AI cuts hiring of young workers more than it causes layoffs

On October 1, Challenger, Gray & Christmas published its September job-cuts report, and one line in it will be quoted for months. Employers cited artificial intelligence in 120,136 announced job cuts in the first nine months of 2026, about 21% of all announced cuts, which makes AI the most common reason employers gave this year. For the whole of 2025 the same count was 54,836. On its own, that line suggests that AI has started to replace workers in large numbers.

The Bureau of Labor Statistics’ turnover survey shows that US employers laid off or discharged 1.6 million people in August alone. Challenger’s nine-month AI total comes to about 7.5% of a single month of ordinary American firing. The layoff rate in that survey has stayed between 1.0% and 1.2% in every month since July 2025. AI has produced no wave of dismissals large enough to show in the official count.

I keep a ledger of AI-attributed job cuts and a job-by-job view of AI exposure, and my read is that AI is affecting jobs mostly through hiring. The layoff announcements are small, and AI has become a comfortable reason for a company to give, sometimes for cuts with older causes. The signal that holds up across independent datasets is narrower, and I think more worrying for one group: employers in AI-exposed work are hiring fewer young people. That is happening in a labor market where almost nobody is hiring much, with inflation near 4% and a rate hike and an oil shock pushing in the same direction, so I will try to be careful about how much of it belongs to AI. In the aggregate, AI probably does not explain most of it, but in a handful of occupations and one age band the pattern is too consistent to dismiss.

120,136 announced cuts, against 1.6 million layoffs a month

Challenger counts announcements, as the firm says in a footnote to its tables: “Some reductions are announced as a range or over multiple years; Challenger records the announced figure.” Through September 2026, technology employers announced 165,925 cuts this year, up 54% on the same months of 2025 and 29% of all cuts. The total was unremarkable, with all announced cuts through September at 573,195, down 39% from a year earlier, and even after removing the government cuts that inflated 2025, they were down 15%. In September itself, AI was the fifth most cited reason, at 3,961 cuts.

An AI citation is hard to interpret, and Amazon is the clearest example. Its October 28, 2025 memo announcing about 14,000 corporate cuts did talk about AI. It said “This generation of AI is the most transformative technology we’ve seen since the Internet,” and gave that as the reason the company needed to be “organized more leanly, with fewer layers and more ownership.” That argument is about speed and layers and says nothing about software doing a particular person’s tasks. The January 28, 2026 memo that cut about 16,000 more roles does not mention AI at all; it describes the round as the rest of October’s work on “reducing layers, increasing ownership, and removing bureaucracy.” It still sits on TechCrunch’s running list of 2026 tech layoffs where employers cited AI, with a June 2025 remark from the chief executive as the citation.

Other big cuts of the period had plainly ordinary causes. UPS cut 48,000 jobs in 2025 and plans up to 30,000 more operational roles in 2026, and its finance chief tied the 2026 figure to lower Amazon delivery volume, as TechRepublic reported. When Microsoft cut about 4,800 roles in July 2026 it said, per the same TechCrunch list, that the roles were “not being replaced by AI.”

Some cuts do come from AI doing the work. Salesforce’s chief executive said on a podcast in 2025 that he had cut customer support “from 9,000 heads to about 5,000 because I need less heads,” as ABC7 reported. Oracle’s headcount fell from 162,000 to 141,000 in the year to May 31, 2026, and its annual filing says, as quoted by The Next Web, that “the adoption and deployment of AI technologies across our operations have resulted” in workforce reductions. Those are specific claims by companies about their own staff, and the ledger counts that kind of filing separately from press-release language for exactly this reason.

The chair of the Federal Reserve described the pattern in January. “You hear large companies, though, saying—many of them saying—that they either won’t be hiring for some time, or that they’re hiring less, or that they’re laying people off, and they tend to refer to AI when they do that,” Jerome Powell said at his January 28 press conference. Shareholders react better to news that software has made a department unnecessary than to news that the department’s customers stopped calling, and for a company that sells AI tools the first explanation doubles as marketing. AI-cited cuts make a poor measure, because the same word covers both a support team replaced by an agent and a slowdown that would have happened anyway.

Hiring has slowed sharply

In August 2026 the hires rate was 3.3%, with 5.2 million hires, and the quits rate was 1.9%, which tells you workers do not think they can find something better. In January, Indeed’s economists wrote that hiring had slowed to rates “last seen between 2010 and 2013”, leaving out the April 2020 collapse, and nothing since has changed that picture much.

The payroll numbers agree. The September report showed 29,000 jobs added and unemployment at 4.2%, with July revised to a loss of 10,000 and August to a gain of 133,000. Over the previous 12 months the average gain was 45,000 a month. The annual benchmark made the earlier picture worse: the March 2025 payroll level was revised down by 898,000, and the preliminary estimate for March 2026 takes off another 79,000. Employers’ announced hiring plans for September, at 90,787, were the lowest for that month since 2011.

A low-hire, low-fire labor market protects people who already have jobs and punishes people trying to get their first one. That is true whatever the cause of the freeze, and it is the main reason I think the AI effect shows up first at the bottom rung. In June 2026, the unemployment rate for recent college graduates aged 22 to 27 was 5.70%, against 4.06% for all workers on the bank’s measure (three-month averages, students excluded, so its all-worker rate sits a little below the BLS headline). The gap between the two was 0.76 percentage points in June 2019, 0.85 in June 2025 and 1.64 in June 2026. Over the first half of this year the all-worker rate on that measure fell from 4.26% to 4.06% while the recent-graduate rate rose from 5.57% to 5.70%, so the gap is widening because the two lines are moving in opposite directions. The bank’s own summary of the quarter says that “Labor market conditions remained challenging for recent college graduates through the second quarter of 2026.”

Two details in the same data make me take the AI reading more seriously than I otherwise would. Young workers without a degree saw their unemployment rate dip slightly over the year, from 7.34% in June 2025 to 7.19% in June 2026, so the weakness does not apply equally to everyone under 28. And in the bank’s annual table by major, published in February, computer engineering (7.8%) and computer science (7.0%) sit near the top of the unemployment list, next to anthropology and fine arts. Underemployment for recent graduates, meaning the share in jobs that do not usually need a degree, was 41.95% in June. None of this proves AI is the cause, but it is the pattern I would expect if AI were taking over the tasks that used to justify hiring a junior analyst or developer.

Three datasets show the same pattern

The best evidence comes from three teams working with different data, and they agree more than I expected.

The first is the Stanford Digital Economy Lab’s “Canaries in the Coal Mine?” paper, which uses ADP payroll records for 3.5 to 5 million workers a month. In its August 2026 revision, employment of workers aged 22 to 25 in AI-exposed occupations “now stands 19% below where it would be had it kept pace with that of their less-exposed peers,” and experienced workers show no comparable gap. The authors say the effect runs mainly through reduced hiring of young workers and not through more separations. Last year’s headline of 13% was a regression estimate, so it does not compare directly. On the simpler measure the paper now uses, the gap widened from 15% at the July 2025 data vintage to 19% by June 2026. The same paper opens with “We find no evidence of widespread, economy-wide job displacement,” which I would like printed on every slide deck about the AI jobs apocalypse.

The second is a Harvard paper by Seyed Hosseini and Guy Lichtinger that follows about 65 million workers at 281,111 firms through résumé and job-posting data. They identify 10,433 firms that started hiring people to put generative AI into their operations, and find that after adoption, “junior employment declines sharply in adopting firms relative to non-adopters, while senior employment trends remain largely unchanged.” The drop is about 9% after six quarters, and again it comes mostly from slower hiring. Their adoption data run through March 2025, so this is the early phase, before agentic coding tools spread.

The third is Indeed’s postings data, which I find the most useful of the three because at first glance it points the other way. Software development postings have been rising since early 2025, up almost 15% since February 2025, when AI coding agents began to spread, while overall postings fell 7%. But Indeed also reports that “71% of the increase in software development job postings between May 2025 and May 2026 is from senior roles, and 37% is due to jobs that mention AI in their title.” The rebound also starts from a low base. On October 2, Indeed’s software index stood at 78.3 against a February 2020 baseline of 100, while the overall index was 103.8. Pay also favors senior roles, with advertised salaries in the most AI-exposed occupations up about 46% since 2021, against 25% in the least exposed, and Indeed finds the premium large for senior roles and “negligible at entry level.”

Dallas Fed research on wages points the same way. Since fall 2022, employment in computer systems design has fallen about 5% while wages there rose 16.7%, and across occupations, AI exposure goes with weaker wage growth where experience earns little and stronger growth where it earns a lot. The author’s summary is the clearest statement of the mechanism I have read: “AI will automate jobs requiring codifiable (textbook) knowledge but complement jobs demanding experiential tacit knowledge.” The BLS payroll survey is consistent with this, with the information sector at 2,739,000 jobs in September 2026, down 120,000 from a year earlier, while health care added about 372,000 over the same twelve months.

Firms that use AI seriously keep and pay their experienced people, sometimes hire more of them, and take on fewer beginners. The textbook part of a junior job is the part the tools do well, and the judgment that makes a senior worth more is the part they don’t. That helps seniors in the short run, but it hurts anyone who expects to become a senior, because the usual way to get tacit knowledge is to spend a few years as a junior.

The evidence against my reading

Some of the confounders are large. The Federal Open Market Committee raised rates by 25 basis points in September, which Governor Lisa Cook said in a September 28 speech was to address inflation of an estimated 3.8% over the 12 months to August, with core at 3.4%. Indeed noted it was the first increase in more than three years. Cook also pointed to higher oil prices and supply disruptions tied to the conflict in the Middle East, and Andy Challenger, the firm’s chief revenue officer, described employers facing “high energy costs, an uncertain war in Iran, a rate hike that could make hiring more expensive” in the September report. Any of these would freeze hiring on its own, and a hiring freeze hurts new entrants first whatever its cause. The weakness in tech also began in 2022, months before ChatGPT launched, after a hiring boom had pushed software postings to more than twice their February 2020 level.

The aggregate estimates of AI’s effect are small. The Dallas Fed calculated in January that the drop in young workers’ share of AI-exposed jobs would explain only about 0.1 percentage point of the unemployment rate even if all of it turned into unemployment. Goldman Sachs estimates a net AI drag of about 16,000 jobs a month over the past year, roughly 0.1 point on unemployment so far. The Yale Budget Lab’s tracker, updated through the August household survey, finds the data do “not provide clear evidence of labor market disruption associated with AI”. In February, Governor Michael Barr said the research suggests AI “has yet to have a substantial effect on aggregate employment or unemployment”, while it may be starting to hurt young people at the start of their careers. Powell, in the January answer quoted above, said there appears to be “some connection” between low hiring of recent graduates and AI, “but it’s not the main, or only, driver.”

The Stanford authors put their own caveats in the abstract. The patterns weaken when they control for education, some divergent trends predate generative AI, and the effect is “more pronounced in the ADP analysis sample than in national survey benchmarks.” The Harvard design identifies adopters by their job postings, which could pick up firms that were restructuring for other reasons. I think both results are probably right in direction and probably overstated in size, though I can’t tell you by how much.

As of the September jobs report, AI is a minor force in the aggregate numbers and a measurable one for young workers in exposed occupations, where it adds to a cyclical freeze that it did not cause. The official forecast reaches about the same conclusion. In its 2025–35 projections, released August 27, BLS expects total employment to grow by 5.9 million, or 3.5%, over the decade, and says AI-powered automation is likely to reduce demand for several office and administrative support occupations, the group it projects to shrink fastest (down 4.0%, or 752,100 jobs). It projects customer service representatives to fall 5% and claims adjusters to fall 6%. Software developers are still projected to grow, by 10%, down from 17.9% two projection cycles ago, a big cut for one of the most exposed occupations.

With the same release BLS began publishing an AI exposure rating for each of its 831 projected occupations, combining five sources: three estimates of what AI could do in theory, and two measures of observed use, from Anthropic’s Claude data and Microsoft’s Copilot data. The agency warns that “An exposure category is not a forecast of employment growth or decline.” I agree, and my /jobs page is built on the same table with the same warning. By BLS’s ratings, 56% of US workers, about 95 million people, are in occupations rated high or very high for exposure, and BLS projects those occupations to grow more slowly to 2035 than the moderate group does. I watch most closely the occupations that are highly exposed and whose hiring of young people is falling anyway.

What would change my mind

Several useful releases are due in the next three months. BLS publishes the October jobs report on November 6. The New York Fed updates its graduate figures generally in February, May, August and November, so the third-quarter numbers should arrive next month. The Stanford team now publishes public indicators intended for ongoing tracking, and the next vintage will show whether the 19% gap is still widening. At his September 16 press conference, Fed Chair Kevin Warsh said the Fed’s AI task force “should report by the end of the year”, and I want to see whether it treats AI only as a capital-spending story, as most Fed remarks this year have, or also as a question of who gets hired.

A layoff rate in the turnover survey above its 1.0% to 1.2% range, with the increase concentrated in exposed industries such as information and professional services, would mean the layoff headlines were right but early. A recent-graduate gap that closes while overall hiring stays weak would mean the bottom-rung pattern was mostly cyclical, and that I gave AI too much credit. Junior roles in Indeed’s software rebound, in serious numbers, would ease the apprenticeship worry, because it would mean firms have found a way to train beginners alongside the tools.

I expect that none of those three results will happen cleanly, though I hold that view loosely. I expect the aggregate numbers to keep saying “not much,” the junior-hiring numbers to keep saying “something,” and the AI-cited layoff count to keep rising faster than any measure of what AI is doing. The group I will keep watching is recent graduates in exposed fields, whose missing job offers never appear in any layoff count, and the first new reading on them arrives in November.