Pascual Restrepo on AI, automation, and the future of work

Artificial intelligence is advancing rapidly, raising new questions about how technology will reshape jobs, wages, and economic growth. For economists, those questions are not entirely new. For decades, researchers have studied how waves of technological change, from factory automation to robotics and software, have altered the structure of work.

Few economists have examined those forces as closely as Yale economist Pascual Restrepo, an Associate Professor who joined the Department in 2023.

“There’s a big core question about trying to understand the ways in which technology shapes society and the economy. What I’ve been doing so far is trying to specialize mostly in understanding how technology transforms labor markets.”

Much of Restrepo’s research centers on how technology reshapes the tasks people perform at work. Rather than replacing entire occupations at once, new technologies typically substitute for specific tasks within jobs. Workers then shift toward other tasks that remain difficult to automate.

That framework is central to modern research on technological change, and how AI will affect labor markets. In an overview of automation research, Restrepo argues that the economic consequences of new technologies depend on a race between two forces: machines replacing workers in existing tasks and the creation of new tasks that sustain demand for human labor.

In influential work with his doctoral advisor Daron Acemoglu—winner of the 2024 Nobel Prize in economics—Restrepo has studied how automation reshapes wages and inequality. In research forthcoming in the Quarterly Journal of Economics, they examine how automation often targets tasks that pay workers wages above their outside options. When those tasks are automated, those wage “rents” disappear, amplifying wage losses and contributing to rising inequality. The research was recently highlighted in a Wall Street Journal analysis of how AI may further tilt economic rewards toward companies and shareholders.

More recently, Restrepo has begun exploring a more speculative frontier: what would happen if artificial intelligence eventually became capable of performing essentially all economically valuable work?

In a recent paper, We Won’t Be Missed: Work and Growth in the AGI World, Restrepo considers an economy where highly capable AI systems can perform nearly any task. In that scenario, the key constraint on economic growth may no longer be human labor but computing power—the hardware required to run those systems.

In that world, Restrepo argues, the structure of the economy could change dramatically: growth would be driven by the expansion of computing resources, and a growing share of income could accrue to the owners of those resources rather than to workers.

“Humans have historically been the bottleneck for production. When you have systems that can perform essentially any task, humans are no longer the bottleneck—you scale the economy by adding compute.”

In a recent conversation, Restrepo discussed how economists think about automation, what makes AI different from earlier technologies, and how researchers might eventually measure its impact on the labor market.

Q: At the highest level, how do you describe your research agenda?

Restrepo: There’s a big core question about trying to understand the ways in which technology shapes society and the economy. I’ve been specializing mostly in understanding how technology transforms labor markets, how it transforms the functioning of the economy, and how it affects standards of living.

There’s a lot of interest in these questions today because we have AI, and everyone is trying to understand how AI is going to transform labor markets and productivity.

Q: You started much of your work studying automation. What do economists mean by automation—and how does that research connect to today’s debates about AI?

Restrepo: I like to think of it as bringing an engineering perspective to what production is. Imagine a manufacturing plant: you have many processes or tasks, and you decide how you’re going to accomplish each process. You can start with everything manual, with workers at each step.

Automation is when you decide to complete a particular task with a machine or a specific piece of software designed for that process because it’s cheaper. If I need to do welding, what’s cheapest: hire workers, or buy a robot that can do spot welding?

Automation happens when the machine outcompetes the worker at that particular task.

Q: Why does it matter to think about automation at the "task level" — and what does that framework reveal about what makes AI different?

Restrepo: It’s very important to think of automation as something that happens at the task level, not the worker level because it’s not that the robot took over all factory jobs. It takes over some components. Then workers do other components in the process that weren’t automated. These things aren’t substituting for labor as a whole. They’re substituting for labor in particular pockets of the economy.

With respect to AI—in earlier waves of technological change, engineers or software developers created machines that were coded to solve specific tasks. Because you need to code it yourself, the tasks you automate are highly codifiable—tasks where you can write a set of instructions.

With AI, we’re not coding systems in a traditional way. We’re not writing instructions for them on how to write. They learn to write by looking at human examples. That means these systems can develop capabilities in tasks that aren’t necessarily codifiable—tasks where humans rely on experience, learning, or tacit knowledge that you can’t put on paper.

Q: Do you believe that AI could eventually automate nearly everything?

Restrepo: As a statement of principle, I believe it. In practice, it’s hard. But in principle, if you can track what a worker does for long enough, eventually you learn to do what they’re doing. If I can observe enough examples and have enough computing power and hardware, eventually I can replicate those capabilities.

Q: In one of your recent papers you imagine a world where output depends on “compute” rather than human labor. What does that mean?

Restrepo: The premise is that you develop systems capable of human intelligence at essentially every task. But to operate those systems you need computing power. Compute is the input that allows the technology to run.

Historically, humans have been the bottleneck for production. If you want to scale the economy, you need more people or more productive people. When you have a system capable of doing all those tasks, humans are no longer the bottleneck. You can scale the economy by throwing more compute into the system. At that point the limiting factor becomes computing power.

Q: There are many predictions right now about AI wiping out jobs. How should we think about those claims?

Restrepo: A lot of these claims are premature. Many of the people making them have an interest in selling a product, so you have to take them with a grain of salt.

It’s also hard to evaluate these claims empirically right now. When you look at the data, adoption of AI by firms is still relatively low. Right now I would describe the labor market as more of a wait-and-see environment. Firms are experimenting and trying to figure out how to use AI.

Q: Even if the technology becomes powerful, why might the transition be slower than people expect?

Restrepo: There are cultural forces and institutional forces. Even if you develop AI that can do something like teaching at a certain cost, parents and students may still prefer human teachers for a while.

But I think the compute constraint is actually underappreciated. You cannot automate everything tomorrow because we simply don't have enough computing power to do it. The infrastructure has to be built incrementally. Companies build more as demand goes up, then build more again. It's not like flipping a switch. And if AI does become as capable as people imagine, the most valuable uses of that compute won't be automating paralegals or economics professors. It'll go toward curing diseases, solving climate change, advancing science. So even with a very powerful system, there are competing priorities for finite compute that slow down the displacement of any particular category of work.

Q: Eventually, how will economists measure the labor-market impact of AI?

Restrepo: In the past we measure the capabilities of the technology, map occupations into tasks, and identify which tasks the technology performs well. Then we look at employment and wages across occupations and industries that are exposed to those tasks. You can imagine doing the same thing with AI—identify which jobs are exposed and then examine whether employment opportunities are deteriorating in those parts of the economy.

Q: One last question. When people hear that AI could lower wages, they often assume that means living standards fall. Is that necessarily true?

Restrepo: People have the wrong intuition when they say that if AI can do my job for ten dollars an hour, then my wage falls to ten dollars and my life is terrible.

A world where AI can do research or teaching at that cost is a world where AI is extremely capable and can produce many other goods and services cheaply. What matters is not the dollar wage but what you can buy with it. If technology lowers the price of many things, purchasing power can increase even if wages fall. Where the pessimistic story can be more accurate is when technology only affects a narrow set of jobs. In that case wages fall for that group while prices elsewhere do not fall.