The claim going around this autumn is that AI’s effect on jobs shows up in hiring before it shows up in layoffs. Firms that adopt it don’t fire people. They stop replacing them, and they stop bringing in juniors. I wanted to know how much of that rests on data that can actually tell a hire from a layoff. The answer is less than the coverage suggests.
Start with the national numbers, because they are the reason the claim needs making at all. The Bureau of Labor Statistics’ JOLTS release for August, published on 29 September, put hires at 5.2 million, a rate of 3.3 percent, and layoffs and discharges at 1.6 million, a rate of 1.0 percent. The bureau described both as having “changed little.” If AI were pushing people out of jobs at any scale, that layoff rate is where you would expect to see it first. One month of “changed little” doesn’t rule out a small effect, but it shows no large one.
So the argument has moved to the other side of the ledger. The Federal Reserve’s Beige Book of 14 October says that “five Districts reported that some firms had either reduced staffing levels (through attrition), delayed hiring, or anticipated less job creation after increasing productivity by investing in AI.” One district’s IT staffing contacts said AI “was not causing layoffs but was dampening the creation of new jobs.” The Dallas report adds that some large accounting firms are delaying new employment contracts because they “may not need as many entry-level employees as initially anticipated, partly due to efficiency gains from AI.” These are anecdotes from business contacts, which is what the Beige Book collects. They line up with the claim. They don’t measure it.
The studies that count both doors
A labor market has two doors. People come in through hires and leave through separations: quits, layoffs, retirements. To say AI works through the first door and not the second, you need data that records both, for the same firms, over the same period. Most datasets record one or neither.
The study that gets the most attention for recording both is by Seyed Mahdi Hosseini Maasoum and Guy Lichtinger at Harvard, “Generative AI as Seniority-Biased Technological Change,” whose latest version is dated 6 May 2026. It uses résumé data covering about 65 million workers at more than 280,000 US firms from 2015 to 2025, and it identifies firms that adopted generative AI from their own job postings for what the authors call “GenAI integrator” roles. Because résumés carry start and end dates, the authors can reconstruct who joined each firm and who left it, quarter by quarter.
Their headline sentence is the claim itself: the decline in junior employment at adopting firms “is driven primarily by slower hiring rather than increased separations.” The numbers behind it are stark. After the first quarter of 2023, adopters hired about 4.0 fewer junior workers per quarter than comparable non-adopters, against a pre-period average of 5.06 junior hires per firm-quarter. That is a cut of roughly 80 percent. Junior separations at the same firms didn’t rise. They fell, by about 1.1 per quarter. Promotions of juniors ticked up slightly. Senior employment trends, the authors write, “remain largely unchanged.” They also caution that adopters differ systematically from non-adopters, in size, workforce and industry.
The falling separations rarely make the summaries. The claim usually reads as “hiring fell while layoffs stayed flat.” This paper says something stronger: at adopting firms, junior separations did not rise, they fell, by about a quarter of the hiring drop. The paper reports counts, not rates, so a smaller junior headcount alone could account for part of that. It is consistent with the authors’ explanation. Cutting hires now, they suggest as one plausible explanation, may be “less costly than future layoffs” for firms that expect to automate junior tasks, a mechanism they formalize in an appendix model with labor-adjustment costs. A firm expecting to need fewer juniors stops adding them. If that is right, the effect is invisible in any series that counts exits, by design.
The same column holds a study that cuts the other way. Peter John Lambert (Warwick and LSE) and Yannick Schindler (Ellison Institute of Technology), in a May 2026 Warwick working paper, use 243 million new hires and 407 million online postings across the US, the UK, Canada and Australia. Measured separately, a two-standard-deviation increase in exposure to generative AI, or in exposure to working from home, each predicts a fall of about 5 percentage points in the junior share of new hires by 2025. Estimated jointly, “the WFH effect remains, while the GenAI coefficient attenuates sharply and is often statistically indistinguishable from zero.” Their appendix tracks how fast junior hires replace junior separations, relative to seniors, and there the AI coefficients are “small, insignificant, or signed incorrectly” for the hypothesis that AI cut demand for juniors. Remote work arrived in the same years and hit many of the same occupations, and most designs in this debate don’t separate the two.
What the other studies measure
Almost everything else in this debate measures something other than flows, and the difference matters.
A widely cited study is the Stanford Digital Economy Lab’s “Canaries in the Coal Mine” by Erik Brynjolfsson, Bharat Chandar and Ruyu Chen, updated on 12 August with ADP payroll data through June 2026. Payroll counts who is employed, a stock. By that measure, employment among workers aged 22 to 25 in the most AI-exposed occupations “stands about 19% below” where it would be had it kept pace with similarly aged workers in less-exposed jobs, up from 15 percent at the July 2025 data vintage. Experienced workers “show no comparable gap.” On mechanism, the update has one sentence: the adjustment “appears to operate primarily through reduced hiring of young workers rather than increased separations.” It is the right sentence and probably the right reading. But the page gives no hire or separation figures behind it, and the authors call their results descriptive, saying they “cannot yet answer” whether AI is the cause.
Chandar’s newer paper with Bouke Klein Teeselink, “How Does AI Change Labor Demand? Evidence from 41 Countries” (September 2026), is larger: 1.25 billion job postings and 154 million employment records. It finds that foreign affiliates of AI-adopting multinationals have a junior share of employment 1.9 percentage points lower than matched controls by March 2026. But the decline, in the authors’ words, “primarily reflects growth in senior employment rather than a fall in junior employment.” Junior employment is down 2.5 percent, “not statistically significant.” Their summary is that AI “changes which workers an affiliate employs more than how many.” That is a composition finding, not a flows finding.
Then there are postings, which measure what employers advertise, not whom they hire. Here the evidence splits. Dodini and Smith at the Dallas Fed (1 September) find that Texas postings for more-exposed occupations fell about 8 percent relative to less-exposed ones by the first quarter of 2025 (US-wide results were similar), and say the overall effect on aggregate posting “thus far has been modest.” Audoly, Guerin and Topa at the New York Fed (14 May) found the exposed-occupation divergence began before ChatGPT and stabilized after 2023, conclude AI “is not the main driver of the slowdown in hiring,” and report: “we do not observe a divergence in labor demand between junior and senior positions within highly exposed occupations.”
Some studies find nothing at all. Using Norway’s population-wide administrative registers through March 2025, Facius and Iacono (16 June) find “no robust evidence of employment displacement among young workers in highly AI-exposed occupations,” with the apparent post-2022 decline reflecting, at least in part, “pre-existing secular trends.” Their young-worker estimates are negative but small and statistically insignificant. Iscenko and Curto Millet, in a January working paper for the Economic Innovation Group, argue the US entry-level slump fits the interest-rate cycle better than AI, since about half the decline in their example occupations had happened by June 2023, before enterprise AI tools were on sale. The Stanford group wrote in February that they “do not find strong support” for the interest-rate explanation, while conceding that with firm-time controls the decline appears only from 2024. That exchange is not settled. A Ramp and Revelio Labs working paper (30 June) goes the other way on adoption itself: among 21,559 US firms, those spending the most on AI grew total headcount about 10 percent and entry-level headcount about 12 percent in the 24 months after adoption, though adopters were already larger and faster-growing before they adopted.
The postings that change shape
The posting data does show something real, though it is about mix rather than volume. Indeed’s Hiring Lab reported on 23 July that as of May 2026, senior postings were up 14.7 percent from a year earlier and entry-level postings down 7.5 percent, and listed AI, remote work, the Fed’s rate hikes and post-pandemic corrections as candidate causes without picking one. A September note by Jack Kennedy, headlined on pay, carries a starker secondary result: among salaried postings in the most AI-exposed occupations, the entry-level share “fell from 29% to 10% between 2021 and 2026,” while the senior share rose from 22 to 47 percent. The same note says postings in more-exposed occupations have generally seen the largest rebound in postings over the past year, so the shift is not one-way.
Two caveats come with that chart. The sample is salaried postings, which skew senior: about 37 percent senior in 2026 against roughly 14 percent of all postings. And these are shares, so the entry-level line can fall because senior ads grew, the same arithmetic that drives the Chandar and Klein Teeselink result. And the source presents these shares as a side result of a pay analysis whose seniority breakdown Kennedy treats as only suggestive.
A precedent from the switchboard
There is a historical case that fits this pattern closely, and it is a good argument for taking the hiring channel seriously. Between roughly 1920 and 1940, AT&T converted its exchanges to mechanical dialing one city at a time. James Feigenbaum and Daniel Gross used the staggered cutovers to study what happened to telephone operators, then one of the largest occupations for young American women. Their paper, published in the Quarterly Journal of Economics in 2024, finds that after cutover the number of young operators fell by 50 to 80 percent, “upwards of 80%” in the authors’ conclusion, “a near-total collapse in entry-level hiring.” The labor market adjusted at the door.
The same paper has the complication that the hiring story usually leaves out. The incumbents, the operators already employed, were the ones most harmed over time: later less likely to be working, or working in lower-paying jobs. And later cohorts’ overall employment did not fall, because clerical and service jobs absorbed them. A frozen entry door moves the cost of automation onto people who never appear in a layoff count, and some of it lands on the incumbents anyway, just later.
What I take from it
The hiring-not-layoffs claim is probably right for the firms that have adopted AI heavily. The Harvard résumé panel supports it, more strongly than the headline does: separations fell. The Stanford payroll work points the same way, and the Beige Book anecdotes agree. But the claim rests on thin evidence: one résumé panel supports it, and a four-country hiring study finds the AI link fades once remote work is controlled for. The broader evidence, from postings and from Norway’s registers, is mixed about whether the junior decline is AI at all rather than interest rates, remote work or a pre-existing trend.
In August I wrote that agents would absorb mid-tier work in knowledge functions, leaving an hourglass, and entry-level roles in frontline operations, leaving a diamond. That was a forecast. What can now be measured is narrower: at adopting firms, fewer juniors come in and fewer leave. If you want to watch for the effect, a layoff tracker will show you nothing, and so will the national hires rate, which isn’t moving. The places to look are hires by age within exposed occupations and separation rates for the same groups, measured at firms that adopted versus firms that didn’t. Very little public data does that yet, which is why a single résumé panel is carrying so much of a claim this widely repeated.
References
- U.S. Bureau of Labor Statistics (2026). Job Openings and Labor Turnover — August 2026. USDL-26-1547, 29 September 2026.
- Board of Governors of the Federal Reserve System (2026). The Beige Book: Summary of Commentary on Current Economic Conditions. 14 October 2026.
- Hosseini Maasoum, S. M. and Lichtinger, G. (2026). Generative AI as Seniority-Biased Technological Change: Evidence from U.S. Résumé and Job Posting Data. SSRN 5425555, version of 6 May 2026 (first version 31 August 2025).
- Lambert, P. J. and Schindler, Y. (2026). The Broken Ladder: AI, Remote Work, and Early-Career Hiring. Warwick Economics Research Paper 1615 / CAGE Working Paper 808, May 2026.
- Brynjolfsson, E., Chandar, B. and Chen, R. (2026). Canaries in the Coal Mine? August 2026 update. Stanford Digital Economy Lab, 12 August 2026.
- Stanford Digital Economy Lab (2026). Canaries, interest rates, and timing: more on recent drivers of employment changes for young workers. 9 February 2026.
- Chandar, B. and Klein Teeselink, B. (2026). How Does AI Change Labor Demand? Evidence from 41 Countries. Stanford Digital Economy Lab, September 2026.
- Dodini and Smith (2026). Job postings show early signs of AI automation impact. Federal Reserve Bank of Dallas, 1 September 2026.
- Audoly, Guerin and Topa (2026). Do Job Postings Show Early Labor-Market Effects of AI?. Liberty Street Economics, Federal Reserve Bank of New York, 14 May 2026.
- Facius and Iacono (2026). Labor Market Consequences of Generative AI: Early Evidence from Norway. 16 June 2026.
- Iscenko and Curto Millet (2026). Looking for the Ladder: Is AI Impacting Entry-Level Jobs?. Economic Innovation Group working paper, January 2026.
- Kharazian, A., Simon, L. and Stevens, R. (2026). A New Look at AI’s Impact on Jobs: Firm-Level AI Spending and Workforce Adjustment. Ramp and Revelio Labs, 30 June 2026.
- Aidala and Puri (2026). The Labor Market Is Tilting Toward Seniority. Indeed Hiring Lab, 23 July 2026.
- Kennedy, J. (2026). AI Exposure Isn’t Squeezing Advertised Pay in the US — It’s Boosting It. Indeed Hiring Lab, 17 September 2026.
- Feigenbaum, J. and Gross, D. P. (2024). Answering the Call of Automation: How the Labor Market Adjusted to Mechanizing Telephone Operation. Quarterly Journal of Economics 139(3): 1879–1939; NBER Working Paper 28061.