Bringing frontier research on AI and economic growth to Yale
Stanford economist Charles I. Jones visited Yale to present new research on artificial intelligence, economic growth, and how economists measure improvements in living standards.
Charles I. Jones (Chad), the STANCO 25 Professor of Economics at Stanford Graduate School of Business, visited Yale this week to present new research on artificial intelligence, economic growth, and how economists measure improvements in living standards.
Jones delivered the biennial Tjalling C. Koopmans Memorial Lecture at the Cowles Foundation, as well a Cowles Lunch Talk, part of two long-running lecture series that bring leading economists to Yale to share frontier research.
In the Koopmans Lecture, Jones presented joint work with Christopher Tonetti examining the role automation has played in driving economic growth, and what recent advances in artificial intelligence could mean for the future. The research argues that automation boosts growth by shifting tasks from slowly improving human labor to rapidly improving machines. At the same time, overall growth may accelerate more slowly than some predictions suggest because production is often constrained by “weak links”—tasks that remain difficult to automate even as many others become mechanized.
“I think AI is the most important thing that has happened in our lifetimes and economists should have a lot to say about it.” - Charles I. Jones
In the Cowles Lunch Talk, Jones turned to a different question: how economists measure progress in living standards. The paper, joint with Philip Trammell, argues that conventional metrics such as GDP can give a misleading picture of improvements in well-being, particularly when economies evolve to include new goods, quality improvements, and services that are difficult to measure. As an alternative, the research proposes using changes in the value of a statistical life—the tradeoffs people make between income and mortality risk—to infer how overall lifetime well-being changes over time.
Together, the talks explored how economists think about the long-run forces shaping prosperity—from technological change and automation to the ways economists measure whether people’s lives are actually improving.
Ahead of his visit, we asked Jones a few questions about artificial intelligence and the future of economic growth (edited for clarity).
Q: How did you select the topic that you are going to present at today’s Tjalling C. Koopmans Memorial Lecture?
CHAD JONES: Koopmans was very much concerned with optimal allocation of resources and how to solve dynamic optimal allocation problems in theory and measurement. The paper I'm going to talk about is very much consistent with those things. How does AI change the allocation of resources and what does that imply about future growth and automation more generally? And then the theory and measurement part is very closely tied to what I do.
So I'm going to do some growth accounting to say, in the past, how much of our past economic growth was due to automation? And that involves writing down some equations, some mathematical structure to use that structure to answer the question. And the structure is actually very important. What exactly do we mean by automation? It's not so clear what we even mean by that. So that's the theory and measurement in the first half of the talk.
And then the second half is saying, well, automation's been going on for like 150-200 years, and AI is just the latest form. Maybe we can use the historical episodes of automation to tell us what AI is going to do in the future. So, write down a model where AI helps you produce ideas and ideas help you automate more things, including how to produce more ideas and how to produce other goods, calibrate that model to the historical evidence that we document, and then run it forward to see what happens in the future. It's not that this is in any way an exact prediction of what's going to happen, but by looking at what happens, you learn some things about the economic forces that are in play. I think that's kind of valuable.
Q: There must be a lot of interest in this topic right now.
CHAD JONES: I think it's absolutely right. So, in economics, everyone, as far as I can tell, is using AI to help us with our work. The economic analysis of AI—people writing papers about AI's effect on the economy—that's happening, but I think the research process, there are long lags, and it takes a while. So people are starting to do this, but I think there should be even more work. I think it's just incredibly important. I think AI is the most important thing that has happened in our lifetimes and economists should have a lot to say about it.
Q: I saw that you started picking up AI as part of your research agenda around 2019. What motivated you? It seems like you were ahead of the curve.
CHAD JONES: It's interesting. So there's a paper that was very important in the economics literature by Joseph Zeira. It was published in 1998, so it was very, very early on. We produce GDP with machines and labor, and you get more machines. So, maybe the number of machine workers goes up, and that's kind of the traditional growth model. And so Zeira said, no, automation is something else.
Instead, automation is the following: There are a bunch of tasks—he didn't use the word tasks, that comes later, but with the benefit of hindsight, this is what he was saying. There are a bunch of tasks that have to be done to produce anything, right? So if you're a factory, you have to come up with the design, you have to hire labor and get all the inputs and get all the materials and the energy, you have to lay out the factory. You have to produce the stuff, you have to check for errors and check for mistakes and repair the mistakes, you have to repair the machines when they break down, you have to sell the goods, you have to market the goods, find the customers, deliver the goods.
What automation is, is learning to do some of those tasks with machines instead of with people. So they got the textile loom in the Industrial Revolution, right? We used to do textiles purely by hand. Then, we invented the textile loom, and then machines could weave and do textiles. Well, the history of automation since the Industrial Revolution is increasingly using machines instead of people on tasks, and that's a different growth model.
I remember seeing that paper and thinking it was very clever and interesting. It had some problems that made it not a good model for understanding the past, but the idea was intriguing. Daron Acemoglu and Pascual Restrepo really picked up on how we should revisit this Zeira structure and start using it to understand automation. They were the first ones to make it a big deal today.
The NBER set up the first Economics of Artificial Intelligence conference in 2017, maybe 2016. This was a conference in Toronto, where they did a clever thing. They got people from throughout the profession to come together and think about AI. They invited me and Ben Jones and Philippe Aghion. We all took ideas we'd been thinking about and put them in this paper, and then that became the thing that got published.
Over time, it's just become increasingly clear that AI has a chance to be the most important technology, right? A pivotal point for me was when the reasoning models came out. I could give it a math problem that I was working on that took me two hours to solve. It would think for five minutes and spit out five pages of algebra that got the right answer at the end. And I thought, okay, that's now useful for me. So I think that kind of moment was where I said, something pretty profound is happening.
Q: I'm wondering how you come to understand enough about the technology to then be able to do the type of modeling and growth work that you do?
CHAD JONES: That's a fair question. Economists are a heterogeneous bunch. We come at all these problems with different skills and different interests and everything else. I would say I don't know that much about how the AI models work internally. I know what I like to do, and I can see, can the model help me do what I like to do? Things that I used to think I was special for, like solving these math problems in elegant ways and getting good intuition, the AI models are better than me at doing that now.
These models are going to be useful in helping us produce ideas. And thanks to Paul Romer, Philippe Aghion, and Peter Howitt, the notion that ideas are key to growth is something that's well accepted. Well, these AI models can help us produce ideas. In the past, we needed people to produce ideas. And if AI models can produce ideas well, you realize that there is the possibility for this positive feedback loop. That kind of flywheel effect could change the dynamics of growth.