Try Free

Recessions, Productivity, and Job Loss Related to AI

Paul Krugman November 20, 2026 7m 1,333 words 1 views
▶ Watch original video

About this transcript: This is a full AI-generated transcript of Recessions, Productivity, and Job Loss Related to AI from Paul Krugman, published November 20, 2026. The transcript contains 1,333 words with timestamps and was generated using Whisper AI.

"My existential fear in all of this is that a lot of historic job loss doesn't happen sort of with new technology, doesn't happen year after year after year in a smooth way. It happens in a very discontinuous way where people get laid off in a recession and don't get rehired. That might be, I'm not..."

[00:00:00] Speaker 1: My existential fear in all of this is that a lot of historic job loss doesn't happen sort of with new technology, doesn't happen year after year after year in a smooth way. It happens in a very discontinuous way where people get laid off in a recession and don't get rehired. That might be, I'm not sure if it was, but it might have been part of the slow recovery from dot-com because we had very high productivity growth, trend productivity growth at that point in time. And so here, that if we have a recession, this becomes the time when all the predictions which have been wrong to date about the impact on labor markets suddenly become true and stay true. That terrifies me. [00:00:41] Speaker 2: We're running a little long, but let's get to that, just this whole productivity job loss. Where are we on that? Because my read is that it's surprisingly invisible in the macro numbers so far. Do we just think that that's lag? [00:00:55] Speaker 3: So I don't think it's invisible. I just think it's hard to see. But if you squint, you can see something. Now, you know, if you squint too much, you can see things that aren't there. But I disagree with some of my colleagues who say there's nothing there. And I definitely disagree with people who say, you know, AI is killing jobs for every college graduate. If you look at the work of Goldman Sachs, economists on this, Joseph Briggs, others, I think they're going to go to Susan Dong, you see they consistently plot these scatter plots with AI exposure on the x-axis and some measure of labor market activity on the y-axis. And those slopes were very flat. And now they're a little less flat or whatever the right slope is. They're a little bit less flat. You know, if it's employment, it's like this. And if you look at the extent to which the difficulty that young college graduates are having, you can see a bit of a correlation there with AI exposure. It's actually true that some who aren't exposed are also having trouble. So it's foggy. But I think there's something there. I think that the point Jason made is one that I've argued and want to amplify. Historically, recessions have been a force multiplier for technological unemployment. If there is a way for an employer to figure out how to not hire you back after she's laid you off, if there's a potential technology that might enable them to do so, they get a lot more interested in that technology post-layoffs. And there is good evidence that recessions that coincide with a potentially labor-replacing technology could be, you know, problematic in exactly the way Jason said. So that's a worry of mine as well. [00:02:59] Speaker 2: I'm actually – go ahead. Oh, no, no, no, no, you, Paul. I was just going to say, you know, my normal – historically, I've tended to say technological unemployment is the last refuge of scoundrels. That it's, you know, whenever there's a recession, people say, oh, it's technological unemployment, which was, you know, that was very much true in the 30s. But it might actually be right this time? [00:03:25] Speaker 1: Yeah, I'm super worried. I mean, look, in general, higher productivity growth is correlated with more job growth, not less. That's when the economy is doing really well, the stock market's booming, there's lots of demand. I hate being a person who sees something happening over and over again and now says the exact opposite is going to happen this next time. So, you know, first of all, people who want to work have been able to work. And, yeah, in the short run, productivity has been good, not bad for employment. You know, at this point, though, you know, do I think there's at least a 15% chance that those past relationships don't hold? Absolutely. Is that terrifying to me? Absolutely. And I think it's probably more than a 15% chance. And we're super early here. I mean, agentic AI is basically one year old. Businesses, individuals, haven't still gone nearly, explored nearly the capabilities the current systems have. And, you know, those of us that don't listen to the people in Silicon Valley about what's coming and just roll our eyes because they sound like a bunch of weirdos, well, those weirdos have been a lot more right about the capabilities, not about the impact. But, you know, don't hold them to the dates in their forecast of the economy. [00:04:42] Speaker 3: I have a different, so I agree with all of that. But I also have this other scary concern, which kind of is on the other side of the axis, which is that I think I might be channeling some of the work of Alex Imus here, which is that there's a combination of Jevons Paradox and Baumol effects, I'll explain what I mean by those, that get you to a place where actually AI, like other advances, creates a lot of new jobs. It's just that they're crappy jobs and that what you end up with is a lot more wealth and equality than we have now, and we already have a boatload of wealth and equality. And this is the idea that AI creates, starts doing a lot of work that people are already doing, but because it lowers the cost of all the things that it can do, and look at Jason's Econ Vitals software, if you want to get a good example of that, then it opens up a lot of space for people to hire new relational kinds of workers, to be chefs, to be therapists, to be personal trainers, to entertain their kids. And there's a world in which we have a lot of jobs supporting the, you know, there's a dark world in which we have a lot of jobs supporting the whims and the entertainment of, you know, are tech overlords. And because of Balmo's effect, which, and that's the Jevons paradox part, the Balmo effect part is that, you know, there's no productivity gain from playing the three-minute waltz in two minutes, right? You can do it, but it's not, it's inherently, like, not good. And so in these kinds of jobs, AI probably doesn't replace them because they're relational, or as I think EMAS puts it, mimetic, which I'm not 100% sure what that means. But the idea is that you just have more of a bifurcation, which feeds into the inequality story, which is already exacerbated. [00:06:43] Speaker 2: I mean, the version of it that, I don't think this is original with me, but that we're quite a long ways from bots being able to empty bedpans, and given our demographics, we're going to need a lot of people emptying bedpans, but on the other hand, if all the new jobs are people emptying bedpans, then we're, that's your world, your dysfunctional world. [00:07:08] Speaker 3: Let me throw in the last one. [00:07:09] Speaker 1: And I just have, I think, you know, there's different sources of inequality. One is how much income goes to capital and how much goes to labor. If I had to bet more is going to go to capital and less is going to go to labor under this, and a wide range of models predict that. I think it's a relatively robust thing. Within labor, which has been the big story in inequality in the last decades, there I'm not sure. It could be that, you know, take somebody who's less skilled and improves their skills, and you're willing to pay them more. And if you're super skilled, it doesn't add much to you, or it could be a false or force multiplier for people that, you know, know how to use it well. So I could tell labor inequality up and labor inequality down stories when it comes to this technology, but almost all of it is capital inequality up stories.

Transcribe Any Video or Podcast — Free

Paste a URL and get a full AI-powered transcript in minutes. Try ScribeHawk →