The push to reskill workers displaced by AI could backfire

AI is reshaping the workforce.

By 2030, the World Economic Forum (WEF) projects that 170 million new jobs will be created and 92 million lost, as AI and other economic shifts take hold. About 54% of global executives expect AI to displace existing jobs at their companies, according to the WEF’s annual survey of over 10,000 leaders worldwide.

Faced with that expectation, many companies are pouring money into retraining programs. In theory, this makes sense. Helping workers learn new skills when their jobs disappear ought to balance out the labor market.

But new research by Professor Seyed Morteza Emadi at UNC Kenan-Flagler Business School suggests that that logic can break down in practice.

“Retraining has limited capacity — the system can only handle so many people at once,” he says. “There are only so many seats at universities, so many instructors and so many slots in certification programs that companies or schools can realistically roll out. If you add capacity in the wrong place or faster than the rest of the system can keep up with, it can backfire.”

Emadi calls this the “capacity-adoption paradox.” As more workers leave their jobs to retrain for new ones, the remaining workers become scarcer and more expensive to employ. This pushes companies to automate those roles even sooner and accelerates the very losses it’s meant to ease.

The paradox does not show up everywhere. It happens when three things line up at once:

• The displaced workers all need the same retraining path
• The technology is far enough along that a small rise in labor costs is enough to speed up a rollout
• There are already too few people left doing the work by hand

When any one of those is missing, adding capacity does exactly what it is supposed to do.

Emadi’s research offers leaders three tools:

• A way to identify bottlenecks in the retraining system
• A mechanism for testing whether more capacity helps or hurts
• A diagnostic test to know when to slow deployment if needed.

He writes about his findings in the new working paper “The Capacity-Adoption Paradox: Managing Workforce Bottlenecks During Rapid Technology Development.”

The induced demand problem 

Emadi’s interest in this topic comes from a few different vantage points. As an expert in operations management, he thinks in terms of resources and what they cost. But he’s also a professor teaching undergraduate and MBA students — and the father of a young daughter. He sees the subject from both the perspective of his students who are on the cusp of entering the job market and his child who’ll be joining it years from now.

“I can see the looks of concern on my students’ faces when we talk about AI,” he says. “They’re excited to learn what these tools are capable of, but there’s a lot of anxiety, too. They ask me, ‘If AI systems can already do this stuff, what’s left for us?’ Which gets me thinking about my daughter’s future and what’s going to happen to her.”

Emadi built a model to find out. He mapped the retraining system as a network of pathways, where seats, instructors and certification slots each limit how fast workers can move through it. He coupled that map with the rate of AI adoption, showing that as the supply of labor declines, companies fast-track AI.

Highway traffic is a real-world parallel. Consider Houston’s $2.8 billion dollar expansion of Katy Freeway. Between 2003-2008, officials widened the Katy to 26 lanes to reduce traffic. But instead, traffic got worse. Data show that morning commute times have risen by 25 minutes and afternoon commute times have increased by 23 minutes.

Economists refer to this as induced demand. Making something more convenient, cheaper or more available leads people to use more of it, says Emadi. “When you increase the number of lanes, people’s behavior changes. They work differently or they live farther away from their jobs or they just drive more often.”

The same dynamic applies to retraining, says Emadi. Adding more training capacity can result in more job losses because it changes how companies behave.

“The shrinking pool of workers creates a temporary labor shortage forcing wages up,” he says. “And when it’s more expensive to keep humans in their jobs, there’s greater incentive to automate faster. Leaders don’t blink an eye to do that.”

The case for pacing AI deployment

Meanwhile, workers get stranded between two jobs: no longer needed in the one that’s being automated away — and not yet trained for the one that’s meant to replace it.

Continuing the traffic analogy, Emadi likens the situation to a city during rush hour. People with old jobs disappearing are on one side of town and new jobs opening on the other, with retraining programs as the roads and bridges for workers to cross. “Not every bridge is jammed, so the solution isn’t to widen every single road,” he says. “It’s to figure out which bridges and which roads need more capacity.”

The goal, he says, is to help business leaders spend their retraining budgets in smarter, more calibrated ways, so workers can transition into new roles. This requires locating the chokepoints in their systems, measuring demand for new roles and being more deliberate about who trains for which ones.

Plumbers have skills that make them well positioned for jobs at data centers doing ducting and cabling, he says. Coders doing simple work could be trained to move into programming roles. And customer support employees can upskill to work in more complex roles with AI assistance.

Still, even the right pathway can get overwhelmed if too many workers try to cross it at once.

That’s why Emadi makes the case for pacing AI deployment. Since companies are unlikely to slow down voluntarily, he argues that policymakers should consider a levy on companies whose AI rollouts put the most pressure on the retraining system. This tax would be similar to a congestion charge during rush hour traffic.

The same logic works inside a single company. A central workforce office can slow its own rollouts through internal approval hurdles, quotas or sequencing, without any money changing hands.

“When tech leaders talk about AI, they often speak in loose and hand-wavy terms. ‘AI is coming for your job, but you can still fight back and the best protection against displacement is to transform yourself,’” he says. “But this issue is complicated and the responsibility for reskilling can’t fall on workers alone. We need to be smart about it. We need a national policy and strategy. And companies need calibrated plans to retrain people.”