AI is changing how people work, and it is introducing a lot of uncertainty for anyone in the workforce. This is not the first time a technology has forced large numbers of people to change jobs. It has happened repeatedly, and the most recent case, industrial robots on the factory floor, was studied closely by economists for the better part of 25 years.

This post walks through that robot research and asks how much of it carries over to AI. The way the damage lands looks likely to carry over. Who it lands on does not. Robots climbed down the education ladder and fell hardest on high-school dropouts. AI exposure climbs up it, rising with skill and income. The first chart sets up that contrast; the rest of the post is the evidence behind it.

What robots did to jobs

The canonical study is Acemoglu and Restrepo’s “Robots and Jobs,” which tracked US commuting zones from 1990 to 2007. It is the most rigorous natural experiment available for a machine standing in directly for a worker:

“One more robot per thousand workers reduces the employment-to-population ratio by 0.2 percentage points and wages by 0.42%.” (Robots and Jobs, JPE 128(6))

That is the aggregate, national figure. Translate the coefficient into people and it lands harder: within a local commuting zone, the same paper estimates one more robot reduced employment by about six workers.

Three features of what robots did matter for the AI comparison:

• Occupation-specific. The hit fell on routine manual work: machinists, assemblers, material handlers, welders.
• Geographically concentrated. Detroit was the single most-exposed commuting zone, with the other most-exposed zones clustered in the industrial Midwest.
• It climbed down the ladder. Every education group lost ground, but workers without a college degree lost far more, and robots showed no positive complementarity even for workers with advanced degrees. Earlier IT waves had lifted the highly credentialed. Robots did not.

One caveat. The sharp negative story is most robustly a US result. Graetz and Michaels (2018) find robots raised productivity and wages across countries while reducing lower-skilled workers’ hours, and the German study by Dauth and coauthors found “no evidence that robots cause total job losses” in Germany, only a reshuffling of job composition (manufacturing losses fully offset by new service-sector jobs), plausibly because stronger unions and a denser safety net absorbed the shock. The policy section returns to that.

For many, the loss was permanent

The aggregate story is mild. Total US employment moved only a small share. That average hides the thing that actually matters to a displaced person, which is that for a large minority the loss never reverses.

The scarring literature measures this. These studies are not about robots. They follow workers who lost long-held jobs in mass layoffs, which is the same event a robot triggers for the worker, and they are used here as the benchmark for what losing a stable job does. The foundational study, Jacobson, LaLonde and Sullivan (1993), found high-tenure workers displaced in the early 1980s still down about 25% of prior earnings five years out (high-tenure workers were chosen because their long earnings histories make the loss easy to measure, not because seniority caused it).

Two later studies traced the longer path, and they anchor the chart below:

• von Wachter, Song and Manchester (2009): displaced workers were still about 20% below comparable peers 15 to 20 years later.
• Couch and Placzek (2010): in administrative data, about 33% down immediately, and still 15% down six years out.

The mechanism is the surprising part. Most of the lifetime cost does not come from a lower wage on the next job. It comes from the trouble finding stable full-time work at all: longer spells out of work, more part-time landings.

That is also why “reemployment” is a misleading headline. The BLS Displaced Workers Survey (January 2024) found 65.7% of long-tenured displaced workers reemployed, which sounds like recovery until you open the number up.

Open that number up and it splits into three outcomes. About a third of displaced workers never got reemployed at all. A quarter came back only at lower pay. And about four in ten found full-time work that matched or beat their old pay. Stack those and the friendly-sounding two-thirds reemployment rate is hiding a long tail of people who are worse off for good.

The damage does not stay inside the pay stub. Displacement is linked to worse physical and psychological health, family disruption, higher divorce rates, and worse outcomes for the workers’ children. And it moves votes. One study found that a single robot per thousand workers cut local average “career value” by about $3,900 in 2004 to 2008, and that automation-hit regions shifted toward populist and Republican candidates in 2016. Displacement is a political event, not only an economic one, and that is where the AI story diverges most.

AI hits the opposite people

Everything above transfers to AI as mechanics: concentrated, permanent for a minority, and hard to undo once it lands. What does not transfer is who stands in the blast radius.

By education the pattern is a clean inversion, the first chart above. Robot exposure was strongest among high-school dropouts and declined monotonically with education; AI exposure ran the other way, lightest on the least-educated and heaviest on college graduates.

Cut the same UK data by wage instead of education and the robot curve changes shape. It is not a straight line down. Robot exposure is an inverted-U, lowest at both ends of the wage distribution and concentrated in the middle, because low-wage work is often too unstructured to automate cheaply and high-wage work too non-routine. AI exposure, by contrast, climbs monotonically up the wage distribution, toward the credentialed knowledge work at the top.

This is the reversal at the heart of the comparison. Every earlier wave of automation, from the power loom to the industrial robot, pushed hardest on manual and lower-paid work. AI is the first to point up the income ladder, toward the credentialed knowledge work long assumed to be automation-proof: software, law, finance, analysis, design. Set what robots did beside the exposure curve and the implication is uncomfortable. Displacement is harsh, concentrated and permanent for a minority, and this time that minority sits nearer the top of the distribution than the bottom.

The optimistic case

The exposure curve shows which jobs AI can reach, not how many workers it will replace, and there is a more hopeful reading of the same data. Autor’s (2024) hypothesis is that AI might reverse the wage polarization that robots and earlier IT produced: because it can put expert judgment within reach of less-credentialed workers, AI could “extend the relevance, reach, and value of human expertise” rather than hollowing out the middle. Whether that happens is unsettled. What is already clear is the direction of exposure, and it points up the ladder, not down it.

Workers have fought this before

Yes, repeatedly, for two centuries. The pattern is remarkably stable: labor bargained over the terms of the transition, occasionally struck over its pace, and rarely stopped the technology. It did best when it had structural power, not when it was left to help individual workers adjust after the fact.

What is genuinely new is a shift from bargaining over automation to demanding it be blocked outright, now aimed at AI as much as at machines:

• Driverless trucks (2023). The Teamsters and UPS contract banned them for the length of the contract.
• Port automation (2024). The International Longshoremen’s Association threatened to shut East and Gulf ports over a demand for a total ban on automating cranes, gates and container moves, framing full automation as existential rather than tradeable.
• AI in Hollywood (2023). The Writers Guild won terms that bar AI from writing or rewriting scripts, and SAG-AFTRA secured rules requiring a performer’s consent and compensation before AI can recreate their voice or likeness.
• A pause on AI itself (2023). An open letter signed by Elon Musk and thousands of others called on all AI labs to pause training systems more powerful than GPT-4, citing risks from mass job automation to a loss of human control.

All that resistance raises a question: did it change the outcome, or was the displacement coming either way? Two readings of the evidence pull in opposite directions:

• Bargaining power, not machines (EPI). The Economic Policy Institute, a labor-aligned think tank, argues there is no good evidence that automation causes mass joblessness, and that the real driver of stagnant wages is the erosion of worker bargaining power. It nicknames the opposite view (that robots are destroying jobs) the “zombie robot” argument, a claim it says keeps lurching back despite being debunked.
• The displacement was real (Acemoglu and Restrepo). Their evidence shows the job and wage losses were measurable and real regardless of how well workers bargained.

Both can be true: bargaining power shaped how much of the loss workers absorbed, and the loss was still there to absorb.

What actually softened the blow

What actually helped the workers who were displaced? Two responses get tried after a wave of displacement. The first is to bring the lost jobs back. The second is to cushion the people who lost them with income support. With robots, the first did little for the workers themselves, and the second is the one the evidence backs.

Bringing the jobs back

One way robots can bring jobs back is through reshoring, moving production home from lower-wage countries. Krenz, Prettner and Strulik (2021) find that robots drove more of it: one extra robot per thousand workers is associated with about 3.5% more reshoring. But the production came back automated.

Average pay can rise in the process, which sounds like good news until you see where it lands. The gains went to professional occupations, not to the routine workers who had been displaced, and Krenz and coauthors tie automation-driven reshoring to a rising skill premium and wider inequality, not a broad-based raise. Higher average wages, a bigger gap.

The same pattern is starting to show up in the AI wave, aimed this time at offshore service and software work rather than factory production. Jobs that companies once sent abroad, in call centers, back offices and IT, are now being automated. The US firm Opendoor is closing its India operations and replacing those offshore workflows with AI run by smaller US teams; India’s IT and business-process sector, which employs millions of people, is bracing for the same substitution, and it reaches past support desks into software itself, where entry-level coding jobs are among the first being cut. The work that offshoring created may be exactly what AI is positioned to absorb.

The safety net

That leaves the safety net, and it is where the evidence is strongest. The most durable cushion was income support: money that softened the fall, paid over the years a displaced worker actually spends looking for stable work.

The clearest evidence comes from an IMF analysis of the robot era: across US commuting zones in 2000 to 2007, the wage decline from robot adoption was about two-thirds smaller in states with more generous unemployment insurance than in stingier ones. The cushion showed up in wages, not employment, and was strongest for workers without a college degree. The mechanism, in the IMF’s reading, was time. The old job was gone either way, but longer-lasting benefits let displaced workers search harder for a job that matched their existing skills, or reskill for a different one, instead of grabbing the first lower-paying opening. Better matching and time to reskill left their wages higher.

The German robot story points the same way, where a denser safety net went with a wave of automation that did not turn into net job losses. With robots, income support did more to cushion displacement than bringing the jobs back did. Whether that carries over to AI has not been tested.

What robots can & can’t tell us about AI

What robots did is the closest precedent we have, but it has to be read with care. The authors of the AI-exposure-index paper flag the limit directly:

“The potential difference in the nature of robots (tangible, rivalrous assets) compared to AI algorithms (intangible, non-rivalrous) remains a limitation of the external validity. Therefore, AI might affect wages and unemployment differently from robots. This distinction is crucial in avoiding false analogies about the effects of AI on the labor market.” (AI-exposure-index paper)

With that caveat, it still says something. Displacement, when it lands, tends to be concentrated, permanent for a sizeable minority, and only lightly softened after the fact; the clearest exception to the pattern is who is exposed, since robots fell on the least-educated while AI exposure climbs the wage and education ladder; and across the robot era the intervention that cushioned the loss best was income support, not the jobs coming back.

The two groups also differ in how they can respond. Robots hit blue-collar manufacturing workers, who had little individual clout but strong union structure, and could still strike a plant or a port. AI is reaching white-collar workers, software engineers, designers, lawyers, analysts, who have the opposite mix: more visibility, but almost no union structure and little collective leverage over how the technology gets deployed. The historical lesson from earlier waves, that labor did best when it had structural power, lands on a workforce that has less of it.

How all of this resolves is genuinely uncertain. AI could hollow out the top of the ladder the way robots hollowed the middle, or it could, as Autor hopes, extend expertise downward and lift more people than it displaces. The direction of exposure is clear; the effect on jobs and wages is not yet in. What robots did tells us how a displacement wave has played out before. Whether AI follows it, we will find out by watching.

Sources

• Acemoglu and Restrepo, “Robots and Jobs: Evidence from US Labor Markets,” Journal of Political Economy 128(6), 2020. (PDF)
• Dauth, Findeisen, Sudekum and Wossner, “German Robots: The Impact of Industrial Robots on Workers,” IAB Discussion Paper 30/2017. (PDF)
• Graetz and Michaels, “Robots at Work,” Review of Economics and Statistics 100(5), 2018. (PDF)
• Jacobson, LaLonde and Sullivan, “Earnings Losses of Displaced Workers,” 1993. (Chicago Fed PDF)
• von Wachter, Song and Manchester, “Long-Term Earnings Losses due to Mass Layoffs During the 1982 Recession,” 2009.
• Couch and Placzek, “Earnings Losses of Displaced Workers Revisited,” American Economic Review 100(1), 2010. (AEA page)
• BLS, “Worker Displacement: 2021-2023,” January 2024. (release)
• Labaj, Oles and Prochazka, “Impact of robots and AI on labor and skill demand: evidence from the UK,” Eurasian Business Review, 2025. (Springer)
• Schaal, “A theory-based AI automation exposure index: Applying Moravec’s Paradox to the US labor market,” arXiv:2510.13369, 2025. (PDF)
• Autor, “Applying AI to Rebuild Middle Class Jobs,” NBER Working Paper 32140, 2024. (PDF)
• Petrova, Schubert, Taska and Yildirim, “Automation, Career Values, and Political Preferences,” NBER Working Paper 32655, 2024. (Knowledge@Wharton summary)
• Krenz, Prettner and Strulik, “Robots, reshoring, and the lot of low-skilled workers,” European Economic Review 136, 2021. (EconStor)
• Bivens (EPI), “The zombie robot argument lurches on,” 2017. (PDF)
• City Journal, on containerization and the long decline of longshore work. (article)
• Brollo, Dabla-Norris et al., “Broadening the Gains from Generative AI: The Role of Fiscal Policies,” IMF Staff Discussion Note SDN/2024/002, 2024. (PDF)
• Bipartisan Policy Center, “What Happens When Jobs Disappear,” 2025. (explainer)
• “The Far-Reaching Impact of Job Loss and Unemployment,” on displacement and family/health outcomes. (PMC)