B (3:16)

That's great. And we're going to talk about that book also in this conversation. But I want to start a different place. On this podcast, I interview a bunch of AI researchers and they have a perspective on AI and the economic impact of AI that differs from how economists tend to think about it. So how do you think economists and AI researchers disagree on? On the future of AI and the economic impact?

A (3:42)

I think it's changed a lot in the last decade or so. The kind of issue that I really cut my intellectual teeth on back in the beginning was a sense that economists were systematically underestimating the capabilities of technology. This was the observation I was making back in 2010, 2011, that in particular a sort of governing idea in the sort of economic literature was that, well, look, machines can perform routine tasks and activities, but they can't perform non routine ones. Things that require faculties like creativity and judgment and empathy. And yet what we could see even back then was gradually but pretty relentlessly, more and more non routine tasks were being taken on by the latest technologies and eventually by AI. And I was interested in why it was that economists were making this systematic mistake in thinking about the capabilities of technology. And what emerged from that was.

B (4:49)

A.

A (4:49)

Realization that economists were using quite an old fashioned conception of how it is that machines and technology works. A view that if you wanted to automate a task, you had to sit down with a human being, get them to articulate how they perform the task, and then write a set of instructions for a machine to follow. And that was true 40, 50 years ago, when people in AI were working on expert systems and it was all very top down. And it's true that back then, if a human being couldn't explain the sort of particular rules or kind of thought reasoning processes that they went through, it was very hard to see how you might automate a task. But of course, what's happened in the decade since is that machines don't have to follow the kind of explicit instructions articulated from the top down by human beings. We make medical diagnoses now through AI, not by trying to copy the kind of particular rules that a doctor follows set down for the system to follow, but by learning from the bottom up through lots of data. And so this was the sort of, I think this was the thing that, the mistake that economists were making. It was not quite appreciating how it was that these new technologies were working and what that meant for the kind of traditional boundaries they'd drawn between what tasks machines could and couldn' not do for computer scientists, I think in the way in which. And you still see this to some extent today, but less, less than you did. But again, 10 years ago, I think the way in which many people in AI and computer science were talking about the impact of technology on work were neglecting a sort of fundamental realization that economists had made in thinking about the impact of technology on work, which was that technology can have two very different impacts on work. On the one hand, it can substitute for human workers, displacing them from particular tasks and activities. And those are the sorts of examples that I think capture our imagination and capture the headlines in the popular press the moment a machine outperforms a doctor or outperforms human driver or wherever it might be. But there's also, and so there's that sort of harmful effect of technology on work. And I think that's what many, you know, technologists were focused on. But there was also a second, far more helpful effect of technology on work, which is that it sort of, it could complement workers as well. It could increase the demand for human beings to do tasks that hadn't yet been automated. And the way in which that process worked was far more subtle. And the actual impact of technology and work depended upon the sort of battle between these two forces and kind of harmful substituting force, a helpful complementing force. And I think a decade or so ago, computer scientists, AI researchers, technologists in general were quite bad at recognizing the sort of dual nature of the impact of technology on work.

B (8:05)

So.

A (8:07)

In short, I think economists have learned a lot in the last few years by understanding far more about how these technologies work and what that means for their capabilities. I think similarly, computer scientists have learned a lot about how the different ways that these technologies can actually affect the work that people do. And as a result, the kind of indeterminate, the kind of uncertain aggregate effect that these technologies can have on work is not as straightforward as focusing on these sort of cinematic substitution effects.

B (8:40)

One complaint I hear from economists is also that it takes that computer scientists and AI researchers in general underestimate how long, how much time, how much effort is involved in implementing these technologies into existing companies and workflows and how long it takes for these technologies to diffuse in the economy in general. So having. Demonstrating something in a controlled lab setting or in a test environment is very different from it being implemented into the economy. Do you think that's true? Because you can also make the opposite case that something like coding, for example, is ready to be implemented basically the moment it's created.

A (9:27)

So I think what you're describing is really important, and it's important for explaining one of the kind of big puzzles about technology and economics, which is that anecdotally, we appear to be surrounded by stories almost every day of technologies taking on tasks that we thought only human beings alone could do. Remarkable stories. And yet when you look at the productivity statistics, that famous line that economists have said over the decades, which is you see technology everywhere apart from in the productivity statistics, I think one of the really compelling explanations for that is exactly, as you say, the lag between the kind of invention, the innovation, and then the amount of time it takes for these technologies to actually kind of find practical use and then diffuse through the economy. And you saw that, you know, say in the Industrial revolution, you have this kind of wave of extraordinary technologies around 1780, you know, the spinning jenny, the roller spinner, the power loom, all of these. And yet it takes many decades for the effects of those technologies to start to here in the productivity statistics.

B (10:40)

And is that the most important metric, productivity? Or what should we be measuring? What should we be noticing? Is it gdp? Is it productivity? Is it unemployment rate? What is most important for us to notice when we want to keep track of AI?

A (10:55)

Thinking about Britain today, sitting here as I look around, public services backlogged and broken, average real wages haven't really risen for 16, 17 years. They're sort of worse. Run since the Napoleonic wars, worklessness rising. There are very few problems in the British economy which would not be solved by more productivity growth. And so if you're looking for, and it's true more generally that almost every measure of human flourishing is associated with a growing economy, and a growing economy requires productivity growth. So I think in thinking about the sort of, the benefits of technological progress, productivity is really, really important. So that's one measure that we ought to be paying attention to and trying to understand, are we, and trying to also understand the limits of the ways in which we measure productivity and this sort of productivity paradox that we see all these technologies around us and yet they don't seem to be making a difference to the way in which we traditionally measure them, which is the productivity statistics. But I also think, I think you're exactly Right. As well, to sort of point to the unemployment. I think a running theme through my work is that we're just not taking the impact of these technologies on the work that we do seriously enough. And as a result, I think paying attention to what I write a lot about, this idea of technological unemployment is one of the things I'm paying particularly close attention to as well. I think it's worth saying, though, and it's quite important, particularly in the shorter run, that it's unlikely that we're going to see mass pools of unemployed people due to technological change. It's far more likely that what we'll see, and to some extent, I think we already do, is technology not affecting the quantity of work, but the quality of work. Not the number of jobs, but the nature of the jobs that are out there. And I think in many respects, whether it's the pay of the work that's available or all the different dimensions in which the quality of work might run, the way in which those are being either improved or indeed degraded by technological progress is really important.

B (13:32)

Is there a way for us to capture the productivity without getting the unemployment? So is there a way for us to steer towards technology, especially AI, that complements labor as opposed to replacing workers?

A (13:48)

Yeah, I mean, this is one of the big hopes, and I think the kind of general philosophy is right, which is that I think very often when policymakers and politicians talk about technological progress, the metaphor they have in mind is that they're like a sort of train driver sitting in a train, and they can kind of push down on the throttle and speed up and get more technological progress, or pull back on the throttle and slow down and get less technological progress. But the sort of. Their direction of technological travel is kind of fixed by the rails that are set down for them to trundle along. I just don't think that's the right. And so, in a sense, the only question that matters is do we want more or less technological progress? And that's not right. And this is a big argument of mine, the most recent book, that actually, a far better metaphor is a sort of nautical one. It's far better to think of ourselves at sea on a little boat, and we can pull up our sail and go faster or pull it down and go slower. But we also have a huge amount of discretion over the kind of direction of technological progress as well, the sort of nature of technological progress. And, yeah, I think that's true of the impact of technology on work as well. Just to go back to that distinction, before technology can have very different effects on work. It can complement, it can substitute, it can reduce the demand for the work that human beings do, or it can increase it. And that characteristic of technology isn't fixed. It can be shaped by the incentives we create in the economy. So just one example, it's really interesting that in Every year since 1981, the effective tax rate on hiring a human worker has essentially been higher than using a machine. In other words, at the margin, the US tax system seems to incentivize replacing a worker with the machine. In other words, it seems to incentivize the development of technologies that substitute for workers. And so you might say that's the sort of incentive that you might actively intervene to change. So there's kind of really interesting and I think important conversations at the moment about how we might steer technological progress, exactly as you say, away from technologies that substitute for workers towards those that complement them.

B (16:19)

But do you think that's ultimately possible? I'm guessing there'll be strong economic incentives for, for developing technology that, that substitutes for, for human workers as opposed to complement human workers. Just because if you fully substitute a worker, you can then cut out the middleman, so to speak, and, and get the productivity without having a human, without having to pay a human worker.

A (16:44)

Yeah. So I do, I, I do think there is a lot we can do with this, but I do think there are limits. Not, not, not the limits that you're describing, but one of the big limits is just practical, which is that ex ante, it's quite difficult to know the impact that a technology is going to have on the labor market. If you say to a computer scientist, well, look, is this going to substitute for a worker or compliment them? They'll say, I don't know, I need to see it in the market, I need to see how people use it. So it's quite difficult to anticipate on that point.

B (17:17)

Actually. What's the best research we have for trying to predicting whether some technology will substitute for human workers or complement human workers?

A (17:32)

I don't really think that literature exists. It's not something. And just to kind of explain why, think about something like the automatic teller machine, the atm, when that was released, just intuitively it feels like this is going to be a disaster for people working in banks. Their job is to hand out money over a desk. Oh my gosh, the ATM automates this. This is going to be the decimation of the kind of, the world of bank employees. And yet if you look at what happened in the United States States once The ATM was rolled out. The opposite happened. You had a sort of surge in bank tellers. And the reason why that happened was, in retrospect, entirely understandable, but ex ante, quite difficult to anticipate, which is that in part, the nature of what it meant to work in finance changed. And so people weren't handing out cash anymore, but they were offering financial advice or offering kind of types of new financial products and so on. And so the nature of the work changed, but also the US economy grew as well. And so there was new demand, new types of financial. So the point again, which is that this kind of helpful complementing force works in ways which are quite subtle and quite hard to anticipate in advance. So that's part of it, which is that ex ante, it's just very difficult to anticipate what impact to take technology is going to have on work. There's also the added complication which is that kind of ex post, the impact that these technologies have on work can change as well. If you think of something like a GPS system in a car, in a world with human drivers in the seat, a GPS system complements human drivers. It makes them more productive at the wheel. It allows them to navigate unfamiliar roads. It's a compliment. But in a world where we have driverless cars, well, these GPS systems just complement the sort of the driverless car instead, making them better. So the same technology, over time, the effect that it has on the demand for work can change. There's also, I think, and this is a kind of deeper and deeper question, which is not a kind of technical one, about whether or not technologies are going to complement or substitute, but a kind of moral one about the implicit assumption in redirecting technology away from technologies that substitute towards those that complement is that there is something important about work that we want to protect. And two obvious reactions. One is that work is an important source of income. It's the main way that we share our income in society. For most people, their job is their main, if not their only, source of income. It's also a way of allocating meaning and purpose as well. Of course, work isn't simply a source of income, but it's also a source of direction and fulfillment and structure. And so you might say because of those things, there are good reasons to want to redirect technological progress away from those that substitute towards those that complement. But as you'll know, there are many people who argue, well, there are other ways of sharing our income in society other than through the work that people do. Work is not the only way of solving the sort of distribution problem of how you share out prosperity in society. And well, hold on a sec. Also, when you think about work and meaning, lots of people really don't like the work that they do. And lots of people would if they could find meaning and purpose outside of the world of work. And so actually it's not entirely obvious that we want to be steering technology away from substituting towards complementing. In fact, we ought to be doing the opposite. And a world with less work, where people get their income from non work and opportunities and find meaning and purpose outside the labor market is perhaps a world that we ought to, you know, welcome. And you know, there are lots of kind of thinkers and writers who argue exactly that. So there are both kind of, I think, important sort of technical and also important kind of moral reasons to wonder whether the project of redirecting technological progress is as straightforward as some of its advocates suggest today. I think, you know, I think there is some merit to it, but I don't think it's the panacea that many hold it out as.

B (22:10)

Do you think these decisions we make on a societal level will play a large role such that it will show up in the economic metrics and statistics? And what I'm asking here is, for example, do you think we'll simply decide that some jobs that we don't want to automate or substitute some jobs?

A (22:32)

There's two questions here. One is we live at a time when the leaders of all the large AI companies say we are going to build a system within a decade that can outperform human beings at every cognitive task that they do. And there are good reasons not to take those claims entirely at face value, not least the kind of extraordinary financial incentives that these companies have to talk up up the capabilities of their technologies. But that said, there are very few examples of technical problems that we have invested as much finance and as much human capital in the pursuit of as the pursuit of AGI we've currently invested, Stuart Russell, the computer scientist, estimated something like 10 times what we did during the entire Manhattan Project in the pursuit of AGI. It's enormous. And so I think asking the question, what if we succeed? What if we build these systems and what work might remain for human beings to do even in a world where these systems and machines could do everything more productively than us, or at least every economically useful task more productively than us. I think that's, that's an important question, that's a kind of technical question about what work from A kind of technical point of view might remain even in a world in which machines can do everything. And I think the answer to that question isn't obviously nothing. I think there are quite interesting reasons to think that there is still work that human beings will do, even in a world in which machines could do everything better than us. But then there's also the question that you're asking, which is more of might there be less from a kind of technical point of view and more from a sort of moral point of view. Might there be certain roles, certain jobs that we want to protect from automation? That it might be these technologies could do them, but we collectively decide that they shouldn't do them from a moral point of view? And I think there are also some of those as well.

B (24:49)

Yeah, actually, the first point, it sounds like a contradiction to say that AIs will be able to do everything, but there will still be work for humans to do. What could be some reasons for that?

A (25:00)

There's a few reasons. I mean, one is the sort of economic reasoning that comes from the sort of Internet world of international trade, which is the kind of just comparative advantage that if you think about what happens when countries trade. If you think about a simplified story with, say, just the US and Vietnam and imagine there are just two goods that are produced, robots and rice. The United States, in theory, has the sort of absolute advantage in both of those things. It could produce robots more productively than Vietnam, and it could also produce rice more productively than Vietnam. But it doesn't make sense for the United States, from a kind of efficiency point of view, to do everything. It makes sense for the countries not to follow their absolute advantages, in which the US has an absolute advantage in everything, but to follow their comparative advantage. What are they relatively better at? And the US is likely to be relatively better at producing robots than rice, and Vietnam's likely to be relatively better at producing rice than robots. And so the countries specialize in producing one of the things, and then they trade in exchange. And that way the kind of collective pie is greater. And the logic of comparative advantage applies equally well, it seems to me, not when we're thinking about thinking about the US and Vietnam, but when we're thinking about people and AI. Even in a world in which AI could do everything more productively than human beings, it doesn't necessarily make sense from an efficiency point of view for AI to do everything. Human beings are a productive economic resource. They should do what's in their comparative advantage. And similarly, AI should do what's in its comparative Advantage the. Now, that's not to say, though it's one thing to say that there might be tasks and activities in which labor retains a comparative advantage. It's another thing to say that there's going to be enough demand for those residual activities to keep everyone in well paid work doing only those. So those are kind of two different observations, but at least from a. There are economic reasons to think that even in a world in which these systems and machines can do everything, labor will still have a kind of comparative advantage in certain things. Now, as these systems become just relentlessly more capable, that comparative advantage might shrivel and the demand for those residual activities might shrink even further. Again, it's not a necessity that it means there's going to be enough work, but that is a possibility. But then there are also, I think, alongside these sort of economic thoughts around comparative advantage, there are also sort of, I think, reasons to think that there are certain tasks and activities that we have a kind of a taste or a preference for how those tasks and activities are performed, not simply by how well they're performed. In other words, the process, not just the outcome.

B (28:21)

Yeah. So this is what you talk about and what will remain for people to do a recent paper and you separate these preference motives into aesthetic achievement and empathy. Maybe we could talk about those in turn.

A (28:35)

Yeah, sure. I should say just this project, I think just to frame the kind of project, you know, I don't think, you know, the challenge for now in the world of work because of AI is not that there aren't enough jobs for people to do. The challenge is that there is work. And for various reasons, people can't do that work. And I call this sort of frictional technological unemployment. And, you know, there's various reasons for it. You know, people might not have the right skills to do the work that's available. They might not live in the right place that work has been created. They might have a particular conception of themselves that's at odds with the available work and want to stay out of work to protect that identity. I think those are the sorts of challenges that we face now. But I do think in light of the kind of technological developments that are taking place and in light of what those at the vanguard are telling us, we also need to ask the question, what if they succeed? What if we build AGI? What does that mean for the world of work? And that's where this paper sits. This paper kind of, and it was published by Columbia a few weeks ago. This paper is in that setting. It sort of takes the Premise of AGI as given and then says, what will remain for human beings to do? And yeah, so there's these kind of economic reasons around comparative advantage, but there's then also these sort of preference reasons, preference limits. And these are all different cases in which we seem to value how a particular task is done, not simply how well it's done. So the aesthetic limits are really interesting. When you walk into the Sistine Chapel and look at the ceiling, you think, gosh, isn't that beautiful? But you also think, isn't it remarkable that a human being did that? In other words, we value the fact that that kind of painting was done by a human being, not simply that it's extraordinarily beautiful. And you might say that so long as we value the process through which a task is done for these sorts of aesthetic reasons, then by definition, these are the sorts of things that these technologies will struggle to do because they are not human. And there's more prosaic examples of it. We value a, you know, the fact that a suit is, you know, handcrafted by a tailor, or we value the fact a, you know, chocolate is, you know, hand molded by a human artisan chocolatier or, you know, more. More prosaically. I love the. There's a great story about a Michelin star restaurant in the UK that sold automated coffee capsules as their coffee. I don't know what machine it was, but it was selling automated coffee capsules for their coffee and without telling the customers in the restaurant. And when the customers found out in the restaurant, they were completely furious. And the reason for their fury is interesting, which is in blind taste tests, people often struggle to tell the difference between a sort of automated coffee and one handcrafted by, you know, sort of great barista. But what they were complaining about wasn't actually the taste of the coffee. It was the fact that they were paying, they thought, for a kind of Michelin star process. They wanted the kind of the artist craftsman making their beautiful cup of coffee and they instead were getting a machine.

B (32:00)

Yeah. I mean, how should we think about this then? Is this simply an instance of the labor theory of value kind of reemerging, where people think that the labor that went into creating something is the value of that thing? Or is there something deeper here? Is it that the labor that went into something becomes in some sense, part of the story, part of the product, part of what you're buying?

A (32:23)

Yeah, I don't think we need to go Marxist on it. I just think it's. People have Taste and preferences, not simply for outcomes, but also for processes. Let me give you an example. So that's this kind of aesthetic one. There's also, I think, interesting reasons of achievement as well. Achievement limits too. You know, anybody interested in AI will have followed over the decades the sort of progress in chess playing machines. And today the very best chess machines could beat the very best human beings at a game of chess. And yet when the very best human beings sit down to play each other at chess, there is huge demand to watch them duel. And the reason is because we like to see, you know, we don't simply care about the kind of efficiency in which the pieces are moved around and on the board. We also care about who's doing the moving or what's doing the moving. We like to see human beings outperforming relative to some sort of standard benchmark of achievement. And so there are lots of areas of our lives where again, for reasons of achievement, we value not simply how efficiently a task is done, but also how it's done. And that is an interesting role, I think, for human beings as well. There's also a kind of, I think a final type of sort of preference limit, which are the kind of sort of empathetic or emotional ones actually, before.

B (34:05)

We get to the empathetic one, the achievement one. So which jobs do you imagine would exist in the future based on our preference for human achievement? Is it that because we can't all be Magnus Carlsen or we can't all be Eugen Bolt or someone who kind of pushes the limits of what humans can achieve. So how will this give rise to jobs?

A (34:29)

Yeah, I think you're exactly right. And again, I don't want to slide from making the claim that there are some tasks and activities that will remain for people to do. The narrow claim to a kind of broader claim, which is that there's going to be enough demand for those tasks and activities to provide everyone who wants it with well paid work? Those are two very different observations. And just more generally, I think it's a mistake that people make when they think about the kind of longer term future of work, which is it's one thing to say this is a task that human beings will always do. And it's another thing to say there's going to be enough demand for that task to provide everyone with a job doing it. I think anything involving kind of degree of competition, degree of sport, kind of degree of kind of rivalrous interaction, these sorts of reasons might, you know, bear out whether it's kind of competition on the sports field, whether it's kind of intellectual competition, whether it's, you know, any, anything that's about again, you know, achievement relative to some kind of standard human benchmark. But by definition, you know, that's exclusive because it's, it's, you know, you are saying you are, you know, valuing the exceptional rather relative to the average. And in a world in which everyone's exceptional, no one's exceptional. So, you know, it's not a sort of, it's not something that is going to necessarily provide everyone with well paid work. Where there's a kind of broader possibility I think is around the emotional or empathetic sort of aspect where we have a taste for a human being for sort of emotional reasons, say end of life care. It matters. The very fact that a human being is sitting there in those last moments of your life is the thing that matters, that it's a human being, a fellow person sitting with you, understanding your thoughts, feeling your feelings. Again, you might say that if that's the thing that you value, then it's difficult to see how a machine could ever do that. I do think, though, in all these different cases there are limits to the limits. And this is another thing that I try and do in the paper. Ask how robust are these limits to these systems and machines just becoming gradually but relentlessly more capable? Just for instance, think about the aesthetic limits. It might be the case that these systems in the future are able to compose music that is so extraordinarily moving or a painting that you see it and you sort of burst into emotional fervor or a piece of text that somehow captures just so perfectly something you were thinking or feeling. It's possible that these systems could achieve aesthetic outcomes which just so dwarf what human beings are capable of doing. That actually our attachment to the fact that a human being was responsible for kind of an aesthetic, an aesthetic outcome might just seem tied and antiquated and you don't have to kind of be sort of high and mighty about it. Think about a tailored suit. Yes, it's lovely to have a suit handcrafted by a human being. But if it's kind of one arm is shorter than the other and it pulls in the back and, you know, it's a bit tight on the bum and, and actually a kind of automate, you know, the kind of sort of, the sort of automated suit is just so extraordinarily comfy, then our attachment to the kind of human craft might, might fade away. So, so I think there's limits to these limits. As well. And one of the things I spend quite a lot of time doing is thinking about what these limits to the limits might be. But there's also a third category, it's worth saying. There's not simply. There's the sort of general equilibrium, these sort of limits, these limits due to the kind of, this idea of comparative advantage, that even in a world in which machines could do everything more productively, it still might make sense to employ labour doing what's in their comparative advantage. There's then these sorts of preference limits where people have a kind of taste or a preference for how a particular task or activity is done. And I think there's various ways that might work. But then there are also moral limits where it's not simply that we have a kind of taste or a preference for a human being performing a particular task or activity, but we believe that a human being ought to do that task or activity from a moral point of view. That even if a system or machine could do the task or activity, they shouldn't do it. That there's something important about a human being being involved in the task or activity.

B (39:45)

And an example here might be a judge, for example, so deciding in a legal case, or we could imagine being a parole officer or perhaps being the person responsible for making sure that military operations are in compliance with international law.

A (40:04)

Exactly right. I think there are lots of kind of micro examples of tasks or activities that have this kind of moral flavor to them where we want to keep a human in the loop, essentially. The. There's also. So there's all these kind of, you know, there's all these kind of task specific sort of moral limits. There's also, though, the broader issue of AI alignment and you know, the, the, the kind of, the sort of, the, the values that sort of direct the AIs that we're building more generally. You might also think that that is an activity that human beings ought to be involved in as well. There are moves, of course, to automate aspects of alignment, to remove human beings from those sorts of judgments. But you might say that's a mistake. But again, I think there are limits even to those moral limits. And the legal one is very interesting. It might be the case that an automated judge is able to reach such a sort of well honed, sophisticated, well designed piece of legal judgment that it becomes morally sort of indefensible to use the sort of flawed human alternative. The human judge who gets hungry before lunch famously and changes their sentencing decisions is stricter or on the battlefield, it might become morally questionable to in the sort of increasingly dynamic and quick nature of conflict, the fact that decisions need to be made so quickly all the time, given the sorts of technologies that are being used, the idea that you might send decision down the chain of command to a human being and then up again, perhaps in the process losing an edge, losing a person, I don't know, that might seem morally. I think there are kind of interesting limits to even the moral limits. And the argument I make in the paper is that what you think about whether there are limits to those moral limits depends in part upon whether you have a moral theory in mind that is sort of process based or outcome based. Is when you think about the role of AI in the legal system and whether it is morally acceptable to use an AI or not, is it the case that your moral theory is only appealing to outcomes? In other words, does this system do a better job or a more effective job of making a sentencing decision? Well, if that system, if that's the nature of your moral theory, that it's purely based on outcomes, then clearly there are limits to those moral limits. Because there might come a time when these systems are just so extraordinarily capable that the outcomes they deliver are better in some sense than the outcomes that a human legal reasoning is able to reach. And so there are limits to that moral objection. But if your moral theory appeals to the process in some way, then that might put a break on some of those limits. And in the extreme, if your view is that it is important for a human being to be making these sentencing decisions, however good an automated alternative comes, that there is something fundamentally important about say, human beings making, you know, only a fellow human being ought to be able to make the decision whether or not to, you know, lock someone up for the rest of their life. If that's your view and it's kind of independent of outcomes. If, if you're just completely attached to process, well then maybe you're, maybe your sort of, maybe the moral limit is robust there because there is. No matter how good outcomes become, your attachment to the human process is going to stand in the way. Now I wonder for those who do hold those sorts of pure process based moral theories for thinking about the impact of the sort of moral limits of AI, whether as these technologies just become more and more capable and the outcomes just become, that they are able to deliver, just become better and better in lots of different domains, whether or not people's attachment to kind of pure process based moral reasoning is going to, to hold up.

B (44:58)

Yeah, it does seem to me that we will probably react to commercial pressures and competitive pressures from, from governments or between governments and between companies and so on. We'll make it so that they're, they're strong, they're strong arguments. They will, you know, we will see strong arguments for implementing these systems even in the cases where we might imagine that there are moral limits. Um, but in the process of doing so and in the process of kind of accepting these arguments, we will, we will disempower ourselves.

A (45:30)

Right.

B (45:31)

If we take ourselves out of the loop in critical decisions. That's, that's something that's difficult to take back just because if you, if you have the automated judge or the automated military system, then, then you are, then you're competitive. And what's the, what's the reason for, for kind of reintroducing humans back into that process? But in some sense we need, perhaps we need the moral arguments, given that we might be entering a world in which AIs are simply better than humans. Do you think that in the end it is the moral arguments that will make the biggest difference to what we end up doing?

A (46:14)

I don't think even in the end. I think already today those moral arguments, the kind of moral limits to automation are almost more important than the sort of technical limits. There are many things that we could use these technologies to do that we don't do. Not from a kind of technical point of view, but often just from a moral point of view, even not also just a sort of cultural point. We just feel uncomfortable. Even without appealing to some sophisticated piece of moral reasoning, we just feel kind of socially, culturally uncomfortable using these technologies in certain settings. And that feels to me like one of the, you know, we spend a lot of time, and as economists, I spend a lot of time thinking about what the technical limits to these technologies are, what they can and cannot do from a sort of technical point of view. But I think these sort of, these sort of moral, cultural, social constraints on technology already bind us a great deal.

B (47:07)

Yeah. My son is in daycare right now and he's going to be 20 years old in 2044.

A (47:16)

How old is he now?

B (47:17)

He's, he's 15 months old. Okay. And, and so when I, when I think of, of the pace of AI progress, when I think of how much better these models have gotten over the last five years say, it makes me wonder what will be left for him to do. What do you think? What do you think the role of 20 year old people in 2044 is going to be?

A (47:42)

I think one of the, Having spent The last decade and a half or so, observing and writing and thinking about the impact of technology on work. I think one of the biggest mistakes that we have made collectively is to think that we are clever enough to say, to sort of predict which jobs are going to have to be done and as a result, what skills and capabilities are going to be most valuable in the future. And there are so many examples of this. There's kind of contemporary examples, but there's also historical examples. Who would have imagined you'd gone back to the 19, the sort of late 18th century and at the start of the Industrial Revolution and sort of whispered to somebody that in a few hundred years time a national health service in Britain is going to employ more people than there are sort of men working on farms in sort of Britain. People. It just wouldn't make sense. There barely was healthcare in the sort of spirit that we have it today. And, and there certainly wasn't kind of public provision, the NHS sort of fifth largest employer in the world or something like that, you know, just kind of. It just would have been unimaginable, you know, the way in which life transformed in the centuries since, you know, the rise of health care, the rise of leisure, you know, just incomprehensible. But. But again, you know, you don't have to go back to the Industrial Revolution think, you know, start of the Internet era. If you had whispered to somebody that in 10 years time people would be finding work as a search engine optimizer, just wouldn't have meant anything. Or even 2019, if you had said, you're going to grow up and be a prompt maximizer, just wouldn't have meant anything. Because the technologies that transformed our lives and the way in which our economies changed, it wasn't simply hard to imagine. It was almost unimaginable. We just didn't have the concepts. And I think this is just as true today, that if we think about the future and given the pace that you're talking about, pace of change that you're talking about, the idea that we can predict jobs in a couple of decades time just seems to me to be incredibly hubristic. And so the kind of principle, one of the kind of running themes of my new book is just the immense uncertainty that we face. And that is the challenge that we have to find a way to respond to the uncertainty. It's just we are setting young people up to fail. If we say, these are the jobs that are going to be available for you to do, and these are the skills and capabilities that you must Learn in order to do them.

B (50:38)

And then perhaps one suggestion that will appear in people's mind is that we should educate our kids in a more general way sense. We should teach them how to learn things. We should give them general, train their general reasoning skills. We should give them very general skills that can then help them adapt to many different future states of the world. Do you think that's plausible or is this a form of kind of cope that we will be able to adapt?

A (51:11)

I think the most important thing is that we teach people how to use AI effectively. I think we need to be spending something like a third of our time in school and university learning how to use AI effectively. And I think that is the most important thing that we can do. And when I say use AI effectively, I don't simply mean how to write prompts and, you know, get the systems, although I think that's important. But I also think it's important that, you know, we teach people, you know, the history of these technologies, where they came from, the way in which.

B (51:52)

We.

A (51:52)

Need to think about problems in order to use them effectively. The sort of the limits of these systems, the technical limits, you know, the fact they hallucinate and, you know, I expect they're going to hallucinate for some time. The fact they make mistakes, you know, that we're able to. And in the event they make mistakes, we're able to understand why and interrogate and then also the kind of ethical and moral issues around their use. I think there is a whole AI curriculum and it's a really exciting project that we need to be writing together, crafting now. And I think that's one of the most important things we could do. And it's not something we can sort of just tack onto the existing curriculum. It needs to be fundamental and it needs to be substantial. And so I think, I think something like a third. And you know, that's not just plucked out of thin air, if I think about. So I spent a long time teaching mathematics and economics to undergraduates at Oxford when I was a fellow at Badiel College at Oxford for some time. And I taught economics across lots of different subjects there. So I taught it to people studying politics, philosophy and economics or history and economics or economics, management. But what I did as soon as the. And I was teaching the college, so I sort of had every year a kind of group of 20 students and small group. But what I did was in that first year, I took a third of their time when they were learning economics just to teach them mathematical methods so that they could then take these tools that they had learned and a big chunk of their time was spent learning it. But take these tools that they learn into all the different domains in economics that they'd then be working, whether it was labor economics or industrial economics, or macroeconomics or micro, whatever it was. But there was a sense in which these sorts of mathematical skills were fundamental and they needed to spend right at the start a big chunk of time learning, to use, learning them so they can then deploy them in all these other settings. And I think that is exactly how I think about AI today. That it's a technology that we need to be asking how we use it in every discipline. And that requires a big chunk of students. Time is spent dedicated to learning about how to use these technologies effectively.

B (54:25)

And that's one path kind of leaning more into AI. Some opposition I've heard to that is that you're a professor yourself and you must have encountered some homework that seemed to you AI generated and then perhaps wondered whether your students are actually learning anything or whether they are simply using AI to get through the homework, to write the essay, to solve the math, and just handing it in without learning much. And so another direction you could go in is to become almost Luddite with regards to AI and go back to pen and paper in order to ensure that students are actually learning something. And perhaps then later, when they have basic skills, when they know when they are good at reasoning, writing, speaking and doing math, then reintroduce or then teaching them how to use AI.

A (55:23)

I think what you're touching on is the kind of one of the fundamental challenges that we face. And it really is one of the big problems that I'm setting out to solve in the new book what Should My Children Do. I think the challenge for all educators in response is not to go backwards, but to go forwards, to ask, okay, well then how do I make what I teach deeper? How do I make it harder? How do I enable these kids to use these technologies to understand ideas, to solve problems, to make discoveries that would have been unimaginable before these technologies came around. That seems to me to be what we ought to be doing, rather than clinging to old fashioned, traditional ways of teaching and educating. So I think the challenge is how we make as teachers and educators, how we make the substance of what we're teaching harder, more challenging. The ball is in our collective court because it's just, I think, unrealistic. It's both unrealistic to think that everyone isn't going to be using these technologies in Years to come and more kind of practically. It's also just. It's unfair because we are encouraging, we are teaching young people in an environment which simply does not. If we sort of strip it of technology, we're teaching them in an environment to be. Does not reflect the world that they're going to enter into. We are setting them up to fail. I think we've got to be asking how we can make it harder, more challenging, more difficult.

B (57:13)

Do we risk losing some of the weaker students if we do that? So if you're faced with a task that is difficult enough for the average student to be able to solve it using AI, is that a risk for the weaker students?

A (57:31)

No, I don't think so. I mean, no more so than.

B (57:36)

The.

A (57:36)

Challenge of weaker students in a world before AI. On the contrary, I think one of the kind of promises of these technologies is that they are able to tailor what is being taught, how it's being taught, to the kind of particular strengths, the weaknesses of different students. I mean, one of the big challenges of the traditional educational model is that it's not particularly tailored. We know that one to one tuition with a human being is incredibly effective, but we were lucky to offer it at Oxford in the tutorial system, and I saw how effective it can be, extraordinarily effective. But most institutions can't afford to have one tutor per one or two or three students. And, and the promise of these technologies is that it can replicate the kind of interaction that you might have with a human tutor, but do so at a far lower cost. And so that's one of the things I think that's very exciting. And you know, people talk about personalized learning and, you know, and things like that, but we really kind of barely scratched the surface and there's a lot more for us to do and explore.

B (58:46)

Do you think there are lessons for parents here about how to. Whether or not to kind of pace your child in a certain direction? For example, ten years ago it seemed like programming, learning how to program, that was the perfect kind of path forward.

A (59:05)

And now it turns out that it's exactly what these systems are good at doing.

B (59:08)

Yeah. Now it's unclear because reasoning models are exceptionally good at programming, at mathematics and so on. So you spoke about the uncertainty of the future and how should parents react to that?

A (59:25)

I think the biggest risk among parents is the legitimate concerns about the impacts of smartphones and social media on young people. And I think there are real issues around social media in particular and the mental health of young people. But what I really worry about is that legitimate concerns about those technologies kind of seep into how parents think about AI as well. AI is not the same thing as social media. And if we bundle technology into a kind of monolithic, indivisible lump of bad stuff, parents are going to let down their kids in preparing them to use these technologies. So that is the kind of biggest warning I have for parents at the moment, which is I share many of the concerns that are out there at the moment about smartphones and social media. But AI is a very different beast, and it's a mistake to conflate them and to conflate the kind of concerns we have about the former with the kind of extraordinary opportunities with the latter. And that's part of the sort of the philosophy of the book.

B (60:37)

As a final question, here there are groups of people, and here thinking about children, perhaps the elderly and so on, that have more trouble than other people, kind of adapting to change. Kids tend to, like, structure. Elderly people tend to dislike change, and so on. How do you think about those groups in society when you're thinking about a more radically uncertain world, a world that's changing faster? What should we do on a societal level to help those groups thrive?

A (61:11)

The new book is focused exactly on that first group, because I think you're exactly right. I think we live in an age of anxiety where almost every day we are told stories of the sort of existential challenges that we face and kind of reminded about our incapacity for dealing with them. And I think there is a hope that has driven many parents on before me, that if they work hard and if they love their families and look after their families, their children's future is going to be better than their past. There's a sense that. And yet I think now, for the first time in some time, there's uncertainty about that. I think many parents are not sure that if they keep their head down and work hard and look after their families, that their children's future is going to be better than their own. And that's quite a dangerous thing when people lose faith in the future. And that is, I think, a situation that many parents find themselves in. And one of the reasons I'm, in contrast, hopeful about the future is because of the possibilities of AI and the possibilities of AI, particularly in the educational setting. I think we can, if we get it right, use it to do really extraordinary things. I think the kind of traditional education system is broken. Not enough people have access to a good enough education. Not enough people are adequately prepared for the world that exists beyond the kind of artificial environment of school and university and the possibilities of these technologies, both in thinking about what we teach and how we teach and when we teach, are really extraordinary. And so that's why I'm spending a lot of my time at the moment trying to gather all these thoughts together so that, exactly as you say, we can help this group of people, young people, really flourish in the world to come.

B (63:32)

Great. Daniel, thanks for chatting with me. It's been a real pleasure.

A (63:35)

Such a pleasure. Thanks so much, Gus.