Back to blog
Future of WorkNovember 18, 2026 6 min read

Employment for 22-Year-Olds Is Flat. The Number Everyone Quotes Is 19 Percent.

AP
Angelo Pallanca
Digital Transformation & AI Governance

TL;DR

  • The August 2026 update of Canaries in the Coal Mine? puts employment of 22 to 25 year olds in the most AI-exposed occupations 19 percent below a kept-pace counterfactual, on ADP payroll records covering 3,5 to 5 million employees a month from January 2021 to June 2026.
  • Over the same window, total employment for that age group is roughly flat, down about 1,9 percent.
  • The 19 percent measures distance from a counterfactual. It tells you where young hiring went, not how many young jobs exist.
  • An NBER working paper from September 2026 went looking for the shock in CPS unemployment among recent graduates in summer 2026 and did not find it, including under a wider definition that counts people who say they want a job.
  • Goldman's 2027 outlook, published 15 November, records that the layoff wave did not arrive and places the displacement in later years.

Assumptions

  1. The ADP panel over-represents large firms and exposed occupations. The authors state this and warn against extrapolating to the whole economy. Solid, in the paper.
  2. Neither study is causal. The Stanford authors write that the work "does not estimate a causal impact of AI". Solid.
  3. The CPS null covers three summer months of 2026 for one group. A null over a short window is weak evidence of absence, and the authors treat it as such. Medium.
  4. Goldman's figures come from a published outlook as reported on 16 November, not from a paper I can replicate. Medium.

Goal

To pull apart the three different claims that have been travelling under one number since August, and to leave four questions that make the difference visible before anyone writes a budget against it.


Two numbers, one age group, no contradiction

Erik Brynjolfsson, Bharat Chandar and Ruyu Chen have been running the same measurement on ADP payroll data for over a year. The August 2026 vintage is the one everybody quotes: for 22 to 25 year olds in the two most exposed occupation quintiles, employment sits 19 percent below where it would be had it kept pace with less exposed peers. At the July 2025 data vintage the same calculation gave 15 percent, so the gap is widening.

In absolute terms, from November 2022 to June 2026, the two most exposed quintiles lost about 11 percent while the three least exposed gained about 10 percent. That is a 21-point spread.

And the whole age group? Down about 1,9 percent. Flat, within noise, over three and a half years.

A gap between occupations and a shrinking total are different events. One number has been doing the work of both since August.

Nothing in that paragraph is a contradiction. Young people are being hired into different jobs than they were in 2022. The authors are explicit that the divergence comes mostly from reduced hiring rather than from more separations, which is what reallocation looks like from the inside. They are also explicit that they find no evidence of widespread displacement across the economy, and that adding a single control for the college share of an occupation cuts the estimate from roughly 0,18 to roughly 0,09. In nationally representative ACS data the same gap comes out around 0,022.


The instrument that was pointed somewhere else

In September, Robert Fairlie and Jane Wu published the opposite kind of study: a null. Using CPS microdata, they looked for a spike in unemployment among recent college graduates in June, July and August 2026, the summer they argue may be the first with AI use widespread enough to leave a mark. They found none, not against previous summers, not against older graduates, not against young workers without a degree. They also built a wider measure that counts the sidelined, people who say they want a job but are not searching. It adds nearly two percentage points to the level and does not change the verdict for 2026.

So one instrument sees a widening gap between occupations and another sees no change in the count of unemployed young graduates. Both can be true because they are not measuring the same object.

The forecast that started all this was measuring a third thing. On 28 May 2025, Dario Amodei told Axios that AI could eliminate half of entry-level white-collar jobs and push unemployment to 10 to 20 percent within one to five years. Eighteen months in, US layoffs ran at 1,6 million in September, about 1 percent of payrolls, and initial claims touched 189.000 in mid-July, the lowest since 1969.

Goldman's response, in the 2027 outlook published on 15 November, is worth reading twice: this year began with fears of an AI-driven layoff wave, the layoff rate stayed historically low, the low-hire low-fire market has about another year in it, and the displacement arrives later. Its own October estimate puts up to 9 percent of workers through the transition eventually, roughly one in eleven, with the caveat that automation bites hardest during slowdowns.

A date moved and a mechanism added. Nobody scored the original.


Why this matters for your business

Robert Solow wrote in the New York Times Book Review on 12 July 1987 that you could see the computer age everywhere but in the productivity statistics. The line stood for most of a decade and then stopped being quoted, because measured productivity growth finally accelerated in the second half of the 1990s. It did not get retired by a new prediction date. It got retired by a number arriving.

If you are making hiring plans, the headline gap will not help you, because it describes an average across a payroll panel that leans large and exposed. Four questions do more, and three cost nothing.

Ask whether a number is a level or a distance from a counterfactual. Ask whether the movement comes from hires or from separations, because a hiring freeze and a layoff wave feel identical in a chart and call for opposite decisions. Ask what happens to the estimate when a control for education composition goes in. Then write down, in advance, the figure and the date that would make you wrong, and track your own monthly hiring rate by occupation against it.

The gap is real and it is growing. It is also not yet the thing anyone promised, and the honest version of the story has a year and a number attached to it.

Want to discuss this further?

Request a proposal