Most AI benchmarks ask how well a machine can do a task humans already do. Erik Brynjolfsson argues that this framing, inherited from the Turing test, pulls technology toward replacing workers when the bigger gains come from complementing them and from inventing goods that don't exist yet. In this episode, Ben sits down with Erik to talk about the jagged frontier, why young workers in AI-exposed jobs are getting hired less, and what it will take for AI's capabilities to finally show up in the productivity numbers.
- Why a more jagged AI frontier is good for workers
- How tax policy and Silicon Valley's founding vision tilt investment toward replacing labor
- What the Canaries in the Coal Mine paper found about falling employment for young workers in AI-exposed occupations
- Why hiring gains at AI-adopting firms can coexist with job losses across an industry
- Why most progress comes from new goods, told through a thought experiment about ancient Greek robots and clay pots
- How a company's AI power users can level up everyone else, and whether AI will widen inequality over the next decade
- 00:50Is AI a normal technology?
- 02:17What makes generative AI different
- 05:01Why self-driving cars have been so slow
- 07:09Round out the jagged frontier, or embrace it?
- 08:14The Turing Trap: complement and extend
- 10:43Does freed-up time fill itself?
- 13:15The CFO who measured AI by headcount cuts
- 14:58Why AI builders chase replacement
- 19:18Are technologists listening to economists?
- 22:22Is AI seniority-biased? The Canaries paper
- 24:25Why jobs fell but wages didn't
- 28:23Is it an anticipation effect?
- 31:02AI adopters are hiring. What about everyone else?
- 34:06Demand elasticity and the power of new goods
- 38:02Daedalus and the clay pot problem
- 40:15Setting a research agenda in the AI era
- 43:17Business process innovation and Workhelix
- 46:37Does AI help the best workers or the rest?
- 49:10Will AI increase inequality?
- 50:32Which dead economist should be studying AI?
Ben: How is AI affecting the economy today?
Erik: Most progress is from new goods.
Ben: Does anyone really think we're done making stuff?
Erik: It's like the mother of all general purpose technologies.
Ben: It's popular in Silicon Valley to characterize AI as a machine that can do everything a human can. Erik Brynjolfsson thinks that's the wrong target, and that the real payoff comes from building things we've never been able to do before. Erik is a professor at Stanford, head of the Stanford Digital Economy Lab, and one of the economists who built the economics of technology into the field we know today. So much of the way we understand how technology affects the economy comes from Erik's research. Here's our talk.
Ben: All right, welcome to The Economics of Work. I'm excited to welcome Erik Brynjolfsson, professor at Stanford, head of the Digital Economy Lab, and an absolute giant in the economics of technology, really pioneering the field as we know it today. Erik, welcome to the show.
Erik: Thank you, Ben. It's always so good to talk to you.
00:50 Is AI a normal technology?
Ben: Likewise. So you've been working on technology for a long time, and when I look at your early work, in some ways it's remarkably similar to what you're saying today. I view that as really prescient and impressive. I also see it as supporting the "AI as normal technology" view. So first of all, is that right? And second, is AI a normal technology as you understand it, or are we in really uncharted waters?
Erik: I love Arvind and the folks who talk about AI as a normal technology, but I don't really think there are just two categories, normal and abnormal. Hal Varian has this fun saying, which I think he got from someone else, that history doesn't repeat itself, but it rhymes. That's how I think about it. There are certain things we can learn from earlier general purpose technologies, like the steam engine or electricity, and from earlier methods of invention, like the telescope. AI has elements of those, but it's also unique in many ways.
Andy McAfee and I wrote The Second Machine Age, where we thought the closest comparison was the Industrial Revolution, which augmented our muscles. Now we're doing that for our minds. And I think our minds are even more important than our muscles, so I think this is an even bigger deal than anything that's ever happened before in history. But we can still learn from the past.
02:17 What makes generative AI different
Ben: By the way, love The Second Machine Age. I assign it to my students. It's right here somewhere by me, so thank you for that. I think that sounds right. In some ways, the world is catching up to these basic ideas: the way technological innovation progresses and rhymes with other technological innovation. That common ground is being learned by everyone, and so the early work is becoming very popular. But in what way is AI, generative AI, distinct from what you would have said back in The Second Machine Age days?
Erik: There are a number of things that are special about it. First, this idea of a tool that augments, or in some cases even replaces, mental work. That's a whole new territory. As a professor, I've always thought mental work is very, very important. Maybe not the most important thing in the world, but it certainly makes a big difference. I was visiting Google DeepMind a little while ago, and they have this modest slogan over there: we're going to solve intelligence and then use that to solve all the other problems in the world. So there's something pretty special here. It's a general purpose technology, but it's like the mother of all general purpose technologies.
It's also happening a lot faster on many dimensions. There are a lot of ways to put metrics on this, but what it costs to match GPT-4 is so much cheaper now. There's really been no technology like that. The steam engine was amazing, but it didn't advance nearly that fast, and neither did any other technology. That's really special. So we're going into uncharted territory.
When we wrote The Second Machine Age, we anticipated some of this exponential change. We had a quote in there about how things change slowly at first and then suddenly. Ernest Hemingway described that as one way people go bankrupt, but it's also true of exponential curves. You barely notice them, and then suddenly they take off. We didn't know exactly which parts would hit the "suddenly" part and which wouldn't. In my view, we're clearly in the "suddenly" part for a lot of generative technology.
Other parts, to be fair, have gone slower than expected. I was pretty excited about self-driving cars in 2012, when I rode in one with Andy from Mountain View all the way up to San Francisco. I thought it was just around the corner. It's taken a little longer. But net-net, I do think we're at an inflection point where people are going, "What the f*** is happening?"
05:01 Why self-driving cars have been so slow
Ben: So at this inflection point, is your understanding of the slowdown in self-driving cars a J-curve-style thing, where it's taking a long time for the investment to pay off? Or is it more about frictions in implementation? How do you understand why self-driving cars have been so slow?
Erik: It's both of those. And it's something I kind of knew, but it really hammered home for me: there's always this long tail of special cases. The way I think about it, machine learning is really great at doing things when there's lots of data. That's been the big unlock, Rich Sutton's Bitter Lesson: if you throw a lot of data at something, you can solve the problem. Every task has parts where you have a lot of data, the common situations. But most tasks also have a long tail of very rare stuff. A black cat walks across the road at night. You don't see much of that in the training data, maybe none of it, so it's hard for the machines to learn.
Humans aren't great at dealing with exceptions either, but we're better than machines. Improvising is one of our superpowers. So there's a natural division of labor. With self-driving cars, getting to 99%, or 99.9%, or 99.99% is great, but given that lives are at stake, you need a lot more nines than that to be completely safe, and that's something people feel uncomfortable with.
I think self-driving cars are a marker for a lot of other things. When we looked at call center data, we saw a similar pattern: some questions came up a lot, and then there was a long tail of more obscure questions the machines didn't have a good way of answering. It's a lesson we're learning: it's just a lot harder to get the last bit. That's one of the reasons I like having humans and machines work together on a lot of things. We each have relative strengths and weaknesses, and we can complement each other.
07:09 Round out the jagged frontier, or embrace it?
Ben: That's what I was wondering about. On the one hand, we've got the Bitter Lesson, we've got this long tail of problems, and we want to solve intelligence so we can solve everything else. That makes me think there are high returns to solving that last mile and getting really good at the things you're not good at. But then you have the centaur model. You wrote this excellent paper, "The Turing Trap," which goes through the advantages of having AI do what machines are good at and humans do what humans are good at, as complements.
In some ways I feel there's a bit of a conflict there. If you can't get much of the return until you've rounded out the jagged frontier, is it better to round it out, or to keep the frontier as jagged as possible and complement it with humans at every step?
Erik: I'm glad you framed it that way, because it sets me up. I have a point of view on this. I know podcasts do better with hot takes, so I'm going to take a...
Ben: No, no, go crazy.
08:14 The Turing Trap: complement and extend
Erik: I love the jagged frontier. The more jagged, the better. As an economist, that creates comparative advantage and gains from trade. When machines are good and getting better at the things they do well, and humans are good at the things we do well, we each benefit from interacting with each other. I think it's a mistake that so many technologists are focused on rounding out the weak spots, trying to get machines to do things that are hard for machines and easy for humans. We should do it the other way, like you said: have machines do what's easy for them, and we do what's easy for us. Then we both benefit. And the more we push that jagged frontier, the more new things we'll explore.
That's what "The Turing Trap" emphasizes. It's an echo of the Turing test, and it argues why that test is wrong. The Turing test was the idea that we try to make machines as similar to humans as possible. It's literally called the imitation game. I have to admit, when I first read about it at maybe 14 years old, I thought, "Oh yeah, he solved it. This is clearly what we should be doing." I was so wrong. It's exactly the opposite of what we should be doing. It's not just a bad test of intelligence. It's a terrible direction for technology.
Let me draw two distinctions in how we can add value. There's complementing, where humans and machines together do things better and create more value. That's really important. But there's another dimension, which is extending: doing new things that were never done before. I talk about both in "The Turing Trap," but I blurred them together a bit, and I really want to highlight that they're different, and that a lot of value comes from extending, discovering and enabling things we've never done before.
So much of what I see executives and technologists doing, and sometimes I do it myself, is gravitating toward making a list of everything humans already do and asking, "Can machines help with those?" But that list is the smallest part of the value. The real value is outside that list, in the things that don't exist yet. Unfortunately, that list is very hard to make, so we're biased toward the first one. But if we're more creative and think about things that haven't been invented yet, that's where almost all the value has come from historically, and I think it will going forward as well.
10:43 Does freed-up time fill itself?
Ben: Do you think that happens organically anyway, though? I think this is true for myself: if half my day were automated and I had more time, I'd think of other things to do. My appetite would grow, I'd want to do more, and I'd find a way to do it. I don't know if I'd need to think about it purposefully. In that way, does supply create its own demand? Does our capacity just fill itself?
Erik: A couple of things. First, it's definitely nice to have annoying things taken off my plate, things I'm already doing, the dirty, dull, dangerous tasks. Let's eliminate those, I'm fine with that, and it'll free up time for us to do more creative things, maybe things humans do better. So I think all of that does happen organically, to some extent.
But I also think there's a lot of room for steering technology. I don't believe there's one inevitable direction that will happen no matter what we do. I'm very much of the view that we have agency, and the more powerful the tools, the more agency we have.
We put out the Stanford AI Index, 520 pages of metrics on how AI is doing. And almost all of those metrics measure how well a machine by itself can do some predefined task. That's just how the benchmarks work. The amazing thing is how good the AI labs are at hitting them. One of their problems is that they saturate benchmarks so quickly that they have to come up with new ones. That's kind of cool. But I'd love to see us think harder about benchmarks in the other direction. What new things could we do? How can we extend science? How can we complement humans? Let's try to hit some of those benchmarks.
When I talk to my technology friends, there's no question they feel they could work on different kinds of problems, make progress on them, and succeed. Some opportunities just emerge, but our own decisions about what we focus on make a first-order difference. In my research, if I work on one kind of problem, I'm more likely to solve it than if I work on a different one.
13:15 The CFO who measured AI by headcount cuts
Erik: And this is true not just of technologists, but of executives I talk to. Let me tell you a quick story. I was talking to a chief financial officer, and she told me she wanted to measure how effectively each department was using AI. I said, "That's great. I love measurement." You love measurement, Ben. So I was excited. Then her next sentence was, "Therefore, we're going to ask each department to report back how much headcount reduction they've achieved from AI."
I was like, "Okay, I see where that's going." No wonder AI is a very unpopular technology, and no wonder her staff resists implementing it, if that's their target. I don't doubt they'll hit it. But I implored her: "You're the CFO. You make a lot of money. Your job is to be more creative. Come up with some other KPIs. How can you improve customer satisfaction? How can you reduce employee turnover? How can you invent new products and services?" There are all these other things you could be doing. She said, "Fair enough, but those are harder to measure." And I said, "Well, that's why you make the big money, right? You've got to do the hard things." I think she will start focusing on those other measures, and that will probably do more to grow the top line than just cutting costs.
Ben: It really is unfortunate that a lot of CFOs are feeding right into the stereotypes of CFOs.
Erik: It's amazing. It's like a gravitational pull: how do we look at existing costs and take them out? It's the easiest thing to grab for. I don't blame them, and there's nothing wrong with being more efficient. Just try to do this other stuff too.
14:58 Why AI builders chase replacement
Ben: I'm also very sympathetic to the point that these benchmarks really are designed around substituting for human labor. My hope is that the ultimate benchmark is actual creation of economic value, and that will hopefully be the arbiter of what gets built in equilibrium.
But going back to this: why is everyone making this mistake? One theory is that the people building AI are like 14-year-old you. They're enamored with the idea of creating humanlike androids, and it's just cool. Another is that if they see their job as a race toward recursive self-improvement, maybe a rounded frontier makes sense for that reason. Or maybe handoffs and task chaining are really important. I don't know. What's your best theory on what they're getting wrong?
Erik: I think it's a combination of things, and I talk about this a bit in "The Turing Trap." Part of it is that, as a society, we've set up some screwed-up incentives. The government has tax policy and hundreds of other policies that reward managers for making certain kinds of investments. According to my former MIT colleague Daron Acemoglu and Pascual Restrepo, if you unpack that, there's about a 5x gap between investing in capital and investing in labor. You end up paying a lot more in taxes when you invest in labor. So the government is putting its thumb on the scale. The first rule of taxation is that you get less of whatever you tax more, so the government is steering decisions toward replacing workers with capital. Is there some societal reason we want to do that? I don't really think so. We should at least level the playing field, maybe even tilt it the other way.
That's part of it. But as I've spent more time out here in Silicon Valley, I actually think it's the least important part. There are these tax incentives, but when I talk to folks, I'm amazed at how mission-driven and vision-driven people are. They want to solve AI. They want to solve all the other problems in the world, like we said earlier. And they listen to people who tell these big visions. Alan Turing told a vision. The folks who coined AGI had a vision that AGI means humanlike intelligence. Nils Nilsson, who was a professor here, said literally that if you want to make powerful AI, you need to look at each task humans are doing and have a machine do each of them. So they set the direction, and a lot of people said, "Oh yeah, sure, that's really cool, let's work on that."
What I'm trying to do is set a different direction, a different vision. How can we expand the pie? How can we do new things we've never done before? How can we complement humans? And it's been gratifying. I gave a talk with this alternative vision at ICLR, one of the big AI conferences, and a bunch of technologists came up to me afterward and said, "We didn't realize it's better to pursue this other vision. We can do that. Thanks for telling us." It was gratifying that they felt they could work on it.
One concrete lever, beyond you and me talking about it in big-picture terms, is changing the benchmarks. Andy Haupt and I have developed what we call centaur evaluations, or centaur benchmarks. It's not "How well can the machine read this medical image?" It's "How well can a human and machine together figure out what's going on?" To be fair, centaur benchmarks are harder to create and harder to assess, but ultimately I think they'll help people concretely steer the technology in that direction.
And I don't think any one of these levers is a silver bullet. I just think that, on average, we're overemphasizing the replacement mindset and underemphasizing the extension mindset, and I'd like to steer things in that direction.
19:18 Are technologists listening to economists?
Ben: It's very encouraging. This perspective is very intuitive from an economics perspective, and it seems very unintuitive for AI people. You've been at the intersection of technologists and economists basically your whole career, and now you're even more at the epicenter of it. Do you think technologists are listening to economists more today than even a year or so ago?
Erik: A little bit. But I want to pick up on your "intuitive" point, because it's really striking to me. A lot of non-economists, even some business folks, have literally told me that, by definition, increasing productivity means replacing workers. That's just mathematically wrong. I don't know where they get this mindset, and I try to correct them. I once had a heated debate on a panel with somebody who insisted you have to reduce headcount, and that's just not true.
Here's an example of how untrue it is. Look at how much you have to pay for an hour of human labor over the course of history. It's gone up and up and up, with some ups and downs, and lately some groups haven't done as well. But an hour of human labor is worth a lot more now than it was 50 years ago, or 200 years ago, or 2,000 years ago. That means technology has mostly enhanced the value of human labor, not driven it down or replaced it. It doesn't always have to be that way, but most technology has created more productivity by making humans more valuable. So that's a big part of it. And I'm trying to remember the second part of your question.
Ben: Are the technologists listening to economists?
Erik: Oh yeah. It is changing. One of the interesting things is that AI, to be fair, used to be this little corner of research that was fun and intellectually interesting but didn't really change the world much. I first started working on AI, believe it or not, back in the 1980s. Tod Loofbourrow and I taught a course at Harvard Extension School on applications of artificial intelligence and building expert systems, and we had a company that built expert systems for banks and other groups. That early kind of AI worked so-so, and it was nowhere near as powerful.
But now AI is affecting every industry to a much greater extent, which means technologists have no choice but to think about the societal implications. That's great. Dario Amodei, for instance, has written some great essays, like "Machines of Loving Grace," about how AI is changing society, and he came to me and asked for some advice on that, which was gratifying. A number of other technologists have reached out to think through how this is going to change society. It's a testament to the technology no longer being a laboratory curiosity, or something that works in a few banks here and there, but arguably the most impactful economic technology of our era.
22:22 Is AI seniority-biased? The Canaries paper
Ben: It's certainly everywhere, which brings us to the question of how AI is affecting the economy today. I want to dig into some of your research: in your opinion, is AI a seniority-biased technological change?
Erik: It looks like there's a lot of that happening. In my paper with Bharat Chandar and Ruyu Chen, "Canaries in the Coal Mine," we went in with no priors. And it was interesting, you guys found some similar things. But let me describe what we saw.
We found there wasn't a big effect on the top line, just a little falling employment for young workers. Where it got really interesting was when we subdivided by AI exposure, using the Eloundou et al. metric or other similar metrics, and by seniority, like you just asked. Specifically, we found that young workers had falling employment, not so much because they were being laid off, but because they just weren't being hired as much. When we first did the paper, it was about 13%. A follow-up found 16%. In the latest version, it's about 19% less employment compared to more senior workers, or compared to less exposed workers. So that does seem to be going on.
The flip side is that the senior workers are doing fine, and people in unexposed occupations are doing fine. To calibrate that: the exposed occupations are things like coding and call center operators. Out of about 900 occupations, we took the top 20%, about 180 occupations. The least exposed are things like home health aides. An LLM can't really put in an IV or help you into bed.
So that's a long way of saying yes, AI is disproportionately affecting some younger workers. And we found a similar thing in a much more focused study of call center workers, where we had more of a natural experiment and could separate out how much was due to AI versus other factors.
24:25 Why jobs fell but wages didn't
Ben: This Canaries result is such an important thing to know about what's happening in the labor market. Major props to you and your co-authors on the paper and the analysis. It's amazing. I want to really understand the mechanism. One point I get confused about, and I'm not sure how it updated in the revised paper, but at least initially, the effects on younger, more junior workers seemed to be primarily on the quantity of labor, on employment, not on their wages. Why is that? You'd think at some point the market would clear. If it's a decrease in demand for younger workers, why wouldn't that hit wages rather than quantities?
Erik: I think it's early, and I fear we will start seeing it more in wages. Historically, wages have been a little sticky, particularly downward. If you run a company, giving somebody a pay cut is a very painful thing to do, so you just don't see it much.
Ben: But that's only for existing workers. For new employees, you can...
Erik: And so what we saw was that a lot of people just aren't being hired as much. Over time, the market will start adjusting, but it's fair to expect the first place you see it is in employment patterns. Young workers, who are more on the margin of being hired or not, are also the ones likely to see it first. Over time, I'd expect some of it to show up in wages. That's part of the answer.
The other part is more intellectually interesting. David Autor and Neil Thompson have a terrific paper on expertise. Suppose you have a job with 20 different tasks. I know you've been a leader in thinking about jobs as bundles of tasks, and we think about it that way too. AI comes in and automates some of those tasks, or takes them off the table. If it takes away the bottom set of tasks, the remaining tasks are higher skill, and the remaining work might actually be higher paid. If AI takes away tasks near the top of the distribution, the remaining tasks will be lower paid. So it's ambiguous, depending on which set. And they showed the same technology may affect one occupation by taking away the top half and another by taking away the bottom half, with different wage effects. So it's not automatic. Even if AI removes some tasks and some demand, the remaining workers may get paid more than before. We shouldn't automatically expect wages to go down.
Ben: Interesting. So the thought is that AI makes younger workers more productive and enhances their high-value areas of expertise, while at the same time decreasing demand for those types of workers.
Erik: Exactly. The people who still have jobs are doing more high-end work. To be concrete, in coding, maybe AI takes away some of the routine stuff that wasn't all that sophisticated. The people who still have jobs are now doing more systems design and overall organization, and the mundane parts of the job are gone. So if you have a job, you're probably creating a lot more value and doing more of the high-end things that are fairly rare, so you get paid more. What we'd see is less employment, but higher wages for those who are left.
28:23 Is it an anticipation effect?
Ben: Interesting. Let me hit you with another theory and see what you make of it. Young workers are always a riskier bet than older workers. You never really know what you're going to get. So if you have a very high discount rate and want to prioritize the present relative to the future, you'd generally prefer older, more senior workers, who are a surer bet. To the extent that uncertainty about what AI will do to your organization is a big factor, that would raise your discount rate, and you'd prefer more senior workers. It would stand to reason that this would be skewed toward the more exposed parts of the organization.
So is there some kind of anticipation effect? Maybe this isn't due to AI adoption at all. Maybe firms are hiring fewer younger workers because they don't know how they're going to use AI.
Erik: I agree with that whole thing. The only words I'd change are "not due to AI at all." It's due to AI in the sense that firms are anticipating it. It may not be affecting them yet, but it's rational to anticipate some changes and to have some uncertainty about them, like you said. Maybe you don't know exactly what they'll be. That seems very plausible. If you look at the press, there's a lot of talk and a lot of concern about this. I wrote a book with Andy 12 years ago, so it's not like nobody saw this coming, and people are beginning to think about it. Geoffrey Hinton said, "Stop hiring radiologists." I think he was a little premature, or wrong, about that. But there's discussion of it, and people will start acting in anticipation, for sure.
So we have at least three hypotheses out there about what's going on, and I think all three are likely partly true. The right thing is what you're doing, and what we're trying to do: better measurement. We've created the AI Economic Indicators. One of my big frustrations, and Tom Mitchell and I wrote a paper in Nature about this, is that we're flying blind into the most uncertain, consequential time, certainly in my lifetime. We need better data. What you guys are doing at Revelio, and what we're trying to do: there's a lot of real-time or nearly real-time data we can make available, and businesses and policymakers need to see it. You and I can come up with these hypotheses, and I like them all, but ultimately let's test them against the data, and get something close to real time so people can know which path we're on.
31:02 AI adopters are hiring. What about everyone else?
Ben: I really want to follow up on the data, because that's fascinating and I'd love to nerd out about it. But one more thought on younger workers. We had a paper in collaboration with Ramp, where we found that firms adopting AI faster than the not-yet-adopters, sort of treated versus untreated, were increasing employment. They were hiring more than the non-adopting firms, and that was true even for younger workers.
What I took from that, and I'm curious to hear your take, is that AI could be displacing younger, entry-level workers in aggregate. But what we can maybe say, and let me know if you're comfortable saying this, is that the channel isn't firms adopting AI and cutting headcount. It may be the non-adopting firms failing to hire.
Erik: Exactly. I'm so glad you brought this up, and I agree 100%. It's unfortunate that the Ramp study and some others with similar findings were sometimes read as, "Aha, AI isn't displacing jobs, because look, these firms are hiring more." That's such a misunderstanding of what they found. It's a disequilibrium result. It's perfectly possible that a firm that adopts AI grows and succeeds, or conversely, that a growing firm is more likely to adopt AI. So you get a correlation between AI adoption and employment. Meanwhile, the opposite is happening in the rest of the industry. The adopters are taking market share from the others. That's part of their pitch, after all. So you see shrinking employment elsewhere.
What we measured in the Canaries paper, and what you measure in the Revelio data, is the broader effect on the industry or the economy. The Ramp data is just within the firm. There's no reason those have to go in the same direction, and it's completely consistent with economic theory that growing firms take market share without the whole industry going that way.
Ben: I also think it has important policy implications. Sometimes we hear the narrative that firms are going to adopt AI and lay off half their workforce. We heard an announcement like that from Block and other companies: increases in adoption followed by layoffs. That's probably not what we have to worry about. We probably have to worry about firms adopting AI and growing super fast while firms that fail to adopt struggle. And those struggling firms may be big employers, which could be problematic.
Erik: We definitely see the laggards in trouble. To be fair, I don't think all firms that adopt end up growing. Sometimes they do cut back parts of the business. We need the data to get all these nuances.
34:06 Demand elasticity and the power of new goods
Erik: You're an economist, I'm an economist. One tool I love explaining to people is the demand curve. A downward sloping demand curve means that with lower prices or more efficiency, you get greater quantity. But it makes a big difference how steep that curve is. If it's very steep, then as you get more efficient and lower prices, quantity increases only a little, you end up spending less overall, and maybe you hire fewer people. That's very intuitive to a lot of folks. But a demand curve can also be very flat. We call that a very elastic demand curve. In that case, a small increase in efficiency can lead to a big increase in quantity, and you may end up hiring more people.
Either outcome, less hiring or more hiring, is perfectly consistent with economic theory, and both happen in the real world. In agriculture, we have fewer and fewer farmers as we get more efficient. In other industries, like airline pilots, jet engines made them more efficient and we ended up hiring more of them, because demand grew even faster than efficiency. There's nothing paradoxical about having more workers or fewer workers. It varies from industry to industry, and you ultimately have to measure it carefully to know which group you're in.
Ben: Let me really get into this, because it's also confused me a little. I'm reminded of something you said to me once, I want to say in 2019 or 2020.
Erik: We've been talking about these issues a long time.
Ben: I know, we go back. It was kind of an offhand comment: "Does anyone really think we're done making stuff, that we have everything we need in the world?" It really stuck with me. Of course we don't. We're not done. That would be crazy.
Erik: Exactly.
Ben: It makes me think about this demand curve issue. Of course, in some industries and product lines, some demand curves are elastic and some aren't, and there are distributional effects for sure. But going back to grad school days, in a closed economy, doesn't the aggregate demand curve of the world have to be unit elastic? If we think about all the demand curves in the world, don't they have to average out? Can we derive some kind of dynamic from that? I don't know.
Erik: I think it mostly goes in one direction, which is toward more growth. Through most of history, and maybe this will change in the future, we've just kept expanding the set of things. I went to a conference in the 1990s that Bob Gordon and Tim Bresnahan organized on new goods, and I came away with an aha, like you had, about new things: most progress comes from new goods, from things we never had before. I'm pretty sure that will continue, at least for a while. That's net expanding, and that's why we're richer over time, and why there's so much opportunity going forward.
So I'm on the side that we should think harder about that. At the same time, and I'm not a psychologist, there's a bias toward looking at what already exists and asking how technology affects it, as opposed to looking at things we've never seen before. We have a cognitive bias in the wrong direction. Part of my job with "The Turing Trap" and other work is to get us to think harder about that other side.
38:02 Daedalus and the clay pot problem
Erik: Let me tell you a little story. You can cut it if it's too long for your podcast.
Ben: No, no.
Erik: It's one I wrote about in "The Turing Trap." Imagine the ancient Greeks and Daedalus, the mythical engineer who supposedly made robots that looked like humans, so you couldn't tell them apart. Suppose he had done what Alan Turing suggested and had those robots perfectly imitate humans and do each of the tasks the ancient Greeks were doing: making clay pots, fixing chariots, burning incense when you were sick so you'd feel better.
The good news is the Greeks would no longer have to work. Everything they'd been doing could be done by robots. Mechanically, productivity would go to infinity: zero work, and you could consume as many clay pots as you want. But when I describe that, you probably think, how many clay pots do I really want? I want mRNA vaccines and jet planes and iPhones.
Almost all of our increase in living standards has come not from cheaper Greek stuff, but from all these new goods. That was true over the past 2,000 years, and I'm pretty sure it will be true for at least a few decades going forward. I don't know how far. We need to think about that side of the ledger.
Ben: It's a very compelling narrative. The question I still have is: if we had automated the production of clay pots, would the ancient Greeks have moved on to better things? Would they keep innovating with their free time, or would they live a life of leisure? I guess this goes back to Keynes's "Economic Possibilities for Our Grandchildren."
Erik: I hope they would discover more new things with that free time. But if we direct all our energy toward cheaper clay pots, we're just not going to get as many mRNA vaccines. It does make a difference what you focus on. So your point is right: we should focus on those second things, and maybe some of it will happen organically. But the reason I run a lab is that I want to try to direct people toward the things I think will have higher payoffs, and hopefully my bets are in the right direction.
40:15 Setting a research agenda in the AI era
Ben: Speaking of running a lab, you mentioned constructing good data. Your lab is doing so many incredible things. I have two questions. First, how do you personally think about setting a research agenda today? It sounds very challenging, and I'm curious about your approach. And second, what's the dream dataset that makes you salivate?
Erik: I've never had as much fun in my life as right now, because I'm out here in the middle of this remarkable change. It's fun being in Silicon Valley, and it's fun being at this moment in history. Part of what I do is spend a lot of time meeting with people on the technological frontier, and they tell me what's in the pipeline and some of the challenges. I also spend a lot of time with senior executives, listening to their problems and the bottlenecks they're facing. My research agenda comes very much from them.
A lot of economists sit behind a computer terminal, look at the data, and think deep thoughts, and that can work. But I feel like there's a cheat code in actually going out and talking to people in the organizations I'm trying to study. Sometimes I have to take what they say with a grain of salt, but they'll tell me about the technological breakthroughs and the challenges they're facing, and then my brain stirs all that up. That's how I generate hypotheses.
To be more concrete: I'm seeing the biggest gap I've ever seen in history between where the technological frontier is and what's actually happening on the ground, and I think a lot of people are seeing it. If anything, people are underhyping how amazing this technology is and what it can do. It's breathtaking. And if we get recursive self-improvement and self-sustaining acceleration, we'll see even more improvement, rapidly. It's not tapering off, in my opinion. At the same time, there's a ton of frustration among senior executives that they're not getting the productivity gains they expected. Overall productivity in the economy is barely moving. I have a bet with Bob Gordon that that's going to change, and I think it will.
Ben: I'm on your side of that bet.
Erik: Well, thank you. Maybe if we work together, we can improve our odds by speeding things up. That gap between the capabilities overhang and what people are actually doing is where the big challenges and opportunities lie, and it's what our lab is focusing on.
And as you know, I've also started a company. I think you had Daniel Rock on your show a little while back. Andy McAfee, James Milin, and I have a company that's trying to walk the walk. I don't just talk about it. We're actually trying to help companies close the gap by showing them what successful companies are doing, and what the successful people inside their own company are doing, to use the technology more effectively. Maybe that will help me win that bet with Bob Gordon.
43:17 Business process innovation and Workhelix
Ben: One thing I say to my students as a kind of mantra, which I think comes from you (I attribute it to you, though I don't know if you were the first to say it), is that technological innovation has to be paired with business process innovation, and that's where you get productivity. So my spiel to students is: think about the business process innovation. What are you going to do to use these technologies and recombine them in ways that actually improve a business process?
To me, that feels, number one, understudied, and number two, hard to even conceptualize. I imagine it's kind of bottom-up, with people using their own initiative to re-engineer the way they personally do work. What do you make of this business process innovation component? What is it, even?
Erik: To be fair, I'm far from the only person who's said it, but I may be the one who's said it most often and most loudly. It was the focus of my PhD dissertation, and I've been harping on it ever since. I've become very convinced that buying technology alone is almost never sufficient. You have to rethink your organization. And like you said, it's really hard to do because it's different every time. If it were the same every time, it would just happen quickly. Probably all your viewers have heard the story of electricity, and how it took 20 or 30 years for factories to re-engineer and get the benefits. We don't want to wait that long. We're trying to speed it up.
There's a lot of truth to your point that it's often bottom-up, because the people on the ground are the ones working with it. That said, we can turbocharge it. That's what Workhelix does. Instead of waiting for people to talk around the water cooler or on Slack and say, "Hey, I've been using this tool to do my coding a lot faster," we speed that up. And sometimes what people share isn't actually true. Some things work and some don't. There's a lot of friction and noise in getting that information around, which is why it takes so long.
What we do is instrument things, so we have telemetry on what's actually working, what's not, and who the leading-edge people are. In every single company we see, there's a power law distribution: a few percent of people are just crushing it. They're getting 10 times more value. They're not necessarily using many more tokens, but they're using them more effectively. Then there's this long tail where the typical user is writing more eloquent emails or, at one company, doing glorified translation. That's okay, but it's maybe 1 or 2% of the potential gain.
That's great news, because it means you can level up. You can take templates from those successful people and replicate them across the organization, and very quickly get a 2x or 5x improvement in baseline performance, whether people are in marketing or coding or sales. So we can speed up the metabolism. It's still kind of bottom-up, but it's turbocharged compared to how it would have happened otherwise.
46:37 Does AI help the best workers or the rest?
Ben: There's so much to unpack there about what that does to inequality. If you have these super users, maybe it's a superstar-biased technological change. Does that sound right?
Erik: It depends. Sorry to be an economist.
Ben: No, no.
Erik: Let me tell you what it depends on. Let me start with the other side, because I think it's the bigger part of the story. In a paper with Lindsey Raymond and Danielle Li, we looked at call center workers and found just the opposite. The less experienced call center workers, and separately the less skilled ones before the technology was introduced, got the biggest boost, about a 35% gain in performance, when we introduced an LLM to help them answer questions. The best workers got close to zero.
Why was that happening? The way the system was trained, it took all the best answers, the ones that made customers happy, from millions of previous conversations, distilled what a good answer looks like, and made that available to everybody. So the less experienced workers were like, "Oh great, this is how I answer this gnarly tax policy question." The best workers looked at it and said, "Well, no kidding. That's what I've been saying all along. You're just telling me back what I already say." So they weren't getting much of a gain.
Since that study, a number of people have found a very similar pattern in coding, writing, and management consulting, where AI disproportionately helps the less strong workers. At the same time, and this is why I said it depends, there are places, in coding and elsewhere, where the best workers get the biggest gain and the weaker workers don't get as much. There was a study of entrepreneurs in Africa that found this as well.
Ben: Dave Holtz's paper.
Erik: Exactly. My hypothesis about what's going on is that if there's a right answer, like "this is how you change your password," it tends to level up. The best people are already giving that answer, so they won't get any better at it. If two plus two equals four, they're not going to make the ball any rounder, as the saying goes. But everyone else learns how to do it, so it levels up. If it's something without a natural ceiling, like scientific discovery or entrepreneurship...
Ben: Like most things, I would imagine.
Erik: I don't know about most, but a lot of things. Then there's room for the best to keep getting better. So you can have either happening, depending on the domain.
49:10 Will AI increase inequality?
Ben: Okay, I've got two final questions for you, real quick. If you were to take a bet, and I know you like taking bets...
Erik: I do take them.
Ben: In 10 years, would you bet that AI increases inequality or decreases inequality?
Erik: Okay. What I'm going to say is that it really depends on our choices. If we stay on the current path and don't change anything, it's going to increase inequality. But one of my missions is to explain to people that it doesn't have to be that way. I call myself a mindful optimist. I believe technology isn't something that happens to us. We're not going to sit back and say, "I wonder what AI is going to do to us." We have to be active. We have agency. What do we want AI to do?
So my message to all your listeners, to the members of Congress and senators I talk to, and to executives is: what kind of world do you want to build, and how can we use these tools to create it? I believe we can create a world of widely shared prosperity, but it's not going to happen automatically. On the current path, inequality is going to go up. I'm going to do my best to lose that bet and get us to a world with not just prosperity, but shared prosperity.
50:32 Which dead economist should be studying AI?
Ben: I love that. I hope you're successful at losing that bet. It's such a beautiful note that I kind of want to end on it, but since I can't help myself, I want to ask one final question. I also asked this to Ronnie Chatterji, and I'm curious to hear your answer. What dead economist do you wish were working on the economics of AI today? Because the live ones probably already are.
Erik: They should be. I was just talking to Nick Bloom at dinner last night. He did a little study a couple of nights ago. He thought about 50% of economists were working on AI, and then he checked the data, and it was only about 15%. So you and I have to change that, Ben. We've got to get it up to at least 50%. Everyone should be working on it.
Look, John Maynard Keynes did an amazing job rethinking how the economy works. Just as Adam Smith and Keynes rethought the economy during their transitions, today we need somebody to paint the big picture of how digitization and AI will fundamentally change the economy. I don't think that book has been written yet, but it's sorely needed, because the economics of information and AI are so fundamentally different from the economics of the industrial age, let alone the agricultural age, that we're going to have to rethink a lot of our assumptions.
Ben: So Keynes, The General Theory of...
Erik: Yes.
Ben: Okay, Keynes is probably the closest. Cool, love it. Thank you so much for being part of this. I really appreciate it. And I've got to point people to your research, your lab, and all the good work coming out of it. Fully endorse.
Erik: I love talking to you. That was a lot of fun. I'm looking forward to seeing this broadcast, and to keeping in touch.
Ben: Sounds good. Thanks so much.
Ben: The Economics of Work is brought to you by Revelio Labs, workforce data for research and benchmarking. If you enjoyed this episode, please remember to rate, review, and subscribe. It'll help us reach other curious listeners. I'm your host, Ben Zweig. Our producer is Cole Wagner. Thanks for listening.





