Computer scientist Konrad Kording and economist Ioana Marinescu have developed an interactive model that incorporates assumptions from both their fields to predict how AI will affect wages, jobs, and the overall economy.
4 min. read
Key Takeaways
Tasks that rely on intelligence will be automated first, pushing human workers toward more physical jobs.
Wages may rise at first as AI boosts productivity, but they may fall later as automation spreads and reshapes the kinds of work people do.
Once automation is widespread, continued expansion of AI raises wages and output to a point, but levels off.
The ultimate constraint on economic growth shifts to physical production with increasing AI.
Artificial intelligence (AI), once regarded as science fiction, is now a part of everyday life, with people turning to ChatGPT and other AI assistants for help with everything from drafting emails to planning vacations. But as AI spreads, fundamental questions on how this technology might reshape the nature of work loom.
Economists foresee slow, steady changes, viewing AI as another capital investment to help workers be more productive. By contrast, tech experts predict explosive growth—“a singularity”—when humans are replaced by AI.
Now, Penn’s Konrad Kording, a computer scientist, and Ioana Marinescu, an economist, have developed an interactive model that incorporates assumptions from both camps to generate meaningful predictions about how AI will affect wages, jobs, and the overall economy. Their work is published in Brookings.
Image: Eric Sucar (left) and Carson Easterly (right)
“Economists and computer scientists have been talking past each other on this for years,” says Penn Integrates Knowledge (PIK) University Professor Kording. “The model gives them a shared language to make their disagreements explicit.”
The model expands the traditional framework of dividing the economy into labor and capital, further categorizing each into two sectors: intelligence tasks and physical tasks. This allowed the researchers to capture the effects of the slow-moving physical piece and the fast-moving virtual intelligence piece—AI—on the future of the economy.
What their model demonstrated was how intelligence saturation—the idea that additional increases in intelligence yield diminishing benefits when physical inputs remain constant—fundamentally constrains AI’s long-term economic impact.
“If I give you a task to restack all of the books onto the table, it is going to take you a certain amount of time,” explains Kording, a professor of neuroscience in the Perelman School of Medicine and of bioengineering in the School of Engineering and Applied Science. “Imagine that you are more intelligent than you already are—could you restack the books faster? Probably. But there’s a certain minimal time that it takes you to restack the books with your hands. That’s intelligence saturation.”
Kording and Marinescu used their model to simulate different scenarios and found that the effects of automation on wages depended on different assumptions and parameters.
In the first simulation, they wanted to understand what happens in the short run in the event of rapid automation.
“In this thought experiment, AI is not growing, but we have a ton of it,” says Marinescu. Initially, wages increase—AI is able to do more intelligence tasks faster than the humans it replaced, increasing economic output overall. “And as it replaces people in intelligence jobs,” she continues, “people move to physical jobs.
“If the economy is growing, everybody benefits, including workers,” says Marinescu.
But this reallocation to physical jobs may then reduce wages as labor supply increases relative to demand, which is the equivalent, explains Marinescu, of having more and more drivers for the same number of cars—particularly if there is a high degree of “substitutability” between the two sectors.
In this scenario, “what you need is complementarity,” she says. “As the intelligence sector does better, it brings along the physical sector because they are strongly coupled, and that prevents workers’ wages from declining.”
But what about a scenario in which AI is allowed to grow quickly after full automation? Now complementarity is not helpful, says Marinescu. In this scenario, “wages and output growth eventually hit a strict plateau—they just flatten out. As you add more and more AI, it just saturates, and there is nothing more to be gained,” she says.
On the other hand, if physical and intelligence sectors are perfectly substitutable—meaning that final production could come entirely from one sector or the other—then wages and output continue to grow because they are not constrained by the fixed amount of physical input—a scenario that, they say, would be unambiguously good for wages but not so much for people.
“At this high level, they’re not very substitutable,” says Kording. “Until we move into cyberspace and stop living in the external real world, we kind of need actual stuff.”
So what does this mean overall? According to Kording and Marinescu, their findings contradict the vision of the singularity narrative, where increasing intelligence yields limitless gains. Yet they acknowledge AI has the potential to reshape the economy.
“AI is going to make some people jobless,” says Kording. Although this is likely to happen somewhat slowly, their findings have real policy implications, he notes.
“There is going to be pain in the transition. Some people will be affected by technology, and they will have to switch their jobs,” says Marinescu. “Historically, we know it can be hard.”
Slowing the pace of automation, they say, could help mitigate potential wage losses—albeit at the cost of reduced economic growth. This would also allow for more time to make necessary investments in physical capital that could boost workers’ productivity in the physical sector, propping up their wages as they transition into these jobs.
“If we are moving towards doing in-person tasks,” says Marinescu, “we’ll need actual buildings—coffee shops, restaurants. If all the virtual jobs are mostly going to be gone, you won’t be able to just sit in your house.”
Nanoparticle blueprints reveal path to smarter medicines
New research involving Penn Engineering shows detailed variation in lipid nanoparticle size, shape, and internal structure, and finds that such factors correlate with how well they deliver therapeutic cargo to a particular destination.
A generous gift from alumni Glenn and Amanda Fuhrman brings the work of internationally acclaimed artist Jaume Plensa to the University of Pennsylvania. The latest addition to the Penn Art Collection expands Philadelphia's public art.
A massive chunk of ice, a new laser, and new information on sea-level rise
For nearly a decade, Leigh Stearns and collaborators aimed a laser scanner system at Greenland’s Helheim Glacier. Their long-running survey reveals that Helheim’s massive calving events don’t behave the way scientists once thought, reframing how ice loss contributes to sea-level rise.
