AI destroyed 200k to 300k jobs in 2025 in the US

I used two different methods, and four different AIs, to help crunch these numbers

The U.S. economy grew at a healthy clip in 2025. GDP came in at 3.8% annualized in Q2 and 4.4% in Q3. Corporate profits were up. Consumer spending held. By most conventional indicators, this was a good year.

And yet payroll employment expanded by only 584,000 jobs. That’s the weakest annual pace since the pandemic year of 2020 and barely a third of what the Philadelphia Fed’s Survey of Professional Forecasters had expected. Employers announced 1.2 million planned layoffs while simultaneously posting the fewest hiring plans since 2010. By December, job openings had fallen to 6.5 million, a level not seen since September 2020.

Something broke the relationship between output and employment. The economy kept growing; it just stopped needing as many people to do it.

The obvious suspect is artificial intelligence, which underwent rapid enterprise adoption throughout 2025. Roughly 92% of Fortune 500 companies reported using generative AI in some capacity. But here’s the problem: nobody is actually measuring AI’s effect on employment in any systematic way. The Bureau of Labor Statistics tracks jobs and productivity with impressive rigor, but it doesn’t attribute changes to specific technologies. Challenger, Gray & Christmas compiles layoff announcements and began tracking AI as a cited reason in 2023, but employers have every incentive to avoid that label. The total number of jobs where companies actually said “we cut this position because of AI” in 2025 was 54,836.

That number is almost certainly wrong. Not because Challenger miscounted, but because it relies on employers volunteering information they’d rather not share. “Restructuring” doesn’t invite congressional hearings. “Replaced by artificial intelligence” does.

So we tried to estimate the real figure using two independent methods, each based on established analytical frameworks, each relying on different data, and each designed to detect AI displacement through its downstream effects rather than through employer self-reporting.

They both landed in the same neighborhood. That’s the finding that matters.

You can download the full report here:

Ai Job Impact 2025 V2

239KB ∙ PDF file

Download

Download

The estimate

Both methods point to roughly 200,000 to 300,000 U.S. jobs displaced or foregone due to AI in 2025. That’s about 0.13 to 0.20% of total nonfarm employment. It’s four to six times larger than what employers explicitly attributed to AI, which confirms what most labor economists already suspected: the official count captures maybe a quarter of the actual effect. The rest hides behind euphemisms, or never shows up in layoff data at all because it takes the form of a job that used to exist, someone quit, and no replacement was ever hired.

How we got there

Method 1: Excess layoffs

The first approach borrows from epidemiology. During COVID, one of the most reliable ways to measure the pandemic’s true death toll was to compare actual deaths against a baseline of what you’d normally expect, then examine the excess. The same logic works for layoffs. You establish what a “normal” year of job cuts looks like given the economic conditions, measure what actually happened, and try to explain the gap.

We built the baseline from Challenger’s pre-pandemic data. The average across 2017, 2018, and 2019 was about 517,000 announced cuts per year. Adjusted slightly upward for the softer growth environment of 2025 versus the late-2010s boom, we settled on 525,000 as our expected figure for a non-recessionary year with GDP growth around 2 to 2.5%.

The actual number was 1,206,374. But you can’t just subtract and call the difference “AI.” That’s where at least one alternative analysis went wrong.

2025 had a massive, unusual shock that had nothing to do with technology: the federal government’s DOGE initiative, which drove 293,753 directly announced cuts plus another 20,976 in downstream effects on contractors and nonprofits. Another 7,908 cuts were explicitly tied to tariffs. All of that has to come out before you start attributing anything to AI. (One prominent alternative analysis failed to make this subtraction, which inflated its estimate by roughly a factor of two.)

After removing policy shocks, the adjusted total drops to about 884,000, which gives you roughly 359,000 excess layoffs against the baseline.

From there, the work becomes a process of elimination. How much of that excess can be explained by things other than AI? The technology sector has been shedding jobs since late 2022, part of a multi-year correction from pandemic-era overhiring. That probably accounts for 50,000 to 70,000 of the excess. Elevated interest rates continued to drag on housing and construction, good for maybe 20,000 to 40,000. Retail restructuring, consumer uncertainty, the ongoing migration to e-commerce: another 30,000 to 50,000. Broader tariff uncertainty beyond the explicitly cited cuts: 30,000 to 50,000 more.

Add it up and you can explain about 130,000 to 210,000 of the excess through non-AI channels. What’s left over is 150,000 to 230,000 layoffs that don’t have a clean explanation, in sectors where AI adoption has been heaviest, during a year when companies were publicly announcing their enthusiasm for replacing workers with algorithms while privately labeling the cuts as “restructuring.”

Method 2: Missing millions

The second approach comes at the problem from a completely different direction. Instead of counting eliminated positions, it starts with a basic macroeconomic identity: growth in hours worked roughly equals growth in output minus growth in productivity. If productivity rises faster than its historical trend while output grows normally, then employment has to grow slower than you’d expect. The gap represents jobs the economy no longer needs.

The BLS reported that from Q3 2024 to Q3 2025, nonfarm business productivity grew 1.9% year over year, output grew 2.8%, and hours worked grew just 0.9%.

The tricky part is choosing what to compare that 1.9% against. The prior business cycle (2007 to 2019) averaged 1.5% annual productivity growth. The current cycle (2019 to present) averages 2.0%. This matters more than it might seem. If you use the older baseline, you get a 0.4-percentage-point “excess” in productivity growth that translates to about 544,000 job-equivalents of reduced labor demand. If you use the current-cycle average, the excess essentially vanishes. One of the alternative analyses we examined used the prior-cycle baseline without flagging this sensitivity. Another correctly identified the problem but then arguably overcorrected by treating the current-cycle average as definitive.

We split the difference with a blended baseline of 1.7%, producing an estimated 272,000 jobs worth of reduced labor demand from above-trend productivity growth. That’s our starting point, not our answer.

We also ran two broader calculations. A GDP-employment elasticity analysis, which asks how many jobs the economy should have created given its output growth based on historical ratios, suggested about a million “missing” jobs. A comparison against professional forecasters’ expectations yielded a shortfall of 1,156,000.

These bigger numbers capture something the productivity calculation misses: the effects of people who simply weren’t there. Net immigration collapsed in 2025. Depending on whose estimates you use (and they diverge dramatically), somewhere between 1.5 and 2.5 million fewer immigrants arrived compared to 2024. The Dallas Fed estimated that unauthorized immigrant employment alone contracted by roughly 310,000 workers on an annualized basis. That’s a labor supply shock, not a technology shock, and it accounts for a big chunk of the gap between the productivity-based estimate and the broader employment shortfall.

After subtracting DOGE, immigration effects, tariff uncertainty, and interest rate headwinds, the residual attributable to AI and related technology-driven displacement came out to 200,000 to 355,000 jobs.

Why two methods are better than one

If you run one model and get a number, you have a hypothesis. If you run two unrelated models on different data and get the same number, you have something closer to evidence.

Method 1 uses Challenger’s layoff announcement database. Method 2 uses BLS aggregate employment and productivity statistics plus BEA output data. The sources don’t overlap. The analytical frameworks have nothing in common. One performs subtractive decomposition on a count of discrete events; the other solves a macroeconomic accounting identity and picks apart the residual. They even measure different things. Method 1 only catches jobs that were actively eliminated through announced cuts. Method 2 also captures “silent” displacement, the kind that happens when someone leaves and nobody bothers to post the opening because an AI tool absorbed their workload.

You’d expect Method 2 to run a bit higher than Method 1 for exactly that reason, and it does. Method 1 came in at 150,000 to 230,000. Method 2 at 200,000 to 355,000. The overlap zone, where both methods agree, is 200,000 to 300,000.

That convergence is more interesting than the number itself. It suggests the signal is real and not just an artifact of whichever assumptions you happen to plug into one particular model.

What we controlled for, and where we couldn’t

Three shocks hit the 2025 labor market at roughly the same time, and all of them could masquerade as AI displacement if you’re not careful.

The federal workforce reduction was the easiest to handle. It was large (277,000 jobs lost from the January peak according to BLS), well-documented, and clearly policy-driven. We subtracted it in both methods.

The immigration collapse was harder and introduced more uncertainty into the final estimate than any other single variable. Published estimates of 2025 net migration span a range of nearly 1.6 million people. Brookings put it at negative 295,000 to negative 10,000 for the calendar year. The Census Bureau measured positive 1.3 million, but for a period running July 2024 to June 2025 that straddles two very different policy regimes. The San Francisco Fed landed around 500,000 to a million. These aren’t small discrepancies. When your biggest confounding variable has error bars that wide, your final estimate inherits that uncertainty whether you want it to or not.

We used a working estimate of 200,000 to 400,000 for the labor supply effect of reduced immigration, drawing primarily on the Dallas Fed’s detailed work tracking unauthorized immigrant employment flows. But if that estimate is off by even 100,000 in either direction, it shifts the AI attribution range meaningfully.

Tariff uncertainty and interest rate headwinds were smaller factors, contributing an estimated 100,000 to 160,000 in combined employment drag beyond what was already captured in Challenger’s explicit reason codes. These are rough estimates, not precise measurements.

What we could not control for is the world that would have existed without AI. Would companies that deployed AI have cut workers anyway? Would companies that didn’t deploy AI have hired more aggressively? The counterfactual is unknowable. The attribution fractions we applied at the final step of each method reflect informed judgment, not empirical certainty.

The real problem: we have no way to track this

The specific number is less important than what the exercise reveals about our ability to monitor what’s happening.

The United States has one of the most sophisticated labor market statistical systems in the world. Monthly employment surveys from two independent sources. Quarterly productivity accounts. Annual benchmarking against near-universal administrative records. A dedicated survey for job openings, hires, and separations. None of it was designed to detect technology-specific displacement. None of it does.

The only direct measurement we have is employer self-reporting through Challenger’s reason-code system. It captured 54,836 jobs in 2025. Our estimate is that the real number is four to six times that. The gap exists because employers face rational incentives to label AI-driven cuts as something else, and because the largest channel of AI displacement in 2025 appears to have been not layoffs at all, but the quiet decision not to replace workers who left.

Challenger’s own data tells this story clearly. Announced hiring plans fell 34% year over year to their lowest level since 2010. Seasonal retail hiring was projected at its weakest since 2009. This is what a “no-hire, no-fire” equilibrium looks like in the data. Companies deploy AI, achieve productivity gains, and shrink their workforces through attrition rather than termination. It’s nearly invisible to traditional labor market monitoring. You can only see it in the widening gap between how many jobs the economy should be creating and how many it actually is.

The methods we used here aren’t new. Excess mortality analysis has been standard practice in public health for decades. Productivity-employment gap analysis is a staple of automation research going back to the 1960s. We didn’t invent anything. We just pointed two existing tools at a question that nobody had tried to answer systematically with 2025 data, and found that they agree.

But these are still blunt instruments. They can’t tell you which specific jobs AI eliminated. They can’t distinguish between displacement caused by generative AI and displacement caused by conventional software automation. They require subtracting confounding shocks whose own magnitudes are uncertain. They are, at best, a proof of concept for the kind of analysis that should be happening regularly, with better data, at institutional scale.

What the moment actually calls for is purpose-built measurement. An industry-by-time panel using occupation-level AI exposure scores, with proper econometric identification and controls for demand shocks, trade exposure, and demographic shifts. The data to build it largely exists, scattered across the BLS, Census Bureau, and private-sector sources. The integration hasn’t been done.

What to watch

If 200,000 to 300,000 displaced jobs sounds manageable for the first year of serious enterprise AI adoption, consider the denominator. Surveys suggest that while 92% of Fortune 500 companies adopted generative AI, only about 5 to 6% successfully scaled their deployments beyond pilot programs. If the current estimate reflects the effects of that 5%, there is a lot of runway ahead.

Early 2026 data is not encouraging. January saw 108,435 announced job cuts, the highest January total since the financial crisis. Hiring plans hit their lowest January on record. Job openings continued to fall.

The single most important data point on the horizon is the Q4 2025 productivity release, due March 5. The Q3 figure of 4.9% annualized, with output growing 5.4% while hours worked grew just 0.5%, was the strongest quarterly signal of technology-driven labor substitution in the 2025 data. If Q4 confirms that trajectory, the case for AI as a structural driver gets considerably stronger. If Q3 turns out to be a one-off, the picture becomes more ambiguous.

Either way, the gap between what we can measure and what we need to know keeps widening. We built statistical infrastructure to track recessions, trade shocks, and pandemics. We haven’t built one for this. The analysis presented here is an attempt to work around that gap with the tools available. It shouldn’t have to be.

---

Data sources: BLS Productivity and Costs (Q3 2025 revised, January 29, 2026); BLS Employment Situation (December 2025, with annual benchmark revisions, February 6, 2026); BEA GDP (Q3 2025 updated estimate, January 22, 2026); Challenger, Gray & Christmas year-end 2025 report; JOLTS December 2025; Dallas Fed, San Francisco Fed, Brookings Institution, Census Bureau, and CBO for immigration estimates. Q4 GDP (BEA, due February 20) and Q4 productivity (BLS, due March 5) were not available at time of writing. Full methodology, data tables, and sensitivity analyses are in the accompanying technical report.