Money Talks: The AI Job Apocalypse is Avoidable

0:04Hello, and welcome to Money Talks, a special extra podcast from Slate Money, where we chat with brilliant and interesting people. I'm Emily Peck. I'm a writer at Axios and co-host of Slate Money. And I'm here today with a brilliant and interesting person, a Nobel Prize winner, Daron Acemoglu. Welcome to Money Talks. Thank you, Emily. It's my true pleasure to be here with you. Darren is, as I said, Nobel Prize winning economist and a professor at MIT. The reason I wanted to talk to you today was because of this paper that we talked about. It was recently, but it feels like a year ago. Time flies by when you're having fun with AI.

0:42About how AI doesn't have to be apocalyptic for workers. It can be positive. So we're going to get into why AI doesn't have to take away all our jobs when we come back on Sleep Money Talks.

1:02Slate Money is sponsored this week by Bill, which is the intelligent finance platform that helps business and accounting firms scale with proven results. Bill is a leading provider of software that simplifies, digitizes, and automates back-office financial processes for accounting firms and small to medium-sized businesses. What does that mean? It means they help businesses pay and get paid by ACH, check, credit card, wire transfers, you name it. It's secure. It has AI-powered accounts payable automation that erases the busy work from capturing invoices, processing payments. It syncs seamlessly with accounting software.

1:39It does your 1099 tax filings, you name it. It's an end-to-end workflow. Its customers come from all industries, ranging from startups to established brands and nonprofits to accounting firms themselves. When business owners choose Bill, they aren't guessing. They're choosing infrastructure that is already proven by millions of users and over a trillion dollars in secure payments. Bill doesn't just move money. It enables the business's entire accounts payables workflow, and it creates the ability for them to control their financial operations securely and at scale. So talk with a payments expert at Bill.com slash proven and get a$250 gift card as a thank you.

2:20That's Bill.com slash proven.

2:27This message is a paid partnership with Apple Card. Apple Card is a no-fee, cash-back, rewards credit card with a ton of great benefits. There's a lot to love, but instead of listing everything, I want to focus on my favorite benefit, travel. You all know how much I love to travel, but even as a seasoned traveler, things can still get stressful, which is why I use Apple Card on my international trips. Apple Card has no foreign transaction fees, so I never have to worry about extra charges when I'm abroad. And with 2 % daily cash back on every purchase with Apple Pay, I'm actually earning daily cash as I travel.

3:07That's 2 % wherever Apple Pay is accepted, which adds up when you're booking flights, hotels, and car rentals. Instead of coming home feeling like I've drained my bank account, I come back with cash back I can put toward my next trip. So start using Apple Card for everyday purchases today. Don't have one yet? Apply in the Wallet app on iPhone. Subject to credit approval, variable APRs to Apple Card range from 17.49 % to 27.74 % based on credit worthiness. Rates as of January 1, 2026. Existing customers can view their variable APR in the Wallet app. Apple Card issued by Goldman Sachs Bank, USA, Salt Lake City Branch, terms and more at apple.co slash applecardbenefits.

3:53Darren, why do you think the notion that AI will replace people's jobs, I think the CEO of Anthropic has said it will create a job apocalypse. Why do you think that narrative has sort of taken hold, not just among the AI industry, but it just seems like we see it everywhere now? Well, I mean, let me first say that that talk isn't idle talk. We are going in that direction. We are moving in the direction of AI replacing people because that's what the tech companies are thinking about. That's what they're targeting. That's what their business model is. That's what their efforts are seeking to achieve.

4:33Our point is that there's an alternative direction, but it's not one that's going to happen by itself. We need to target that in order to get that more pro-worker direction to be a possibility. And the centerpiece of the argument is that artificial intelligence is quite different than human intelligence. And when two things are different, a natural way to combine them is in a complementary way, not try to replace everything that one does with one type of intelligence using the other type. Could you unpack that? I mean, because stepping back, it seems like a lot of the things I'm reading now are like, yes, AI is different from human intelligence because it's, in fact, better than human intelligence.

5:18So it's actually pretty novel to argue, no, it's just, it seems like you're saying between the lines, it's not better, it's just different. So can you sort of, you know, unpack that? How is it different, not just better? Well, if you look at how children learn, you'll see that it's different from AI and not inferior. Children don't have access to a huge training database. Right. Nor do they have billions of tokens in their minds. They cannot keep in their brain training weights and update them. Right. But they're doing a fantastic job of learning. This was in your paper, right? You're like, my child didn't need to see a thousand cats before she understood what a cat was.

6:00It just took one or two, probably just one. And then they're good. How do they do that? Well, they do that, first of all, because our human brains work very differently. And we would be hopeless, absolutely hopeless, at taking something like a medium-sized encyclopedia and find the relevant information in there in one hour or two hours or three days, really. So we're not very good at taking large amounts of information and sifting through it and recognizing patterns, but we are very good at understanding the context, generalizing, developing hypotheses, extrapolating from a small number of things, and then learning, continuing to learn by trial and error, by social learning, meaning other people are doing, being in arguments or being in discussions or imitating other people.

6:51So there are lots of things that humans do that seem to work extremely well in terms of very quick learning. And that kind of quick learning, out-of-the-box thinking, is very important for creativity. The other thing that humans do very well is we are able to combine multiple types of reasoning for judging whether some new ideas are good or just are crazy. whereas the artificial intelligence is not a one-shot learner it's not good at trial and error and it has no judgment so when models hallucinate they hallucinate badly when humans hallucinate that's like when we're being creative we think of new things and then we decide oh no that's not a good direction let me experiment with this one so there are many things in which the two types of learning could be very complementary to each other yeah so okay so now let's talk about the complimentary piece.

7:45Because I have to say, even in just the past few weeks, I've been working more and more with Claude or Anthropic or whatever, and I'm finding it really useful. And I actually, I don't think it could take my job at all. It's just sort of helping me. It's almost functioning for me like when I call someone like you and ask a lot of questions and then go, you know, come back away from the interview and write something. That helped me, but I still need to do that thinking and writing and all that work on my own. I can't just, you know, publish what you say into the world. I mean, maybe I could. Right, great.

8:17And in fact, imagine what you could do if Claude and all of these other large language models weren't designed for automation, but they were in fact explicitly designed for being complementary to you. Right now, they are explicitly designed for automation. It's for two different reasons. One is that they're burning desire is for artificial general intelligence, let alone the fact that AGI, artificial general intelligence, doesn't seem to be well-defined. But there is just this singular obsession with reaching that in all of the leading AI labs. Second, the way that we train these models, we don't train them to work with humans.

9:00We train them to act autonomously. Like the way that they are trained is, okay, here are some questions. Are you going to get the right answer to these questions? How are you going to get scored when we ask questions and test you on these questions? Whereas if they are trying to be useful to humans, the way we should judge them is how can you team up with humans and be useful to them? It's not that we need to act like you are the lawyer passing the bar exam. How can you provide the useful type of information with the right context, with the right reliability to human lawyers? Why the obsession with creating artificial general intelligence and not in creating partners for workers?

9:45There are two broad reasons, and within each one of those broad reasons, we can also dig a little bit deeper. The first one is economic. The main companies at the forefront of AI have business models that are based on automation or things that don't involve human decision-making centrally. They either make money from digital advertising or selling software or software services to large companies so that they can automate some of the tasks in their jobs. So that's the economic reason that they are targeting. that. That's what they know in order to make money. Creating human complementary technologies has been very active during different periods of history, certainly feasible, certainly potentially profitable, but you need to create a different business model.

10:34You need to market them differently. You need to babysit those models differently, et cetera. So that's the first economic reason, and there are additional subcategories there. Our tax code, for example, encourages this because we tax labor, we subsidize capital, so it makes it more profitable for firms for artificial, non-natural reasons that they should use algorithms for automation rather than rely on human workers and algorithms together. But there's also an ideological element. These companies have underestimated human intelligence and have, in my mind, been mesmerized by the potential of science fiction-like machines to take over human tasks.

11:18These two are complementary, synergistic, but distinct. Underestimating human intelligence shows itself in their thinking that you're going to be able to automate wholesale occupation with simple generative AI tool. The statements that there won't be any accountants left, there won't be any paralegals left, there won't be any customer service representative or interpreters and translators or journalists left. Those are all statements about wholesale occupations being within reach of these models. Whereas if you think of what most people do during the course of their occupations, there are so many complex judgment decisions, so many edge cases, so many places where you are routinely using combination of judgment and creativity.

12:05And you're doing the very difficult task of cognitive reasoning interacting with the physical world. All of those are very difficult things for artificial intelligence and very natural for humans. So if we accept that, then it becomes much less likely that we're going to be able to automate wholesale occupations anytime soon. So we would be more in the realization that actually doing something complementary would be useful. Second, you know, from the time of Alan Turing's famous imitation game and all of the science fiction that followed, including all of the AI. Remind us the Turing game. So the Turing thing was, you know, started from the mathematical and philosophical endeavors that Alan Turing was pursuing.

12:53and at some point he started arguing that because the human machine is a kind of computer, once we have a powerful enough computer, that will be able to do everything that the human brain does and then to prove or develop the argument that machines could think like humans. He also devised this imitation game, which then became part of our general culture when it became a movie where the imitation game is, you know, you're testing an AI model or a computer against a human, asking questions. And when you can distinguish that this is a machine, this is a human, then the machine has failed. Whereas the moment when the machine can imitate a human well enough that you cannot distinguish it from a human, that's when the machine has come to the thinking capabilities of humans.

13:44So that became the basics of the Turing test, et cetera. So all of these, together with all of the science fiction that has been produced since then, in which machines quickly come to the same or superior capabilities as humans and then start threatening our civilization, have been the diet upon which many of the people in this industry have grown up. So they are naturally gravitating towards that kind of science fiction, I think, in their aspirations as well. I am. I think about this. I've been thinking about this so much since I first spoke to you. This idea that not only is there's this inspiration from science fiction, but also this idea that a lot of these researchers and tech industry folks are underestimating human intelligence, just like underestimating people's humanity.

14:31And just as someone who's been sort of like reporting and writing on the labor market for so long, it's actually not surprising at all. Because this is how big tech and companies often treat humans as lacking in humanity, underestimating them, treating them often like machines. Like that's not even a big tech thing. It's a vicious circle. The moment you start thinking, and this is true both for tech companies and executives, the moment you start thinking humans are worthless. You start treating them as worthless. Yeah. And then their potential is not realized. And now you're spending hundreds of millions of dollars to create these data centers and driving all this capital investment.

15:13Yes, I'm sorry, billions, maybe even trillions, I think Sam Altman said at one point, in these data centers and building this technology that will maybe, but isn't yet, be as smart as humans. and systematically underinvesting in literal humans all the time in our education and workforce and policy, it's really kind of like, blah. I don't have a specific question for you, but maybe that leads me to what I also wanted to talk about with you, and I'll ask you when we come back.

15:59Slate Money is sponsored this week by Upside. There are certain things that you just have to spend money on, like gas, groceries, probably even eating out. Even on the tightest budget, you can't escape them. But this is one thing you can control. You can get cash back every single time you fill up, hit the grocery store, or go out to eat. If you use the Upside app, you just open up the app, You claim the offer and you pay with your normal credit card. You get all your normal credit card points and rewards. And then on top of that, you get cash back, not points, not credit, real cash, money you can transfer straight into your bank account.

16:35There are over 100 ,000 participating locations. And that means that once you start checking the app, you'll realize how often there's an offer right where you're already going. If you're going to the gas station, you're going to fill up with gas. You may as well open the Upside app first, select that gas station you're going to anyway. And then on top of getting your gas and getting your credit card rewards, you also get cash. Independent studies have confirmed that Upside users earn significantly more than traditional rewards programs. Upside has already given back$1 billion to its users. To find out how much you could earn, download the free Upside app and use promo code SLATEMONEY to get an extra 25 cents back for every gallon on your first tank of gas.

17:17That's an extra 25 cents back for every gallon on your first tank of gas using promo code SlateMoney.

17:47Get an additional$100 off select laundry pairs. The season's best lineup is here at Lowe's. Valid through May 6th while supplies last. Selection varies by location. Mulch offer excludes Alaska and Hawaii. See an associate or lowes.com for details. Amazon Health AI presents Painful Thoughts. I, um, I can't stop scratching my downtown. Yeah, but I'm not itching to go downtown and tell a receptionist I'm here to talk about my downtown. Some things you'd rather type than say out loud. There's no question too embarrassing for Amazon Health AI. Chat your symptoms and get virtual care 24-7. Healthcare just got less painful.

18:37so what i wanted to ask you was you talked a little bit about policy and you talked about the tax code but what are some ways that policymakers could think about encouraging pro-worker artificial intelligence because from what i hear when talking about policymakers and ai all i really hear get talked about is sort of regulation or lately you know i don't even know picking favorite companies or stuff like that. But I don't hear anyone sort of putting forth a positive theory, which is also why I liked what you guys wrote. Thank you. Thanks, Emily. So what are some ways policymakers should be thinking about pro-worker AI?

19:14Well, let me first try to articulate how we could go to pro-worker AI. And my belief, and this is just a belief, is that if we woke up tomorrow and 50, 60, 70 % of the very talented AI engineers and creators working for these leading companies decided what we want is not the super duper large language model that can imitate human speech, but we want to create pro-worker AI tools that expand what humans do, that create new tasks for them, that increase their ability to solve problems and make better judgments and decisions, be more creative. I think we would get there. There is enough talent, enough energy, enough unity in this industry that if they set themselves a feasible goal, and I very much believe, and we can talk about this, why pro-worker AI is very feasible.

20:16It's actually much simpler to achieve than human level or near human level intelligence. In fact, it's within our possibilities, even with the models right now. If they set their mind to doing that, we're going to get there. So the question is, how can we change enough of the focus and prioritization of people in this industry? And there are two big levers for that. One is public pressure. So if the debate shifts, and that's why programs like this are very important, if the debate shifts to saying, no, no, this discussion of whether we are good for humanity because we put some roadblocks against autonomous weapon systems, that's not the main area that we should be focusing on.

21:01We should be focusing on whether you're actually furthering this pro-worker agenda versus trying to just sideline humans. That would really have a big impact because I completely believe that many people within these organizations, especially in companies like Anthropic, really are well-meaning. It's just that their well-meaningness has been squeezed into this narrow space. Well, we're going to AGI, so the only thing we can do is make sure that AGI, when it comes, doesn't kill us all. That's a very, very sad sort of space for us to be altruistic for the future of humanity, I would say. Yes. So how can we get there?

21:36I think public discussion, but also a little bit of policy push would help. Before we talk about policy push, I think maybe I should ask you more about how AI would work in a complementary way to people in their jobs. You had some good examples in the paper, especially in the healthcare field, I think. Yeah, I mean, I'm happy. I can give examples from any field that you like. Let me give, actually, education, which I think is very easy to understand. Education is one of the areas where technological change has been very slow. And there has been a lot of emphasis recently on using AI in education.

22:12But if you look at what most companies are trying to do is they're trying to replace teachers. They're trying to take things that teachers used to do and either centralize it or pass it to children by themselves or to children and their parents. So when you have automated grading, automated teaching plans, or teaching that's by AI tools rather than teachers, those are all directly trying to replace teachers' teaching role. If you look at students having access to ChatGPT, you're trying to replace what teacher-student interaction did in terms of teach students learning how to learn and how to write essays and how to develop their own reasoning skills.

22:56You're trying to do that with the students and perhaps students and their parents without with much less input from the teachers. None of that has proven very successful so far. And I think it's going to be very difficult to make those things become really transformative. I cannot imagine. I mean, just having watched my kids suffer through online school in 2020, they didn't learn. I mean, nothing happened. But here is a different way to use AI. Imagine that we provide a background AI tool for the classroom, which as a taking inputs, responses to quizzes from students, or perhaps other things that students are doing during the class, understands which students are having trouble following which parts of the curriculum.

23:42On the basis of that, suggests to the teacher ways of creating different groupings of students so that some groups can be taught at the right level and others can be given more advanced material together with suggestions for overcoming some of the difficulties that students have. Because why can the AI tool do that? Because again, you fed in the right curricular material and you have fed in use cases from very experienced teachers dealing with specific types of problems. So the AI tool has more experience on some of these things than the teacher can. But it cannot teach as well as the teacher because there's a social aspect.

24:18There is an interactive aspect to it. So the AI tool and the teacher are together now achieving something that people have aspired for more than a century, which is individualized teaching, but you're doing it for much cheaper. But critically, you don't need less of the teacher. You, in fact, probably need more teachers to do this. You need very skilled teachers, and you might need two teachers now for this class because now you can divide the class into four groups, and each teacher can take two of the groups. So you can see the big differences. is in one case, you're amplifying what the teacher does, increasing the demand for the teacher, and you're actually going to get much better productivity outcomes.

24:53We know that from a few studies that have followed very costly individualized education programs. Now you can do that in a much cheaper way, in a much more effective way. That doesn't sound that different from what we've seen in past technological advancements or shifts. Did you already mention accountants? I mean, I remember people always say when spreadsheets came out, you know, everyone said, oh, this is the end for whoever, accountants, financial analysts. But what actually happened was everyone just started using spreadsheets and started just doing a lot more stuff with spreadsheets and became more productive, right?

25:27Yeah. I mean, there are some technologies that are flexible enough that you can use them in different ways. They will automate some tasks, but create complementarities in other tasks. And there are other technologies that you can design them specifically to be complementary to workers. So many of the early manufacturing advances needed to be complementary to workers because there was definitely not the possibility of these machines acting autonomously. So you needed a lot of technical skills on the side of the workers so that they could work with these machines. So they needed to be upskilled at the same time as the machines were getting better.

26:02But you can also have technologies that are purely for automation. Robots, great. They've actually improved productivity a lot in manufacturing, but they are an automation technology. They have been designed as an automating technology. Perhaps in the future, we may have core robots that work with workers, but that's not where we are. So technologies are a wide spectrum. Some of them are going to fall in the more automation category. Some of them are going to create new tasks or other complementarities to workers. And yeah, we should say, I should ask you, you're not anti-automation. You're not, right?

26:31I mean. I'm not anti-automation. I think automation is very important. Thank God we don't have to carry things on our backs anymore. Right. Thank God for the washing machine. Thank God that many people don't have to do the very dangerous tasks with machines that could take their fingers apart. Yeah, right. But if all we do is automation, this is going to have horrible effects on workers. Even moderate automation, if it is not coupled with an effort to create jobs, good jobs for workers, can have pretty bad effects. For work I have done with Pasquale Restrepo, for example, shows that robots had fairly negative effects in many communities, especially on blue-collar workers, because they used to be essential for the production process and got paid quite well.

27:20And they lost those jobs because they were displaced by robots that did these things more cheaply. Firms benefited, but that led to 20, 30 years of lower wages, lower employment in those communities. Right. And that's now, I mean, that's the fear. That's going to potentially what we are aiming for is something much bigger with AI. Now, let me also say that the same things that I tried to argue earlier, that they are underestimating human intelligence, human capabilities, human dexterity, also means that the spread of AI models is not going to be quite as fast as what some people in the industry are thinking, even if they just aim for automation.

28:00So it's not like we're going to have a job apocalypse tomorrow. But even if only 10 % of jobs are automated away with AI, that's going to have fairly large distributional consequences. So actually, one more beat on the large distributional consequences if 10 % of jobs get automated away. I mean, if the goal is among these companies is to drive more and more profits, isn't it counterproductive to get rid of that many jobs because you're getting rid of your consumer base? I mean, isn't that sort of... Well, it depends. So the consequence of large-scale automation is a combination of three things.

28:39One is there is a productivity gain. Firms, I mean, there could be circumstances in which that's very trivial or even not present. But in most of the time, there's either a small or a large productivity gain. And that's additional earning power for somebody. Second, automation changes the balance between capital and labor. Labor earns less, capital earns more. So capital owners will have more money. Third, there is a greater inequality within labor. You know, when we had robots, the labor share went down relative to the capital share. But also engineers, technicians, managers did well. Blue-collar workers didn't.

29:17So then the question is, one group is losing purchasing power. The other group is gaining purchasing power. Now, if the people who are gaining purchasing power are Elon Musk, Sam Altman, and Bill Gates, well, they have so much money, they're not spending much of it. If, on the other hand, if it's engineers and technicians, they could spend more. Now they could start buying luxuries and a second home and very expensive wine, et cetera. It may not be what you want as a society that some people are hungry and the rest are buying, you know, vintage French wine. but the earnings might still translate into consumption.

29:53If it all goes to Elon Musk, well, he can't spend$3 million a day. So that might actually depress consumption. Okay, so then get back to it. You talked about a policy push. How might a policy push towards pro-worker AI, what would that look like? So I think there are two legs to it. One is get rid of existing distortions that make such an agenda difficult. And the other one is take facilitating enabling actions. In the first bucket, I would put the tax code. Our tax code artificially encourages automation because we tax labor quite heavily, more than 25 percent, payroll taxes, health care contributions, income taxes, whereas we subsidize capital close to zero percent.

30:40So that creates a very big wedge between capital and labor and makes it attractive but not productive for firms to use capital at the margin rather than labor. So we could have a more neutral tax policy where capital is not subsidized, and that would help. Second, we have a hugely concentrated industry in the tech field, more concentrated than many of the areas like steel, railways, oil were at the beginning of the 20th century, which started the progressive movement and all the antitrust actions. And that matters not just because it creates monopolies, but if we want new technologies often come from new entrants that have new business models and try different things.

31:27And that becomes harder and harder in a concentrated industry. So creating a more competitive environment and getting rid of the distortions is one thing. A second thing is we can also have the government encourage more pro-worker areas. How would it do that? it would definitely be a horrible idea for the government itself to become an entrepreneur or a venture capitalist and decide what technologies should be produced. But like what we have done in the area of solar panels or other technologies that are undersupplied by the market, the government can provide certain subsidies or create tournaments or competitions where, you know, prototypes that are doing things that the market isn't doing could be rewarded.

32:10The government has done that very successfully, even in the area of AI and robotics. The early challenges that the government had for robots that can do this or that under the auspices of DARPA were very important in kickstarting big investments in robotics. So we could do that the same thing for pro-teacher technologies, pro-nurse technologies, pro-electrician technologies. So there are a lot of things that with a small amount of money, but leveraging its leadership role, the government could move the dial. It seems hard to imagine right now. It seems hard to imagine. Really hard to imagine. I mean, the one thing I was thinking of, because as we're taping, we're in the middle of this war in Iran.

32:49And I was thinking war is generally advanced technology. And I was wondering, like, how the use of AI in this war is going to change, you know, or advance AI. But I think if you're actually, if anybody is making note, The experience of using AI right now should encourage them to go in a more human complementary direction because they are seeing it every day that AI is providing useful information to humans. But if you actually let AI choose targets, it's just not up to that. Imagine, for example, we were running an experiment and we said in this city, in Denver, from now on, air traffic control is completely delegated to AI.

33:29Would you want to fly to Denver? Absolutely not. Never again, no. But same thing. Yeah. Imagine you told the generals, from now on, AI is going to choose the targets. No human is going to interfere. But your life and your court martial is on the line. Would they agree to it? No. So that should be a lesson about how we should try and strive to use AI. Yeah. So much of it is just, it's narrative. It's just trying to change the story from, you know, science fiction to a different kind of fiction, at least. 100%. Well, thank you so much for coming on Slate Mind. This has been great. My pleasure. Anytime.

34:04Thank you, Emily.

34:32a private tutor in your pocket. Start speaking with Babbel today. Get up to 55 % off your Babbel subscription right now at babbel.com slash acast. Spelled b-a-b-b-e-l dot com slash acast. Rules and restrictions may apply. It's deck days at Lowe's and the savings are stacked. Right now, pros get 15 % off all in-stock composite decking from top brands like Trex, TimberTech, and Decorators. Plus, get a free DeWalt 20-volt max 5-amp-hour battery when you buy a select DeWalt tool. The deck's stacked in your favor with brands pros trust. Our best lineup is here in Lowe's. Valid through 422, while supplies last.

35:18Selection varies by location.