Exploring AI Matters
Our mission is to help the policy community understand the breadth and richness of AI and the potential for such technologies, wisely applied, to augment all sorts of human endeavors.
Some AI tools are able to assist humans in performing tasks faster, more accurately, or more efficiently. Some, however, are inaccurate and unreliable. Who or what we hold accountable for these flaws, and what incentives we do or do not create for their correction will influence AI’s hand in how we work.
In this series we will refine, sharpen, and clarify your understanding of AI.
Exploring AI Matters
Episode 2 - AI as a Prediction Tool
Use Left/Right to seek, Home/End to jump to start or end. Hold shift to jump forward or backward.
Send us comments and questions.
So far, advances in AI are not bringing us real “intelligence.” Rather, these advances are bringing us a key part of intelligence: prediction. This enables businesses to make predictions faster and more precisely to improve their business models and marketplace advantage. In this episode of Exploring AI Matters, Avi Goldfarb, an economist at the University of Toronto’s Rotman School of Management and one of the authors of “Prediction Machines: The Simple Economics of Artificial Intelligence,” will explain the economics of AI and how it can lead to better and cheaper predictions. [2022-06-14]
Welcome to Exploring AI Matters. This podcast series, previously known as Mind the Gap Dialogues on Artificial Intelligence, will continue to appear in the ABA series to the extent that. In addition, all of the episodes, old and new, will now appear under our new podcast name, Exploring AI Matters. Thank you. At least that's what we read and hear. It's processing our job applications, our children's college applications, it's deciding whether we qualify for a loan and which inmates are considered for early release. The last few years, it is getting even closer to us. Our cars use it. Our smart homes use it. Our doctors use it. Weather forecasters and even dating apps are using it. Artificial intelligence has stepped up to the challenge of these prediction problems. They're all examples, and many, many more, that are just performing some kind of prediction. I'm Charles Palmer.
SPEAKER_00I'm Ama Adams. We are your hosts for this episode of Mind the Gap Dialogues on Artificial Intelligence.
SPEAKER_04Today we're talking with Avi Goldfarb, an author along with A.J. Agarwal and Joshua Guns, of the book Prediction Machines, published by Harvard Business Review Press. Dr. Goldfarb and his co-authors are economists, and they bring economic thinking to their study of artificial intelligence. By the way, Avi, my students love this book. And they're a mixture of plain old vanilla security folks and computer science students as well as AI pros. And uh it's the one book they assure me they will never sell. So kudos. Welcome, Avi. Um you don't often see economists collaborating on a book about AI. And it's the first time I've seen it. Um tell us a little bit about this and your co-authors and how you came to work on this area.
SPEAKER_06Thanks, Charles, and that's great to be here. We were professors in the business of the University of Toronto. And um down the street from where we sit is the computer science department. And about 20, well, actually more like 35 years ago, our computer science department made a bet on an emerging field of computer science uh called neural nets. That field became the research area that led to today's excitement around AI. And so out of our computer science department came the future heads of research at Apple, at OpenAI, um, and elsewhere. And so we had this hype and this excitement in our computer science department. Now, for those of you who know how a university works, all that we would normally see in the business school is that would show up in the school newspaper or the bulletin, we'd open it up and say, wow, look at that. There's fantastic things happening in our computer science department. They just won the Turing Award, isn't that great? And then we leave it at that. Uh, but there's something a little bit different, which is a Jay, Joshua and I run this organization called the Creative Destruction Lab. What the Creative Destruction Lab is, is a uh program that helps science-based startups scale. In the very first year of our lab in 2012, we had this company called Atom Wise, uh, run by a PhD student out of our computer science department, of Jeff Hinton's lab, that said they were using artificial intelligence for drug discovery to predict which molecules bind with which proteins. The next year we had a couple more AI companies. And pretty soon we had this flood of AI startups coming through our lab, largely from um you know the computer science department in Toronto, to some degree, uh, you know, Montreal and Alberta and a few other places. And at that point, we realized this this technology was a thing. It was exciting, it was worth digging into. And we put you know back on our hats as scholars, economists who think about technology and technological change, and tried to apply that lens to this new technology of AI.
SPEAKER_00So, as you talk about sort of this thing and this sort of evolution and this flood of AI startups, you know, looking back and thinking about this, there's been a lot of hype and noise about what AI is, what it does, what it doesn't do. And the name of your book, Prediction Games, sort of cuts through that hype in sort of a very simple, focused, measured way. Can you talk to us a little bit about why you focused on prediction?
SPEAKER_06Sure. We call it artificial intelligence. That sounds extraordinary, right? And we think about science fiction, whether you know uh machines that listen to us and do everything that we want, like the Jetsons, or machines that don't listen to us and we're terrified of, like the Matrix. Um, that's not impossible, but it's very important to recognize that's not what we're talking about today. That technology, an artificial general intelligence, has been 20 to 50 years away since the very first AI conference at Dartmouth College in 1956, and it continues to be 20 to 50 years away. That's not what we're talking about today. If you dig into the technology that's generating the current excitement in AI, you realize it's computational statistics. Computational stats sounds a lot less exciting than AI, but it still can be transformative. Why can it be transformative? If you dig in, you realize computational stats is prediction technology, and it's prediction that's a big deal. And once you recognize that, oh, it's predictions gotten better, faster, cheaper, and we use predictions in order to make better and better and better decisions all the time. That better, faster, cheaper prediction is going to transform decision making, which is gonna help organizations change society and have a huge impact, we believe, over the next 10, 20, 30 years.
SPEAKER_00You said something there that I think is really important to focus on. I think people tend to assume that prediction is perhaps the same thing as decision or decision making. Um, and can you just kind of expand on that? Because I think that's a critical component of understanding what artificial intelligence is.
SPEAKER_06This so prediction is the process of filling in missing information. So it's giving you some likelihood that different things are gonna happen. That doesn't tell you what to do, that just tells you what might happen. In order to understand what to do, you need to decide what you value. You need to decide what matters. Uh, my favorite example of distinguishing between prediction and what we call judgment, which is knowing what matters, comes from the movie iRobot. I don't know if you've seen this movie. It's a pretty good movie. Uh Will Smith is the protagonist and he hates robots, right? And so you can kind of see where the movie's gonna go. And why does Will Smith hate robots? Well, there's this flashback scene where Will Smith and a little girl are in a car accident and they're both sinking in a river, and it's pretty clear that they're both about to drown. And then a robot comes along and saves Will Smith and not and not that little girl. And that's why he hates robots. What's interesting is because it was a machine, he could audit it and he could figure out why that machine save me and not the girl. And he learned that the robot, the machine predicted that he had a 45% chance of survival. And that girl only had an 11% chance of survival. And that's why the machine saved him and not the girl. There was this prediction of 45 versus 11. And then the machine saved him and not the girl. And he says, well, no, 11% was more than enough. And a human being would have known that. That's saying something about judgment. That's saying something about what we value. He's saying that that girl's life was worth more than four times his life, right? 11%, 45% to save her, if that was the right decision, her life has to be more than four times as valuable as his. I don't know that all human beings would know that, but that's judgment. That's understanding the reward to an action in a particular environment. That's understanding the payoffs. And it's only the combination of the prediction, the 45 versus 11, with the judgment of how valuable their lives are relative to each other, that we can come to a decision.
SPEAKER_00So we've got the prediction, we've got the judgment, and that sort of leads to, as you were talking about, sort of the outcome, the action and the outcome, right? Sort of the final choices and all of those pieces coming together. Exactly. Exactly. That's it. Yeah, going back on what you were just saying, as um sort of putting the economic lens on this, sort of you were talking about as predictions get cheaper, right? We're able to make more predictions. And, you know, the more we are able to make predictions, right, we can perhaps be getting better outcomes, being able to make better decisions and make better outcomes. So, with that as sort of the backdrop, what are some of the areas where you think AI is most effective at performing?
SPEAKER_06Okay, that's a great question. So the first applications of machine prediction or good old fashioned prediction programs. So what we're seeing today is the application of AI in places where we always knew prediction was valuable. So uh you run a factory and you're worried that your machines might break down. And you may have had predictions before, but when you need to do preventative maintenance and so that they don't break down. And so increasingly we're using machine learning tools, we're seeing AI, this prediction technology, prediction machines, sub in for whatever they were doing before to predict when machines are likely to break down in order to do preventative maintenance. Or you walk into a bank and you want a loan, and the loan officer has to predict whether you're gonna default or not. Uh, a long time ago, they used to walk into the bank, they'd look you up and down, maybe they'd look at your collateral and they'd make some kind of prediction about whether you're gonna pay them back. Over time, that was became a little bit more technical with credit scoring. And now increasingly they're using machine learning tools to figure out to predict whether you're gonna pay them back. Those are good old-fashioned prediction problems. But what's changing is as prediction is getting cheaper, we're starting to realize some things you didn't used to think of as prediction are actually prediction. Image recognition is prediction. How do your eyes recognize a familiar face? They they take in light signals and then they provide a label, right? For you know, based on the context. That's that's prediction, filling in the missing information of a label. Um, medical diagnosis is prediction. What does your doctor do? They take in data about your symptoms and they fill in the missing information, the cause of those symptoms. That's prediction. We didn't used to think of it as prediction, but we now realize it can be reframed engineering-wise as a prediction problem. And so as prediction is getting cheaper, we're starting to realize there's more and more places where there's missing information. And filling in that missing information uh helps us solve the prediction problem. In the legal context, we can see that redaction uh and all sorts of other areas that we used to think of them as cognitive, but actually what they are is there's some missing information. Should that be blocked off, blocked out or not? And we can have a machine do that with prediction.
SPEAKER_00So are there any sort of um real life scenarios or experiences that most of us kind of take for granted? Whereas you were saying, if we had better prediction or um, as you say in the book, prediction machines, better known knowns, right? You use that concept in the book that would result in better options or choices for us, like as everyday consumers, for instance.
SPEAKER_06Right. The world is full of what we call hidden uncertainties. So there's these things that just exist. These organizations exist, these structures exist that you might not even realize they exist because uh because we have bad predictions. So um, my favorite example of this is the airport lounge. So remember when we used to travel, we'd go on planes, and uh, if we were one of the airline's good customers, they would take us to this special area called the airport lounge where we get free coffee and drinks and some other things. Um that was considered great customer service. But if you think about it, it's actually terrible customer service. What would the ultimate customer service be? It would be you walk on, you get to the airport, you walk right on the plane, and it takes off. You only need an airport lounge because there's uncertainty about how long it's going to take to get to the airport, and uncertainty about how long it's gonna take to get through security. In the absence of that uncertainty, if you had good predictions of the time from home to the gate, then we could serve customers much, much better in that industry and get rid of what we thought of as great service to recognize it's a hidden uncertainty, it's actually terrible service. Yeah.
SPEAKER_04Yeah, I've spent way too much time in those rooms. Uh well, okay, so AI is gonna help us in these places where we need prediction and better prediction, more accurate, reliable prediction. Are there some areas that it's just we just don't see it happening, or it's not gonna be particularly good at?
SPEAKER_06Yes. So the thing to remember is these are prediction machines. They're not decision machines, they're not causal inference machines, they're prediction machines. And so what does that mean? It means they're gonna be very good at filling in missing information when they have data that's relevant to that missing information. So, where are they gonna be bad? They're gonna be bad number one at rare events. Things that happen only once, you know, every every few years, prediction machines are not gonna help you solve. Who's gonna win the presidential election in 2028? Who knows? 2024, who knows? This is the world changes too fast between presidential elections to make a good prediction that far out. It's a rare event. The other thing that prediction machines are bad at um are things that they don't have data about because um they didn't happen in the past. So, for example, if you naively feed data on um hotel prices and hotel capacity, how many people are staying in the hotel on any given night, what your prediction machine will tell you is that when prices are higher, you get more customers. Okay. So, right, so you know, a pure prediction machine, if you naively say, well, what price should I charge based on the correlation between prices and the number of people in the hotel, it'll tell you price infinite, price really, really high, because that will get you more people. But obviously, you're missing something important, which is when capacity is high, you know, when you're constrained, when demand is high, you get more people who want to come to the hotel, and then you can charge higher prices. And so where prediction machines fail is with respect to causal inference. Let me give you another example of that. Um so I think you know, we wrote our book prediction machines. I think it's pretty good. Um, and we ask in the book, is reading this book a good idea? Okay, now imagine you could see your future self. Okay, and your future self, after having read the book, is fantastically happy. You're let's say you know you're spectacularly wealthy, your personal life is fantastic, and you think it has something to do with your understanding of AI. Okay, so not are you not only are you happy in every way you can possibly imagine, but you can link that to your understanding of AI and your interest in AI. Can you then say reading the book was a good idea? I'd love to say you can, because I wrote the book, but you can't. Why can't you? Because you don't know what would have happened had you not read that book. That counterfactual you is missing data. And because the counterfactual is you is missing data, the prediction machine can't tell you whether having read the book led to a good outcome. And so that's all you know, two different ways of saying these machines work well on the support of the data they have, and they can fail and often fail spectacularly when they don't have data that's like the situation they're currently predicting.
SPEAKER_04Yes, it's it's all about the data, it sounds like. And then there's quality of data. I mean, suppose I've got tons of data, but it all stinks, you know. Well, you get what you get, right?
SPEAKER_06Absolutely. Um, and so there's versions of quality of data that things are rare, there's issues around quality of data, like it's not in the counterfactual, and there's all sorts of other challenges on data quality. So, you know, there's I have to think of two broad categories, which is one is the data going in is biased in some way. Okay, so there's uh an example of a hiring prediction machine that Amazon built and never deployed to their credit, uh, where they were predicting um who they should hire. And what were they using in predicting who they should hire? They were using past hires. They were trying to predict what their human HR people uh would select based on CVs. And what they what their prediction machine was doing once they ran it and started experimenting with it, is it was rejecting all women. Okay. And not just rejecting women, rejecting men who mentioned on their CVs that they'd done things like coached a women's soccer team. So they tried to fix it with various clutches and then decided you know to throw out the whole project and start again. And where they presumably started again was with their high human hiring practices, recognizing there was a bias in the data going in that uh led to you know, a bias in the predictions of the machine. So issue number one is if the data is based on biased decisions that's fitted into the machine, you know, say biased human decisions, the machine's gonna be no better than human and make the human issues happen as you know, concerns at those human biases happen at a higher scale. The other area where uh the data can go wrong or you can have bad data is if you have an adversary, say a competitor who's trying to mess with your data. Okay. So a strategic challenge, you know, both in the corporate world and in the geopolitical world, is that you can have adversaries who are trying to mess with your data in order to give you bad predictions. And that is even worse because they're gonna be doing that so that those bad predictions happen at the times that they're most consequential. And so, again, we have to be very wary about using prediction machines, um, you know, especially automating the results of prediction machines, uh, in the, you know, in situations where the stakes are very high, and in situations where either your data might be biased or an adversary could mess with it.
SPEAKER_00Given the critical role that data plays in the function or in the realities of prediction machines, um, and you walk through some of the limitations that we can run up against with respect to data, you know, bias in data, maybe for privacy concerns, you don't get full access to a certain subset of data, right? Which kind of limits what the machine, the prediction machine can do in terms of prediction. Are there particular areas that you see AI being ill-advised in for use at this time, or where the risk-reward calculation might steer away from AI from a particular situation?
SPEAKER_06Um, it's part of me that just wants to say no. Uh that's not quite true. So uh the particular situations are when you don't have the data. Okay, so the data might be biased, the data might be messed with by with a competitor, uh, it might be illegal to collect the data for privacy reasons. Uh, in those situations, you want to be very careful about using an AI for prediction. There's a second point, which is we have the AI and the prediction side. Then we have to decide once we have a prediction, uh, do we automate and sort of pre-specify the judgment in advance? So the AI, the AI doesn't make the decision, but sort of a human makes the decision in advance by pre-specifying the judgment. Or do we have a human see those predictions and then make a decision one by one as the predictions come in? And so a lot of the risks on AI, I don't think, are on using AI for the prediction. A lot of the biggest risks are on the automation side, which is you know, should you have a human in the loop applying judgment as each prediction gets in? Or can you just say to the machine, uh, we know what the judgment's gonna be, the appropriate judgment's gonna be in this situation. And we know sort of everything in all the different predictions what's gonna happen forever. And so forget the human. Let's have a human today, when you clear up and forget the human, let's have a human today make the decision to go along with all those future predictions. Um, that creates extraordinary risk because you know the human might not understand all those potential issues. Um and you know, in terms of scale, and uh you know, the human might have different values than you would expect sometime in the future. And that scaling, or the human might not even realize. They might think, oh, I'm making the deceived machine decide, I'll pre-specify judgment. They might not even realize the power that they have as they're doing that reward function engineering.
SPEAKER_00So, given that backdrop, curious, it how do we know if an AI or prediction machine is doing something badly? How can we how can we judge that? How do we know that or can we know that?
SPEAKER_06Um, what's fantastic about machines is we can audit them. Okay, so if we have humans who are making biased, horrible, bad decisions, it's very hard to go back and figure out what the decision process was that led to those decisions. It's very hard to figure out, you know, to aggregate up over you know, we humans don't make that many decisions, so aggregate up and see if it really is a bias, or if they can say, oh no, it was just you know, that particular case was like this. So, with what's nice and uh hopeful and optimistic about using technology is number one, we can audit it. So we should have auditing processes in place to check against known biases or anticipated issues. And number two, it doesn't only have to be retrospective, it can be prospective. Because with a prediction machine, not only can you audit, but you can simulate uh opportunities. So it can be prospective, not just retrospective. So you can say, we are worried that the machine is going to be biased against uh certain genders or uh races or countries or people with particular tastes or particular literary interests or whatever else it might be. If we anticipate and we can hypothesize what those biases might be, we can prospectively, proactively simulate uh to reduce the chance that the machine has uh these kinds of issues. So I'm actually quite optimistic about the potential of machines because we can audit them and we can simulate.
SPEAKER_04Yes. So one of the things that does trouble me though is how broadly is this gonna be available? That is, is it is it gonna be like all the other technologies that shook up the world and help some more than others or sooner than others, or will some just never see the benefit? How's this gonna roll out?
SPEAKER_06So here I'm less optimistic, uh, which is there are good reasons to anticipate that AI is going to increase inequality. Now, I should, you know, there is a narrative that a lot of the people who do prediction and whose jobs are prediction earn a very good living, whether it's financial analysts or medical professionals. Uh, and so maybe it'll you know reduce inequality. But there's two forces working the other way. Force number one is we should anticipate, I think we can anticipate, that using AI well is gonna be like the past 50 years of information technology. It's going to require skill. It's gonna be skill-biased technology technology, skill-biased technical change. And so the people who have the skills to use it, who typically are better educated and already doing well, are gonna increase their relative income. So, like uh past generations of information technology, we should expect this to be skill-biased. Uh, this is you know, ideas in uh Claudia Golden and Larry Katz have uh been the main uh proponents of these ideas, and we should expect that AI is gonna be like IT. So that's inequality increasing. Reason number two is it's capital. It's so it's a machine, and somebody's gonna own the machine. And now we have, well, the people who own the machines are going to get the rents for them, they're gonna earn money. And the people who you know earn income from labor who tend to be poor are not. And so it's another reason to expect increased inequality as a consequence of this. And so we have a very difficult question, a very difficult trade-off here, uh, in thinking about the future on wealth versus inequality. Now, a lot of the narrative um is a lot of the narrative around AI talks about is AI going to replace jobs? So um I'm not worried about AI replacing jobs, not worried about it for two reasons. First, uh there's good reasons to expect that new jobs will arise, just like many of our ancestors were involved in agriculture. Most of us are not today. We still have jobs. The uh as one sector becomes more productive, we get more work in other sectors. Okay. That's doesn't always have to be true, but it's historically been true. The bigger reason, the more important reason um, I think to not worry about jobs is jobs are called jobs, they're not called, you know, the word is job, it's not a fun, right? We we the major victory of the 20th century in many ways for uh humans in the United States and Canada and elsewhere in Europe and elsewhere has been the rise of leisure time. We work 40 hour weeks, uh, we uh have time to be children, we get to retire, we get vacation. That's all good. So, you know, and you know, in contrast, if you think about the movie The Matrix, you've seen that movie, every human in that movie has a job from the day they're born to the day they die, right? Their batteries. That's not an optimistic future. So this the narrative on AI replacing jobs, I think is just the wrong narrative. There's reasons to worry, but the reasons to worry are around inequality. If the people who benefit are those who are already successful or already uh wealthy, already educated, or in contrast, maybe it's just the future. Maybe they're not wealthy today, but they might be in the future, it's still going to increase inequality. Uh, and that uh you know that creates a tension between the growth, economic growth that AI could bring, that economic potential on just the growth level, but increasing inequality.
SPEAKER_04And there's there's individual uh inequalities, like we just like you just said. But then there's also um geopolitical or you know, this country, that country. Uh are we about to step into, or maybe we're already in an AI technology race, like the space race of the 60s.
SPEAKER_05That one requires some thought. Okay. I'm gonna be careful in my response.
SPEAKER_06Uh so there are there are certainly politicians and technologists uh in the United States and China and elsewhere that think we're in a race. Okay, an AI race, and there are billions and billions and billions in investment going into AI because of this race. Um, Vladimir Putin a few years ago said, you know, whoever controls AI is going to control the world, or something uh ominous like that. Um and we've seen not quite as blunt but similar statements coming out of both the United States and China. Um, so in that sense, there are reasons why, you know, geopolitical reasons why we're seeing massive investments in AI and lots of excitement, both both the military potential and the economic potential of this technology. That said, there are reasons to think that that um competition is misleading and that story is misleading. Uh, reason number one is this technology is pretty open. Okay, so uh so I'm sitting in Toronto and a lot of the technologies were invented in Toronto. Okay, that sounds wonderful, but I didn't, when we said you know, the country's leading the world in AI in terms of the economic potential or the military potential, I didn't hear Canada. Okay. Why didn't I hear Canada? Because you know, it was published and it's not that hard to deploy software in many ways, and it's been American companies and Chinese companies and some others that have gotten economic, the biggest economic returns. You know, Canada's you know uh you know doing a little better than they normally would in the in high tech, but that's you know, it certainly has crossed the border. So reason number one in thinking about should we really think about this as a you know as like the space race, is it's not obvious that the rents, the benefits are highly localized the way they might be in space. Reason number two is um in particular, you know, in many contexts, military contexts and others, uh, you adjust to your competitor. And so just because the United States, say, is great at AI doesn't mean its adversaries will say, well, the best way to beat it at AI is to get even better at AI. Uh it might be to do something different and to identify a different strategic advantage. And you might be thinking in the military context, but that's also true in the economic context. Okay, so um the you know the uh a useful story around that economic context is thinking about the you know the 19th century, believe it or not, and Switzerland. Okay, so Switzerland in the early parts of the 19th century did not have patents. And so you might think that wouldn't have been good for Switzerland as an innovative place. The rest of Europe had patents, they protected those patents, the US had patents, and so the best innovations were coming out of the rest of Europe in the US. But what happened was Swiss innovators innovated in areas um where patents were unnecessary. They got great at clocks, for example, which required you know uh tacit knowledge and skilled trades much more than patents in order to build great clocks. And so in AI we may see the same thing, which is some countries may specialize in building economic returns and economic uh powerhouses, companies around AI. Other companies, it's not the only technology out there. And so others may specialize in something different. And that's uh so you know, maybe there'll be a race because we want there to be a race, but it's not clear that uh that kind of a race has a winner uh because there's lots of different opportunities around for technological change.
SPEAKER_00I like the spin you had there on that narrative around sort of the AI race and sort of changes the way you think about the dialogue and the conversation. And that leads me back to a discussion point that we were just going through. Sort of, as you noted, one of the biggest sort of maybe fears or I'll say paranoia about AI is that it's going to put us all out of work. Um, kind of taking us down that thought process um and sort of putting some distinct points around that. You know, in the book, one of the things that you and your co-authors note is that prediction machines are better than humans at taking and factoring in complex interactions around different integers, I think is how you kind of spell it out. Um, could you elaborate a little bit on why that's a particularly good use of AI or why AI is better than that, better than that, better at that than humans? Sorry, better at sort of factoring in complex interactions, sort of taking in all those sort of data sets and interactions.
SPEAKER_06Okay. What today's AI does really well, what this um the main technology, deep learning, convolutional neural nets and others, does really well is take in lots of different variables and put them together to make a prediction about something else, one on another period. So um it does that with images, it does that with text. Our human brains actually with images are quite good at that because you know we have a whole bunch of pixels and we can interpret that as a face or you know a book or whatever else. But in other contexts, when the data is sort of unrelated, our brains aren't good at making sense of it. It's just we don't know how to interpret a hundred or a thousand different data streams in order to make a prediction. And um machine learning tools are better at that. And so they are going to be good at combining data streams to make a prediction, as long as that prediction is on the support of the data you care about, of the decision you care about. And so as long as we're not doing causal inference thing, doing a prediction problem, uh, that's fine. So back to your will AI then put us all out of work. Well, no, but work will change. Okay. The aspects of your job that are prediction, uh, particularly prediction based on rich data or that could be based on rich data, are going to increasingly be done by machine. Uh, but there's still plenty left for us to do. Where we can uh help figure out what data should go into the machine. We can use judgment once the prediction comes out to figure out what to do. We can work in areas that for whatever reason are just inherently human, where we think there's real value in having a human performing that service instead of a machine, whether it's you know raising children or um perhaps taking care of old people. That was an aside. Um, Daniel Kahneman, uh, the Nobel Prize winning, he won the Nobel Prize in economics, but he's a psychologist, as he uh would regularly point out. If I called him a Nobel Prize winning economist, you'd correct me. Um said, I think actually when we're old, we're gonna want machines to take care of us because machines won't get hungry or tired or frustrated. They'll know how to, they'll they'll be able to predict what kind of actions uh we'll respond positively to. And therefore, so we still we'll still want our grandchildren and our great-grandchildren to come visit us, but we won't thought we won't want them to nag us about whether we take our pills. We'll just want them to come visit us and hug us, and then the machine will will take care of our needs in various ways. So, okay, so I know that was a a digression, you know, sort of me disagreeing with myself. But the the overall point is there are reasons to think that um AIs will increasingly do prediction tasks. There's lots of other tasks. And as the prediction component gets cheaper and cheaper and cheaper and cheaper, we're gonna see humans and more human jobs in those other tasks.
SPEAKER_00Yeah, but I think you the way you framed it is excellent. It's really we need to kind of change how we evaluate the way we look at sort of the job, right? It's a series of different tasks that go into defining what is a job, an occupation, a career. And AI may be good at some parts of that. Humans will continue to be better at some parts of that. Um, and that's how the synergies work together to get us to a better outcome.
SPEAKER_03If I could add a add a sort of an interesting metaphor, um, I would much rather ride in an elevator that's controlled by automation than an elevator that's controlled by a person because those human-controlled elevators were weird.
SPEAKER_04Okay, but but uh to that, since we're riffing a little here, uh I was in Washington and I went to the Marriott to a meeting, and I asked what meeting the floor were what floor the meeting was on, and they directed us and off we went. And I went to the elevator lobby, and there was a bunch of people standing around looking totally perplexed. And you know, I just thought, geez, uh have they never seen elevators before? And I walked over and there were instead of the button up and button down, there was a keypad, and it said, type in the floor you want. And well, I I'll follow directions. I did, and then it said, okay, the most efficient car is going to be that one, and it put it, it was all there, and so perhaps by now, a few years later, everyone walking into the Marriott gets it, but right then they had no clue.
SPEAKER_06And yeah, the there's a training aspect to it, and we need to learn how to how to work with machines. Um and you know, a hotel where people are coming in and out, and new people are coming in and out every day is is gonna be a very difficult place for yeah, it's gonna be it's gonna be a hard place to be using AI in that way. Where an office building or residential tower, where most people are there every day, by day two, they figured it out. Um, and so it'll be a little more efficient.
SPEAKER_01Can I just add one other question along this line? Um, IT staffs have grown enormously as computers have proliferated. As AI proliferates, will there be a need for huge AI support staffs that choose the data, uh, identify what questions AI should be asking, et cetera? Can you give us what you would imagine might be the shape and substance of those staffs?
SPEAKER_06Um yes, you will need some AI staff. Some of them will be business people, like the business staff, the leaders who think through strategically, how can we use prediction well? Um, and those exist in you know in the IT context too. But we'll need people to organize data, to structure data, to figure out what kind of data goes into the AI, to identify what things can be predicted, how good those predictions are, uh, put bounds around the predictions have statistical confidence, they're not perfect, uh, understand the risks of that uncertainty around the statistical confidence, apply the judgment to all sorts of tasks uh that are separate from the prediction task. And in many contexts, in today's workflow, the prediction task happens like that. We can just make it way, way better. But we don't even think of it as part of a person's job, uh, although a person might do it not as well as the machine. And if the machine does it well, we now we may be able to have more humans in that workflow. So for example, um in triage in emergency departments and hospitals, uh, that's it's a prediction problem, really, in terms of where does the person go? But that prediction problem does not take a lot of time. That the time is often in other tasks like entering data and stuff. Those tasks aren't going away. But the prediction problem of where to triage the person will now be with a well-functioning AI, should be much better than it is with that tired person who's been working that shift for 16 hours. Um, that should improve medical outcomes. You may even need more medical staff because you're doing the right thing. Um but uh but you're gonna have AI doing prediction and doing it better. Um, another challenge, though, that I you know that comes up in this context is sometimes the predictions are done by as part of training. So you know, in the legal profession, actually, often your predictions, you know, your training process is read these documents and figure out what's relevant. Or read these documents and black out what somebody shouldn't see. And that teaches people, you know, junior lawyers, what matters and what doesn't. It teaches them the law, and then over time they can apply deeper judgment in the human skills that the machine can't do. But what happens when we have machines doing all those tasks that we use to train the junior people in the profession? Well, now we need to transform education. We need to think about well, how do we how do we get people up there when they don't? We need to think what do we do with junior lawyers before they're ready for those higher level tasks? Um and and how do we maybe restructure those aspects of the profession? It's an open question. We don't have an answer to it. I know law schools are grappling with that question today, uh, but I haven't seen a Ah, here's how we need to change legal education so that when people graduate, they can work with the AIs and act as if they've been lawyers for five or ten years.
SPEAKER_01We're not there yet. But we need to get there. When the IT era began, only large companies could afford computers. Then they proliferated into desktops, laptops, our phones, our wrists. Do you think the same thing will happen with AI or will it remain the the owned and run by the companies that can afford the large machines and can afford to assemble the large data sets?
SPEAKER_06I think that depends on regulation. What do I mean by that? Um right now, when we think about um privacy, we tend to think about it corporation by corporation. So if you think about GDPR and Europe's regulation, or uh even some of the regulations around data in the US, um they focus on Google can collect data and use data, for example, uh, but they can't share data with others between companies. And in that world, it gives an advantage to being big. So under a world where data flows between companies are restricted, then it's very clear that there are going to be advantages to being a company that collects a lot of data. In contrast, if uh the regulation is such that companies are able to combine data, small companies say, are able to combine data uh and you know in efficient ways without a large legal barrier to that, then small companies can benefit from AI just like big ones. Because there's no technical reason, there's no real technical reason that gives you increasing returns to scale and data. In fact, as a math topic, it's the opposite, which is improving the square root of it. So there's decreasing returns to scale and data. And so where we worry about uh big companies dominating AI, I think there's a direct line from there to how do we regulate the exchange of data between companies. As long as small companies can exchange data with each other, I don't think we're leading to a future where the big companies continue to dominate.
SPEAKER_03A funny side uh note on this uh on predictions being tampered with by adversaries. If you go back to World War II, uh one of the challenges that the British found with the bombing of England uh by the Nazis was that the Germans were remarkably good at targeting their bombs and nobody knew how to what was going on. And a fellow named RV Jones, who who ran the uh scientific intelligence unit for what became the RAF, um studied that and figured out that what was happening was that the German bombers were following a radio beam, uh, and they had a technology which let them stay on that beam, and then they were crossing beams, which told them when to wake up, when to you know open their bomb bay doors, and when to drop their bombs. And what uh so if you think about it, these were predicting in a certain in a very static, very limited sense when to do certain things and what to do. And what he then did was having figured this out, uh, set up a set of transmitters that warped the beams that the Germans were following as a result, causing them to drop their bombs in empty fields rather than on factories and cities. So it was a fascinating uh historical thing where if you look at it with the lens that I think the the that the the your your book, Prediction Machines uh brings, uh makes it really an interesting way to understand some of our history. It's a great story, Mark.
SPEAKER_00Thank you, Avi. I just wanted to sincerely thank you for joining us today for this episode of Mind the Gap. And we really enjoyed the illuminating insights that you've provided us today. So thank you.
SPEAKER_06Thank you for having me here. Uh, it's been great to chat with you and take care.
SPEAKER_04The views expressed in these podcasts are solely those of the speakers and not of their employers or organizations. We thank the American Bar Association for their generous sponsorship and support of the production of this podcast. Our theme music was composed and performed by the very talented Ben Rosenblum. We welcome questions and comments from listeners. Send email to comments at MindoftheGap Dialogues.com. We read all comments and questions and will try to respond in the letters section of a future episode. If you're writing about a particular episode, please mention the specific episode number. And please also include pronunciation tips to help us properly say your name when we reply in a subsequent episode. See you next time on Mind the Gap Dialogues on AI.
SPEAKER_02Thank you for listening to the AVA Business Law Sections Podcast Series to the extent that the section offers a robust collection of content. To explore more about this topic or to learn about joining the section, visit ambar.org slash bizlaw. That's B I Z L A W.