The Economist Who Believes AI Will Be Great for the Middle Class

25:21Well, it's not for the frontier experts. They already know how to do these things. Nor is it necessarily for the rank amateur. You don't want to go to YouTube and say, well, how do I install and wire in a brand new central house air conditioning? I've never done anything like that before. right? If you went to YouTube for that, you would quickly get yourself into trouble because if you don't have some foundational skills, that could be a problem. On the other hand, if you were handy and you had some experience with electrical work, some experience with plumbing, some experience with carpentry, but you've never done an AC installation before, well, now you could go to YouTube and that would get you further.

25:56So you could think of YouTube as kind of like a mini AI that provides guidance in guardrails. I feel like Tracy has watched many YouTubes in the last year to fix your Connecticut house. This example hits home so hard. And I'll give you a specific anecdote, which is my husband and I are currently building a shed and we're trying to put a roof on it. And we thought like, OK, we put the plywood on the roof and then we get some joist tape. We put the joist tape down over the edges and then we put on the shingles. And we watched many, many YouTube videos on how to do this. Turns out that you can't use joist tape when it's less than 50 degrees Fahrenheit outside.

26:37A problem in Connecticut. Which it was, which of course, none of the YouTube videos that are filmed down in Florida or wherever actually mentioned. And then secondly, it turns out that the ability of the joist tape to actually adhere to the plywood varies enormously depending on what plywood you're using. So there are all these subtleties and nuances that you don't necessarily get from a 10-minute YouTube video. Maybe that's not that surprising. But on this note, so you mentioned training and you've spoken a lot at this point about the idea of AI being able to provide guardrails and context around decision making that maybe can resolve the bottleneck of expensive decision makers, as you put it, by creating more of them or allowing more people to tap that function.

27:23I guess my big question is, how much of this is just going to be, well, we add a new layer of training that people have to do? So you can use AI, but you still have to know how to use AI. You still have to understand the result that it's spitting out and interpret that. You still have to know how to actually apply and use that result. Are we basically just replacing one skill set with another? It's a good question. We want it to require skills, right? If everyone is expert, no one is expert, right? It's important. The question is whether it can speed the acquisition of expertise or whether it just gets in the way, another thing you have to certify on.

28:08We now have a bunch of evidence on AI in specific applications and where it works well and where it doesn't. So, for example, some students of mine, Shkhead Noy and Whitney Zhang, published a paper in Science last year, where they gave ChatGPT three and a half to people who were doing advertising, writing, and marketing plans. And these were people who were college graduates who do this for a living. And one group just used the standard tools, which is basically the internet, word processors. Another one actually used the chatbot. And this was early enough that most people didn't already have it.

28:38And there were a couple of really nice results. So first thing, it saved everybody time. It cut the time. It took people to do this work from about 30 minutes to about 18. The second is it improved the quality on average. So the output of the people using this tool was judged by other college graduates who were not confederates in the experiment to be more precise, more concise, and more accurate. So improve the quality of work and save time. But then the most exciting result was if you looked at the quality range of the work people did, it basically made the least capable writers using ChatGPT were about as good as the median writers not using it.

29:16So it kind of leveled up the bottom. And we've seen this in other places as well, folks doing customer support. The example I'm thinking of is a kind of a enterprise software product and the customers chat in through a chat window. And then the company installed a tool that suggests responses to the customer's chat. You don't have to use them, but it will also not just suggest technical responses, but polite responses and so on to keep the customer from getting overheated. And the result is that it speeds the rate at which people learn. So it used to take people 10 months to reach peak capacity.

29:53Now it takes them about three months. They're somewhat faster when that's done. So it's not that it eliminates the training or learning. Everyone starts off bad at this job, but they get faster. They converge towards expert level more quickly with this tool. And also really interestingly, people quit a lot less. And the reason is, you know, customer service work is actually really difficult. It's very heavy emotional labor. And you have to take a lot of incoming abuse actually from customers. It's hard to keep your cool. And the sentiment analysis of this tool, of the chats that occurred through it, is that it basically reduced the level of hostility from customers to workers and from workers to customers.

30:32So it actually did a lot of the emotional labor. So it didn't eliminate the need for skills in doing this work, but it enabled people to become more efficient, more rapidly, with less stress. And so that's the good scenario. There's a lot of work that needs to be done. And right now, what are the most expensive things, the things that are growing more and more costly all the time, are education, health care, legal services. Why is that? Why are those things getting so expensive? Well, during the industrial era, we got really efficient at manufacturing goods, right? So TVs, automobiles, coffee makers, right, mobile phones, these things are actually remarkably good and relatively cheap.

31:16Why? Well, we've automated them and the labor content is relatively low. On the other hand, healthcare, education, law, right, we've not gotten any more efficient to those things. And they require people who've gotten more and more expensive over time, because as we've automated the other work, the people who are the degreed professionals or have become the bottleneck. So that slows the growth of productivity. It makes the cost of living higher for the typical person, right? Typical person is not a lawyer. It's not a professor. It's not a doctor, but they're paying for all those things. So if we could enable more people without as much training, and I don't mean no judgment.

31:53I mean some training. If we could allow paralegals do more legal work, if we could allow nurse practitioners to do a larger range of medical tasks, if we could enable people who are working as contractors also to do more design, right? If we're enabling people who don't have computer science degrees to do more software development, not only would that reduce the cost of these expensive services, but improve the quality of work that people could do. It allowed them to take some expertise and make it go further. So that's the good scenario. Joe, I like the idea of using AI to reduce emotional labor.

32:27I wonder if I can start automating some responses on Twitter to toxic Bitcoin maximalists. That's interesting. Tracy, the block button is right there. So there's so many different questions now that I have in my mind. But, you know, look, we're only near the beginning. I mean, ChetGPT, which is sort of what's brought this all into consciousness, was unveiled to the public in late 2022. So not even two years into that. This sort of breakthrough that enabled it is just a few years older than that. The concern would be, well, yes, at this point, some training plus AI enables many people to become much more productive and have the sort of output that was previously associated with people with years of experience.

33:12Like the fear would be that in multiple generations down the road, you don't even need that initial training. First of all, I fully agree. We're just at the beginning. The tools are only so good, they're going to get much better. Our understanding of how to use them is also very primitive. We often don't know how to interact well with AI. In fact, I could give you examples of cases where it goes pretty badly, even though the tool is good. So I think there are sort of two concerns built into what you said. One is that basically for now, it's a helper, and then eventually it's just your replacement.

33:42And the other is that even if it just makes everyone more efficient, eventually, well, we just saturate the world with whatever that thing is, and then it's super cheap, right? So there's only so many PowerPoint presentations the world can tolerate. And if you get really fast at making them, eventually people will pay you to stop. Yes. We're there now maybe. But anyway, keep going. It's quite possible. So I think that will occur in some cases. There's no question that in some cases the tool will initially be a supplement and eventually be a replacement. So maybe air traffic controllers would be an example like that, where eventually almost all the air traffic control will be done by machines.

34:17But I don't think every job is like that. I don't think that's the case in medicine. Medicine will be a hands-on occupation for a very long time. So will law, where there's a lot of high-stakes decision-making. So will design. So I don't think that we're going to automate everything away. I know people think that, and I think it's a valid concern. I don't think that's the most likely scenario. But I also want to stress something that's said too little in these discussions, which is when you think about what you can do with a new tool, most people think, well, what can I automate? what is the thing that I'm doing now that I could now have the machine do for me?

34:52And that's important. And we do a lot of automation, but automation is not the primary source of how innovation improves our lives, right? Many of the things that we do with new tools is create new capabilities that we didn't previously have, right? So airplanes did not automate the way we used to fly. We just didn't fly before we had airplanes, right? The scanning electron microscope didn't automate the way we used to look at subatomic particles, we simply couldn't see them without that microscope, right? So think of the thought experiment of automating everything in ancient Greece, you know, 2000 years ago.

35:29Even if you automated everything in ancient Greece, it wouldn't be modern America, right? It wouldn't have electricity. It wouldn't have computers. It wouldn't have airplanes. It wouldn't have penicillin. It wouldn't have a million tools, technologies that we take for granted, right? So the most important applications of technology are to enable capabilities that didn't previously exist. And I think AI will do that as well. So, you know, we couldn't be having this conversation were it not for our computers, right? If someone took my computer away from me, I couldn't even do my job, right? It just, my job wouldn't exist in its current form.

36:02And so what we do with new technology is create new capabilities. And then human expertise is often needed to support those capabilities, We didn't have pilots before we had airplanes, and we didn't have pediatric oncologists before we had all kinds of tools and knowledge to treat cancer or cancer in children. And so as we instantiate these new capabilities, we often require new human skills and expertise that are valuable. And so much of what we do with these tools is to change our lives by pushing out the possibility set rather than simply just automating the things that we already do. And I think AI will also be really important for that.

37:08I only pay for what I need. Right now, with MyBizPlan, get our best price. As low as$25 a line. Visit verizon.com slash business to get started today. New lines only. Price per month with five plus lines. Includes auto pay and paper free billing and promotional discount. Taxes fees, economic adjustment charge, applicable add-ons, prices, and terms apply. Guarantee applies to base monthly rate and stated discounts only. Add-on prices additional. Offers in January 5th, 2026. Support for the show comes from Public.com. You're thoughtful about where your money goes. You've got your core holdings. Some recurring crypto buys.

37:34Maybe even a few strategic option plays on the side. The point is, you're engaged with your investments, and Public gets that. That's why they built an investing platform for those who take it seriously. On Public, you can put together a multi-asset portfolio for the long haul. Stocks, bonds, options, crypto, it's all there. Plus an industry-leading 3.6 % APY, high-yield cash account. Switch to the platform built for those who take investing seriously. Go to public.com slash market and earn an uncapped 1 % bonus when you transfer your portfolio. That's public.com slash market. Paid for by Public Investing.

38:09All investing involves the risk of loss, including loss of principal. Brokered services for U.S.-listed registered securities, options, and bonds in a self-directed account are offered by Public Investing, Inc., member FINRA, and SIPC. Crypto trading provided by XeroHash. Complete disclosures available at public.com slash disclosures. One thing I wanted to ask you is you are very, very clear in your piece that this is more of an informed thesis than an actual forecast. And here, I am actually leaning on chat GPT when I asked it to poke holes in your argument. One of the ones it spat out had to do with this exact question.

38:46But are there specific measures or policies that we could be doing right now to make the probability of this outcome better rather than the sort of like destructive AI jumerism outcome that everyone is worried about. Yeah, so I appreciate your saying that. The future should not be treated as a forecasting or prediction exercise. It should be treated as a design problem because the future is not like the weather that we just wait and see what happens, right? We're making our own weather. We have enormous control over the future in which we live and depends on the investments and structures that we create today, whether that's democracies, whether that's education, whether that's how we use tools and science, whether we use fissionable material to make bombs or to make energy, right?

39:32We have lots and lots of agency here. So in terms of using AI well, so first of all, let me say, what would be a metric? How would we know we were using AI well? Because it's not like carbon dioxide where, you know, we say, oh, we know we're reducing carbon dioxide. You can just measure it, right? How would we know we're using AI well? I would say we know we're using it well when we see people who don't have four-year college degrees doing work that we would think of as expert decision-making work, whether that's coding, whether that's medical vocational work, whether that's design and contracting, or even whether it allows skilled repair people to work on a broader range of products or tools or engines or whatever.

40:12So that's my metric of success, that it opens up new job opportunities to people who are not at the absolute elite of a field. How do we get there? So I think that's a super central question. And I think most thoughts about, you know, policies about AI are about regulating, controlling. And some of that has to happen. And I feel reasonably confident that it will. This is much more about investing, right? So you could say, look, you know, in the United States, for example, about 20 % of GDP, two in$10 goes to education and healthcare. More than half that money is public money. So in fact, we have a lot of control over how education and healthcare are delivered.

40:54So healthcare would be the best place to start to say, all right, let's redesign the tools or invest in the tools in a way that enables more people to deliver this work. And not only would that make better jobs, it would also improve access to healthcare, potentially lower those costs. We could do the same in education. How can we make education, you know, make better use of teachers, provide better services to students, and also make education more accessible, immersive, engaging for adults, right? We have lots of adults who need to learn, and traditional classrooms are really not the best place to do that.

41:25So I do think you have to think about these moonshots, and governments can invest in them. That doesn't mean the government has to run them, but governments often fund basic science. Governments fund education. Most health innovation in the United States is paid for by the National Institute of Health, which is much, much larger than the National Science Foundation, for example. So I think that is the biggest challenge, is to look for those opportunities and then design with the intention of creating a more effective way to structure work that uses the tools and uses human skills better. And let me say, you might say, well, why doesn't this apply equally well to the last era?

42:03So first of all, we didn't design and probably we should have done more, but essentially computers are good at following rules. And so they could replicate a lot of work that was just that, but they weren't good at supplementing skills, enabling people to do these high stakes decision-making tasks. So it's important to understand AI is almost the inverse of traditional computing, right? If I told you I have the most advanced technology in the world, but it really can't do math and it's not reliable with facts and figures, you would say, well, what kind of technology is that? And I would say, well, that's artificial intelligence.

42:35It is really quite the opposite. So I think it has quite different capabilities. And in some sense, you could say traditional computing was really complementary to the the most elite professionals. And it's quite possible that AI will enable more people to compete with them. And that's a really good thing because that improves the quality of services and improves the quality of jobs for people who are not at that leading edge. This is, I think, the key thing, because in your piece and in other testimony you've given, you've talked about this idea of collective decision-making. And when I think about modern American society or modern society in general, I don't necessarily think that collective decision making is something we're particularly strong on.

43:17So if the future depends on making good collective decisions, then that makes me anxious. But, you know, you talk about investment, but it sounds like the other element here. And you mentioned that the rise of the nurse practitioner had to happen over the kicking and screaming of the American Medical Association, which represents that a top strata of healthcare professionals, the elite doctors. How much of this is going to be a political fight ultimately in which the doctors and the lawyers and the podcasters collectively resist other people who are using these tools to do our jobs? And how much is that really like where the collective fight is going to happen?

44:00Yeah. If we have to take on the podcasters, I think we're doomed. Yeah, we're going to fight this kicking and screaming for sure. The AMA is one thing. Yeah. But the podcasters, that's a whole different army. Some of that will absolutely be turf warfare, right? The professions, we think, oh, you know, that, you know, oil companies and so on don't like competition and they're always trying to rig the market. But in fact, the professions rig the markets as well, right? Yeah. They, what a profession is actually, what it means is an occupation that gets to certify its own members and decide who's in and who's out, right?

44:30And so it's the medical profession that creates training standards and certification standards. It's universities that decide what skills enable you to have a PhD and therefore become a professor. So it absolutely is going to be a challenge. Like lawyers will try very hard to say, well, that can't be a legal document unless a lawyer has signed it, someone with a JD and has passed the bar, right? So that will be a source of resistance for sure. On the other hand, if there's a really good competing alternative, if you can say, look, these nurse practitioners can do a lot of this diagnostic work, you know, they work well with doctors, but they can do some things that doctors would be more expensive doing, and you can make that case.

45:10Or a paralegal using the software can create a lot of routine documents, or a software developer using GitHub Copilot can go pretty far. Then that creates a lot of economic pressure that tends over long periods of time to erode these gills. So I think that they will not go quietly into this dark night. But if the models are successful, it does create a strong incentive for eventually that to become adopted. I think part of the concern around AI has to do also with how any productivity gains are actually distributed and whether or not people are compensated for doing more. and I asked chat GPT obviously to provide a summary of Das Kapital before I came on here.

45:58No, I do think there is this concern about, okay, in an ideal scenario, we're all more efficient in terms of our labor and maybe some types of work are even better to perform. Maybe we reduce that emotional labor. But aside from that particular benefit, how do we distribute the additional productivity gains. Is there any evidence or any reason to believe that these benefits are going to go to labor, to actual workers and individuals versus to companies and capital? Yeah, good. So let me give you two answers to that question. One is it really does depend on institutions, not just on decentralized labor markets, right?

46:38So if you compare, you know, the U.S. versus Germany versus Scandinavia, right, we have so much in common. We have the same technologies, we have the same Asian populations. We have the same rising education levels. We have the same China as a competitor. We have lots of immigration. And yet these countries have baked very different cakes with the same ingredients, right? The U.S. is kind of cowboy capitalism, very high levels of inequality and disparity, and not so much sharing with workers. And if you look at Scandinavia or Germany, it's much more cuddly capitalism, right? It's not nearly as unequal.

47:06And that's really a question of tax regulation. It's a question of the role of labor unions and labor voice. and it's a question of social norms. And so I guess we should not take it as inevitable that the outcomes we have are the only ones the market could tolerate. But at the same time, we should recognize that without those sort of countervailing forces, the outcomes can look pretty bad, right? So I do think, you know, I'm happy about the rise of collective bargaining again in the United States, although it's from a very low level. I'm happy that more states are passing minimum wage regulations.

47:41I'm happy that the Biden administration is trying to sort of beef up the Occupational Safety and Health Administration and the Equal Employment Opportunity Commission and so on. So I think those things matter a great deal. So one should not take it for granted that just because productivity rises, workers benefit. In many countries, that's true, but not so much in the United States. I want to press you right here on this point, because why doesn't this undermine much of the argument? If these different countries, whether it's Sweden, Germany, the U.S., can have very different sort of distributional outcomes with the same cake ingredients, with roughly similar technology in labor markets.

48:18Why then take the assumption that it's the technology that has the distributional impact rather than just those policies themselves? Okay, that's an excellent question. So I think the technology provides headwinds and tailwinds with which policy can work. So all of these countries I mentioned have become more unequal. Okay. All of these countries have seen a decline in middle-skill work. All of these countries have seen the mean wage rise relative to the median, meaning the upper wages have risen more than the center. But the degree to which countries have pushed back against that is a function of their institutions.

48:53In the prior era, prior to computerization, all of these countries saw their middle classes grow together along with the upper class and lower class. And so the industrial era prior to computerization was very friendly, sort of intrinsically towards the middle class. The computer era was much, much less so. And then policy helped ameliorate those impacts and much less so in the United States. So I do think the technology plays a role. I just, we should simultaneously believe that these underlying forces of technology and globalization create strong pressures in one way or another, and then policy can shape how those pressures play out.

49:27It won't undo them, but it can channel them more or less effectively. So you're asking both the right questions. And I think the answer is both are true, but we should think it's not one or the other. And in some periods, those forces are very favorable and policy has to do less hard work. In other periods, they're relatively unfavorable and policy, if it's working well, has to do more work. The other point I want to make, and this is why I'm so focused on expertise, is expert work is intrinsically well paid. It's scarce and it's necessary. And that's why if we live in a world where all the work can be done by machines, we're completely dependent upon redistribution, right?

50:05The people who own the machines to share with everyone else. And I'm not so optimistic about people's excitement about sharing with everyone else. And even when people say, oh, we'll have universal basic income, they really mean universal basic income within the borders of the United States. They don't mean universal basic income for the rest of the world, right? So really, people's notion of sharing is very limited. So I do think it's extremely important that labor remains valuable. And that's actually an achievement of the industrialized world, that so many people can make a good, reasonable standard of living based on their skills.

50:40And so technologies and tools that make human expertise more valuable by allowing to go further are really favorable towards income distribution. Technologies that just automate away work, even though they raise productivity, are not favorable towards income distribution because it means it goes to ownership of capital. And ownership of capital is intrinsically more centralized than ownership of labor because in a country that doesn't have slavery and doesn't have labor coercion, everyone owns one worker themselves. And so that inherently creates some tendency towards equality when labor is valuable.

51:17The efforts of the Biden administration to re-industrialize the U.S. and sort of counter some of the effects of the last 20 years that you wrote about, Do you have any optimism that those trends can be reversed? I know this is a very simple, straightforward question that you're going to answer in about 30 seconds. So good luck. I don't think they can be completely reversed, but you can stem the tide, right? So it's not that this is stabilized. The U.S. continues to lose industrial capacity, right? Whether it's in semiconductors, whether it's automobiles, whether it's in aircraft, thank you, Boeing, and so on.

51:48So I think reinvesting can help solidify those sectors. and I think it's very important to do so because now they're not just a question of jobs. It really is about leadership of the key profit and idea generating activities in the modern world and we don't want to lose a leadership place in those activities. Good, concise answer to what probably could be multiple future episodes. David Autor, thank you so much for coming on Outlots. That really was a fascinating conversation. We probably could get multiple episodes out of this conversation with you, but really appreciate your time. Thank you very much.

52:26Nice to speak with both of you. Have a good day.

52:39Tracy, I'm convinced. I think everything will be fine. I'm no longer worried. Well, first of all, I would say it was nice to hear a slightly more optimistic argument from David. there were a lot of quotable sentences in there. So I like the idea that everyone's their own individual capitalist in the sense that we each have one worker to direct and get the most money out of. So that's how I'm going to start thinking. Cuddly capitalism. Oh, cuddly capitalism, which, as our producer Kale observes, is a much more appealing name than the Swedish model. I like that. What I would say is, again, putting on my cynical journalist hat, and I guess I don't have an opinion because I am a journalist whose expertise is about to be automated away.

53:24But my non-consensus take here or my sort of hot take here is that I agree with David that we are going to get more jobs out of AI and probably more than a lot of people currently anticipate. I guess I'm less convinced about how useful those jobs are going to be. So going back to his point about how do we measure how well we're using AI, I have a feeling that a lot of it is going to end up basically creating a whole new layer of BS jobs that don't actually do much. So there's going to be all these decision making bodies attached to AI. There's going to be big discussions about how you implement AI, fairness, litigating its results, and things like that.

54:09I guess I'm a little bit pessimistic about the ability of AI to generate additional bureaucracy in addition to additional productivity. The other term that was great was when he said the future is not like the weather. Yeah. But also, like, I am worried about any notion that to achieve the good outcome, the good equilibrium, we have to make good, correct, collective decisions. Yes. Because I have almost zero confidence in whether it's just the U.S. specifically or globally to make collective decisions. I do think, like, going after, like, these guilds, like the American Medical Association, which for all of the rise of nurse practitioners, it doesn't seem like we're doing great on, like, bending the cost of healthcare jobs or really having a healthcare capacity, that's going to be like really tough.

54:52And those fights are going to be really intense, whether it's with lawyers, whether it's with doctors, whether it's with teachers, whether it's with podcasters, whether it's professional architects, etc. Like those fights are going to be extremely intense. But like the basic intuition sounds very compelling to me. The other thing is like, you know, this idea of like, oh, yeah, you have some training plus AI, like, I am worried, Like maybe you just won't need the training and maybe it's just AI from the start. So I don't know. Well, in some respects, I think that would almost be a better outcome in terms of democratizing AI.

55:27But yeah, there are so many questions. Uncertainty, as you mentioned in the intro, lots of different takes at the moment. I guess we'll see how it plays out and whether or not you and I have jobs in 10 years time. We'll see. We'll have David back when we're just like... When we're automated voices. Yeah, exactly. All right. Should we leave it there for now? Let's leave it there. Okay. This has been another episode of the Odd Thoughts podcast. I'm Tracy Alloway. You can follow me at Tracy Alloway. And I'm Joe Weisenthal. You can follow me at The Stalwart. Follow our guest, David Otter. He's at David Otter.

55:58Follow our producers, Carmen Rodriguez at Carmen Armin, Dashiell Bennett at Dashbot, and Kale Brooks at Kale Brooks. And thank you to our producer, Moses Andam. For more Odd Lots content, go to Bloomberg.com slash Odd Lots. We have a blog. We post transcripts and we have a weekly newsletter. And you can chat with fellow listeners 24-7 in the Discord, discord.gg slash oddlots. There's even an AI room in there where people are talking about all these things. So imagine there will be some conversation about this there. And if you enjoy Oddlots, if you like it when we do deep dives into AI, how it works, what it means for the economy and society, then please leave us a positive review on your favorite podcast platform.

56:39And remember, if you are a Bloomberg subscriber, you can listen to all of our episodes absolutely ad free. All you need to do is connect your Bloomberg subscription to Apple Podcasts. Thanks for listening.

57:43Transcription by CastingWords bad or just another billionaire. Good bad billionaire from the BBC World Service. Listen on BBC.com or wherever you get your podcasts. Residents at Brightview Senior Living Communities enjoy enhanced possibilities, independence and choice. Brightview, Dulles Corner and Great Falls offer vibrant senior independent living, assisted living and memory care services through various daily programs, chef prepared meals, safety and security, transportation, resort style amenities and high quality care. Take financial possession of your apartment by December 31st, 2025 and save.

58:17Discover more at brightviewseniorliving.com. Equal housing opportunity.