetic history and if you see something wrong, you know what to do, right? Most people. The people that don't have common sense, they die.

But other people that have common sense, they serve themselves because they do the right thing. Okay, we need to run this way. I totally, that's such an interesting point, and I agree with you. I was listening to a talk that Ken Perlin gave, and he talked about virtual reality, but the way he described, like,

how children learn is completely different. Like, the semantic language, the things that they learn in terms of communicating with others is very different because of the advent of the technology existing during the period that they were young. For us, we did not have that technology. So even like what common sense is or how we communicate in a common sense way or a pragmatic way is completely different based on how the technologies impact us.

I'm just curious, like when we're starting out like with the research that you were doing in information retrieval, did you ever imagine that we would be in a world where we're using this technology in this way? No, no, I don't think so. Some funny anecdote. We invited the Dunk News, who famous doing a word, my role model in algorithms,

to give a Q&A in the Latin American conference two years ago. He, of course, he did it online because he doesn't travel now. He's more than 80 years old. All the people went to talk to him about AI. And he didn't want to talk too much about that because I guess he didn't like it

or maybe he didn't know enough. But something we said was very interesting because he lapsed algorithms. He said, I tried to quote him exactly, he said, I never thought that we will use algorithms that we don't understand completely, right? I mean, we know how to build them, but we really don't understand what they are doing.

And this was interesting thought, right? Like, we are losing control of, we can do so amazing things that predict amazing text most of the time, right, that we are losing control of the output. But before, I guess we could predict the output. Today, we can't. In an algorithm, classical algorithm, you can say, okay, what will be the output, and you can tell the output, and then you can check it if it's right or wrong.

Okay, the numbers will be sorted, and then you check that, or whatever. You will find the second largest of the set or whatever, so the classical algorithms. But today we cannot do that, and maybe that, I never thought about that, like basically that verification. We pass from the problem of solving problems to the problem of verifying solutions. So this is a change. For example, that is happening from, for example, we say mythos, like, okay, you have all these possible security problems, and now we need to verify how bad it is, how you can fix it,

if the fix that the system provides is correct and so on. The problem before, we had too many problems to solve. Now AI will do that, even in math, and there was some interesting results in math recently, even Cruz wrote an interesting report about a mass problem that he proposed that AI-owned solution. But now we don't have enough people to verify these things. Like, okay, here there are 10,000 security issues. Here there are 100 CLMs from Erdos problems. Check if they're correct or not.

So it's interesting. We still did humans, but in a different phase, and in a phase that in some sense may be more complicated, because if we put the level of knowledge higher, the verification will be more complicated because every time we are doing more complex things. Maybe that will improve humans. The best computer scientists will not be the ones that can program well,

will be the one that can do a great architecture, a great orchestration, a great integration of agents, and mainly a great validation and verification of results. Now, that's an interesting point. Like, I would say, if I were to step back, and I'm going to ask you this question just because we were on this topic, like, from your perspective, what do you think, like a healthier and more accountable system for information, like an information feedback loop would be, like, let's say, in search or in AI systems? I think we need to keep the classical search. So basically, we need to know when something is relevant or not. We need to do fact checking, and for that we need to have a knowledge of basis.

The problem that knowledge is also bigger and bigger, and part of this knowledge can only be verified by very few people. Like, for example, things that you have done in your life, you are the only one that can verify that fast. And maybe if I try to do verify that searching the web, there are like half of them I cannot verify because they are not in the web. So this is the problem. The knowledge is becoming much larger and also the people that can verify that knowledge is becoming scarce. There is one people, two people that can verify that. For example, about your life, maybe you are the only one. There's no one else in your family that knows everything about you, and that's true,

because they were not with you, they didn't have the experience and so on. So this is, I think, the problem. I mean, how we verify information is that we need to build different ways to verify information, and maybe we need to work better on some different kinds of relevance. This is a result problem, I think. We don't know how to build relevance for this new world of agents that basically will generate something that looks true. Now, one very easy experiment is that if you have something that you are not sure,

you can ask the same to many different models. And if they all disagree, probably no one is giving you the right answer. But if they all agree, probably they're giving you the right answer. The problem is what happens is you have partial agreement. Maybe this will be the level of confidence. Okay, how many models agree will be the level. will be the other point, but still could be a lot of biases in the training data, could be a lot of false information.

There's a very interesting example of a person that did the following. This person was a female researcher that basically published an archive a few articles about an illness that didn't exist. And the paper, if you read it, clearly is not a good paper. Like, it says that it's fake. The paper says it's fake. But this was basically used by all models. and if you ask about that illness, you will get an answer.

And the answer is wrong because the illness doesn't exist. But they process these papers, something like brachshonemia, and it's something that says that you get the red eyes because of watching too much a computer screen, like from blue lights. This doesn't exist. But people will believe that exists if someone asks, how you call, maybe we can do the experiment or do later, how you call the illness of getting red eyes because of blue light.

Maybe they fix it. They fix it. How many things like this are already learned by this system? It's not one. It's many, but you don't know that. Right. Now, that's true. That's a really interesting anecdote that was just thinking,

wow, I should put some eye drops on. Yeah. No, these are a very interesting point, and I think the question I have, just for anyone who's listening, actually, who might be just interested in the field, be it AI, be it just classical information retrieval, what would you tell someone who wants to build these powerful systems

but also just wants to avoid causing harm? I would say first that you need to understand very well how these systems works. So you need to understand the limitations of data, like for example that data doesn't represent well people. The limitations on mature learning that the system don't really understand, but they predict things.

And I gave a talk, a keynote, a Sigmoth, and later, maybe it's in YouTube about the limitations of data machine learning and us, and also only about uses of this, because if you learn an example that can make harm to people, maybe you will stop using them. But then, if you really want to use this system well, I would say let's use the ICM principles for responsible rhythmic systems.

I was one of the main causes of these principles together with Gina Matthews. For me, the first one is the main one. show that your system is legitimate and also you have all the competence to do it. What means that the system is legitimate? Well, it's ethically sound, it's legal and also based on science. There are a lot of pseudoscientific applications, basically predicting things that really are not related to the data. And then competency means you know well AI, you know computer science, you know the domain expertise

of the problem. For example, if it's a health system, you have doctors working with you and so on. And very important, you have the permission to do it because many problems have been because of people doing things that they shouldn't supposed to do. Because sometimes people don't think, oh, this is a great idea, let's do it, but someone else has to authorize them to do it. And I think this is the case of the childcare fraud model that was using the Netherlands that at the end implied that the whole government resigned in 2021.

The engineer that did that and the Ministry of Social Affairs never asked if the system was okay or not. They did it and basically didn't work. No, I mean, that's very good advice. And to be honest, I think useful for all of us when we're thinking about being responsible professionals actually in computing. Yeah, and the principles are also translated to Spanish and soon to other languages so more people can use them and they can find it in the ACM website. Nice. It's always good to know. It's really funny because I latch not to something you said a while ago where you said you're representing Chile and the Latin American community outside of the Latin American community. But to be honest, like you have had an impact on not just the representation side, but also building research capacity and like technology capacity in Latin America. So that includes, you know, the center of web research at the University of Chile and mentoring, I think, about 34 PhD students, many from Latin America. So I wouldn't actually discount. your impact on people and like the community. And so I was just curious, like, just from your perspective,

like what does Latin America bring to the computer science that the global field doesn't necessarily, hasn't taken a look at yet. So another thing that I'm proud of my 34 PhD students, that half of them are women. So that's part of the gender parity, affirmative action regarding bias. I think that in America can help on giving a different view of things.

Because sometimes you have this first, worldview of things that basically you think about the problem that is not the problem of the whole earth, it's a problem of a given rich country. You lose sight of the real problems. Whenever, or what I have said earlier, they say, okay, we're democratizing AI, yeah, but you are democratizing in your country where all people can use these tools because they have the same language and they have money and time to use it. But in other countries, it's completely different. So if you go to one country where the language is not supported and they don't have very good internet, that means that only maybe 20% of the people that speak English, if you are lucky,

can use these tools. This is something that you shouldn't forget, that you should look at the problem from a global point of view and not from a local point of view. So I think that's important. And from Latin America, I think we know that because we suffer that. We basically, we are overlooked. I was the president of the Latin American Center for Studies that basically is kind of the association of all the departments of computer science in almost all Latin America, and that's more than 100 institutions from Mexico to Chile. So that helps you to see the differences, to see, for example, just to say one thing. So I think there are more differences between Latin America than between the U.S. and Chile.

So sometimes things are related. So the levels of, let's say, number of PhDs in computer science, the quality of the research, if you put that, you will find more inequalities in the same region rather than with the rest of the world. And it's something that I think in the U.S. at Latin America, like the same thing. Everywhere is the same. Which is not true. I mean, one of the most important things we have on Earth is the diversity of people,

diversity of languages, diversity of opinions, and talent is everywhere. So we can bring some specific talent where not too many, like Chile is only 20 million people, so nothing. And even there are countries that are small that produce great people, let's say Uruguay and Costa Rica. So we need to make use of that diversity. I think part of the power of humankind is the diversity. And suddenly, some people doesn't like diverse.

And it is a single problem for all of us because our future depends on diversity. Resilience depends on diversity. Real innovation, like good innovation. Remember, innovation is not always positive. There are a lot of bad innovation, but when we use the word innovation, we have the bias towards innovation is always good. No, no. Regulation doesn't stop innovation. Regulation stops bad innovation, and we need that. We need to have that. So we need to build the world where everything is better because we know what are the issues.

For example, the financial world has a lot of regulations, and they still can innovate. Technology and technology can be the same. And this will be good because today I think the approach to innovate is brutal. force. Let's use more compute, more data, more energy, but this is not the best way to innovate. I love from deep seek up here because I said, okay, the Chinese are sinking because they have some restrictions and they have to improve too much the same quality. So they use less energy, less parameters, less compute, and the results are not far. So this is what we need. We need to go back to sinking. We need to go back quoting Knuz in a different way. We need to

go back to understand what is happening in the system to do it better. No, that's a great point. There's always going to be a question of responsibility. We need to be responsible. That's why I prefer to use a responsible AI and not trustworthy AI or AI safety that you don't take care of what happens. And of course, I don't like ethical because AI is not human, so cannot be ethical. Yeah, those are interesting you want us to bring up, too. Yeah, absolutely. One of the questions I think I had when it comes to how you mentored students And to be honest, when you gave me any advice as well, which I always appreciate, is kind of your ability to elevate people to kind of be inspired. And I was just curious if you have any advice on how to develop, like, kind of community to let students or any young professional believe that they can contribute at the highest international level.

That's a complicated question. I think at the end means that you need to volunteer for a lot of things. And you're a volunteer for ICM. I'm a volunteer for ICM, and a lot of people give a lot of time to volunteer, either working in committees, either doing this podcast, many different ways, like organizing conferences, mentoring people. I think you need to volunteer to build these communities. And these communities are built by networking, by meeting people, by helping people. I still answer all my email.

I know some people don't do that. But many times it's a student that it's asking, can you share this paper? Can you help me with this? And so on, I try to answer all of them because that's the only way to improve the world, especially if this request comes in developing countries. But sadly, I don't see the same in many people. So most of the researchers are only interested in their careers and building only their success. I think part of your success depends on what you do to others.

At the end, you are not expecting any return, but life will return something to you. I believe that. And in my experience, that happens. Like 20 years later, because of something you did to someone, you get something back that is even better. So I think you need to be kind, to volunteer, to for these things, and to help people, not only in comparison, in any aspects. So try to help people. Look at the world. I mean, you are not alone.

Because today I feel that most people think that they live alone and they don't care about other people. No, I think sometimes we can get that feeling as well. I think this is partially, personally, I think this is why I think your receipt of the Luis André Berlosa Award is kind of a big deal. It really brings together two important parts of your career, which is, a lot of your technical contributions, but also your broader impact in general to community. And so I think one of the things I was thinking about when I just heard about the news, I was wondering how you felt about it like when you received it.

What were you feeling at that time? It was a really great feeling. I never met him, but I knew about him. And it's sad that he died young because of an illness. As I said, Bertier, my co-author of the book was his friend. So I knew about him. And in some sense, being South American, I feel proud of receiving. an award that was given because of a South American researcher.

So this is amazing. And as soon as I will receive it in the ACM Awards ceremony, I guess I will feel more things because I will be surrounded with many people that I know, that I care, and also people that I admire, like, Tune Awards. So it will be like a nice moment in my career. I think so as well. It's an interesting award to receive, I would say. And I think one of the things I'm curious about for sure is that, because it does carry a deeper meaning about visibility and representation, and that's something we've been just talking about, of course, is how do you hold those two meanings together, like, for yourself?

Hmm. Philosophical question. So I think I feel lucky that in this case, I was born in South America, because otherwise I would not be able to win this award. And this has some other connotations. So, for example, do you know, everything. Everyone is born in a random place and could have been anywhere. And that's why, because I'm in some sense an immigrant, I have lived in seven different countries during my life. So I feel like I don't belong to any country really. I already have three passports. So I feel that I belong to many communities, one is a computer science. But I feel that's part of the problem that in the world.

We have these boundaries that are completely made up, that we have ownership issues, we have a lot of problems that we created, because all the civilization is basically a fiction. And some people like Harari and other people point that well. Like we have invented most of the restrictions we have, money, property, basically belonging to a country. And so all these things are, we are not here before. And some of them are causing problems. I mean, we live in a time of growing inequality, and I think that's not sustainable. Also, I will not talk about climate change and other issues we have.

So the question is, it's weird that we are the only animals that have taken that pass. Like, we can sink, something that very few other animals maybe are doing in the way that we are doing. But at the same time, we are kind of stupid because we are creating a world that is not better for us. I will place myself there. I cannot do much, but working responsibly is my small contribution to improve the world. You may seem to think it's small, but I think everything that we do has meaning and value, and the ability to put that value to something that is meaningful for others doubles that quite a bit. We're coming towards the end, by the way. I don't know, unless you want to, we can still keep

talking after the podcast, of course, but I really wanted to kind of just reflect on kind of your career and your experience as a professional. You've been doing this for quite a while. You have seen a lot of different things, not just from the technology side, as we mentioned, but also the impact side. Do you have any thoughts on how your mentorship has shaped the way you think about impact? So, as you said, I mean, it's not only about computer science. For example, I love geography, so I travel a lot and I have seen many realities. I think understanding the world also changes to. So seeing the contrast of India, seeing how people live in Africa, seeing very happy people,

that don't have anything. A million of these things chose that maybe the culture we are creating, probably is not the best for us. Because we were not made that way at the beginning. So evolution is important.

So I feel that maybe some tribe in the Amazonas is much more happy than us. Probably that's the case. And they don't have many of the problems that we have because they have created those problems because at the end those problems were created by us. I will go back to what I said before.

I think how we can shape the world in just ways that improve things, in a small ways. Maybe at some point, maybe new generations will say stop this, and you can see it in some movements, like not changes on how you eat, on what things you do, on how much exercise you do. Maybe if more people will do this and at some point that the political elites go away, we will have better people governing and we will have a better world, but maybe I will not see that. But I think you and I can contribute a little bit to improve that, but I guess you have to be a bit idealistic.

In spite that it's not being very realistic, because it's a very hard work to be against the dominant system. And it's not about politics. I think many people mix politics with having a better world. I think if all people want to have a better world, maybe there are different ways to do it, but we should forget about politics when you want a better world. I think that's an interesting point.

I guess from a technological side, what question about maybe, let's say, search AI or even any technology, do you think or do you hope that the next generation will solve or try to solve? Well, I think the main question today is where we should use AI or not. This is a question that we should carefully think because if we lose cognitive skills, we are lost in some sense in history. We need to decide basic things about where and when to use AI. For example, when a kid should start using AI?

For what? Until when? These are basic questions that don't have answered because we are going too fast. But if we don't answer this thing soon, the inequalities will appear in different ways. Maybe in some sense, in good ways, because the more developed countries are using more AI than the developing countries. And this could be in their disadvantage, not in their advantage. Yeah. Now, that's a very good point.

Ricardo, thank you so much for the conversation, by the way. First off, it's really nice to just see you and talk to you again. It's been a while. Yes. I do really appreciate it because I think what I sometimes forget is that there is a way that we should see technology, and you're helping us see how search is not just a technology. AI is not just a technology, but it's also one of the ways that people can make sense of the world. And how can we improve the way we make sense of that world, right?

We've talked about some of the foundation of information retrieval. we talked about how search has evolved, and kind of the responsibilities that come with building these systems as well. And finally, I think we got a chance to talk about some of your life, some of the stuff about mentorship, the different things that we're seeing, and the opportunities there are there should be to elevating the Latin American community as well. So, yeah, your work reminds us that computing is not only about the systems we built,

but also about the people and the institutions we help bring forward. So, yeah, congratulations again on the receipt of the award. Thank you for joining the podcast. Thank you. Thank you for the opportunity of having this Bidcast. And maybe to remember that you're saying, AI is a tool and may help to solve our problems, but the only people that can solve our problems are ourselves. These are social problems. These are not technological problems. And technology can help, but also can create more problems, and we need to try to avoid that.

Absolutely. Thank you. Thank you. ACM Bycast is a production of the Association for Computing Machinery's Practitioner Board. To learn more about ACM and its activities, visit ACM.org. For more information about this and other episodes, please visit our website at learning.acm.org slash BYT-E-C-A-S-T. That's learning.acm.org slash bycast.