Who’s losing their jobs to AI

Morgan Frank's job usually has him working with reams of data and answering big questions - questions like how technology influences the economy or what the future of work will look like. But early in the pandemic, rather than swimming in spreadsheets, you'd more likely find him calling up an obscure statistician at a state unemployment office.

The result of that decidedly low-tech effort may help solve one of today's most anxiety-inducing mysteries: Is AI really taking people's jobs? And if so, whose?

"This is the million-dollar question," said Frank, an assistant professor in Pitt's School of Computing and Information. "Economists, technologists, people in industry, people in academia: Everyone wants to study this question."

Even before large language models started taking over the world, it was a question that caught Frank's interest. An expert on skills, automation and work, Frank first published on the labor impacts of AI in 2019, years before ChatGPT was made available to the public. More recently, he turned his attention to the skills and employment of fossil fuel workers, another area where the march of technology is colliding with an existing workforce.

And if this latest generation of AI represents a unique disruption to the labor market, then that also makes it a unique opportunity for labor researchers - a survival-of-the-fittest moment for economic theories about skills and employment. So it was only natural that Frank would be first in line to study it.

Two sides, little clarity

If you're confused about whether AI is causing job losses, you're not alone. You can find experts in respected news outlets arguing either that all white-collar jobs are under threat or that the concerns are overblown. Tech CEOs cite AI as a cause for layoffs, while others cry "AI-washing" - suggesting leaders are blaming the technology for staff reductions that would have happened anyway.

"How can these two narratives be so prominent, and how can there be this confusion?" Frank asked.

The likely problem, he explained, is a lack of data: If AI is having an impact, then we may just not be equipped to see it. Any losses would likely be concentrated in job types where AI directly competes with employees, so to spy an effect, it isn't enough to look at the overall economy. You'd need to see changes across categories that are finer than those in the government statistics usually used by economists.

[Read more about Frank's work in the School of Computing and Information.]

One place that would have that data, however, is a state unemployment agency.

"If my job is to help you find your next job, the very first question I'm going to ask is, 'What were you doing before?'" Frank said. "And of course, states are asking this, so it was a matter of trying to figure out where we could get that data from." Prominent labor economists didn't know: One Nobel laureate even told Frank it was a "silly question." If no such stream of data existed, Frank realized, he would simply have to claw one together one himself.

That's what led to him doing it the old-fashioned way. He and his team called state offices to track down who was keeping the right records and convince them to make the data public. Combining those sources gave them what they needed.

"Tell me where you are, tell me the month, tell me your job, and I'll tell you the probability that you claim unemployment," Frank said. "This data is super valuable and super new."

The team then used those numbers to test several different models: ideas about which kinds of jobs should, in theory, be most vulnerable to AI. They found models disagreed across many areas, and each alone didn't do a good job of predicting job loss to automation. But when combined into an "ensemble," they could predict close to 20% of changes in employment. The team published the results in the journal PNAS Nexus in 2025.

The data may help researchers better tune those models of AI job competition, and its potential goes far beyond looking at the particular effects of this technology. Having an eye on job loss in particular areas and industries could, for instance, help policymakers better target interventions and policies meant to benefit specific types of employees.

"I'm hoping one of the long term outcomes of this is to show its usefulness for not just our use cases, but that other researchers will use it for their own things they'd like to study," Frank said.

Linking up LinkedIns

Unemployment is just one way Frank is looking at AI, especially because the technology will likely have many more subtle effects, like how it changes the way people do their jobs or how long it takes them to find one.

A collaboration with Microsoft's AI for Good initiative gave Frank access to millions of LinkedIn profiles, allowing him and a team of researchers to study the post-graduation experience of job-seekers. 2022, the same year ChatGPT launched, appears to be an inflection point. Recent graduates saw a 16% loss in salary, greater than the overall rate of 7%, and recent graduates in AI-exposed fields spent almost a month longer job-hunting than their peers.

The only catch? The changes started in early 2022, while ChatGPT didn't launch until November of that year.

"Unemployment risk rose, especially for AI-exposed workers around the launch of ChatGPT, but it started before ChatGPT was widely available to the public," Frank said.

Rather than being a result of AI, Frank's best guess is that the shifts have to do with interest rates - rising rates made tech startups a less attractive investment, and they had to shed workers to stay sustainable. So the economy is showing mixed signals, but it's early days yet. Frank further plans to combine his data with a database of more than 3 million course syllabi to see how the AI skills students learn affect their job search. It's a topic close to Frank's heart, as an avid user of the technology for writing, coding and more.

The usefulness of AI is clear to Frank - and even if the broader impacts of the technology are less-than-clear now, he expects that they'll eventually be impossible to ignore. When they are, he'll have his finger on the pulse.

"I think that we're still in this transition period where we have this new tool that I think can be used in lots of different ways, but we need it to be embedded into the interfaces we already use," Frank said. "Thats when we'll start to see broader and broader changes in the economy and in the nature of different jobs."

Photography by Tom Altany