The missing rung: AI’s young-worker employment gap hits 19%

I.The gap that keeps widening

Erik Brynjolfsson, Bharat Chandar and Ruyu Chen published the first version of Canaries in the Coal Mine? in August 2025. On 12 August 2026 they revised it with ADP payroll records through June 2026, and the central estimate got worse rather than better. Employment among 22–25-year-olds in the two most AI-exposed occupation quintiles is now roughly 19% below where it would sit had it tracked same-aged workers in less-exposed occupations — up from about 15% in the July 2025 vintage.

The composition of that gap matters more than the headline. Young workers in the most exposed quintiles saw employment fall about 11% between November 2022 and June 2026, while their least-exposed peers gained about 10%. Over the same window, total employment in the sample rose roughly 6%. There is no aggregate job destruction here at all — there is a redistribution of hiring away from one cohort in one set of tasks. The sample is substantial: 3.5 to 5 million employees per month, drawn from firms covering over 26 million US workers, across roughly 600–700 occupations.

II.Technology that favours the incumbent

Most technology waves have favoured new entrants. Whoever learned the spreadsheet, the browser or the smartphone first arbitraged an advantage against people whose expertise had been built on the previous tool. This one runs the other way, and that inversion is the real landscape shift buried in the August revision.

If generative AI substitutes for codified knowledge, then the asset that a fresh graduate brings — a recently certified, formally documented body of knowledge — is precisely the asset whose marginal price is collapsing. The asset that appreciates is the one that cannot be written down: judgment about which client to push back on, which anomaly in a model matters, how this particular firm actually works. The Stanford data shows exactly that split, with mid-career and senior employment growing faster in tacit-knowledge occupations. The revision also finds women face higher average AI exposure than men, consistent with the systematic review of 94 studies that puts clerical work — roughly two-thirds female — among the most exposed categories, and flags outsourcing-dependent developing economies as a second exposed group through erosion of client-country demand.

The strategic consequence for firms is uncomfortable. The cheapest labour input just got cheaper still, but the input that AI complements — experience — can only be manufactured by employing inexperienced people for several years. Every firm has an incentive to buy that experience from someone else's payroll rather than grow it.

III.The mechanism moved from layoffs to hiring

The revision's most consequential finding is procedural. The divergence is driven by reduced hiring rates, not elevated separations. Firms are not firing young workers in exposed roles; they are not opening the requisition. That distinction determines everything downstream, because a hiring freeze is invisible to almost every instrument we use to detect labour-market stress.

The new analytical wrinkle is a split between codified and tacit knowledge — formal, documented, teachable expertise versus the kind acquired only by doing the work. Entry-level employment grew more slowly in codified-knowledge occupations; mid-career and senior employment grew faster in tacit-knowledge ones. The authors call this “directionally consistent” with generative AI substituting for codified knowledge while complementing tacit knowledge. Survey evidence points the same way: Gartner reported on 27 July 2026 that 22% of 110 HR leaders had at least one business leader stop entry-level hiring because of AI automation — while 95% of organisations had deployed AI and only 20% reported significant or transformational value from it.

IV.The same question, three answers

Here is where honesty is required. The 19% figure is not robust, and the authors say so first and loudest. Their own baseline coefficient for the most-exposed 22–25 cohort is −0.179 log points. Add a control for the college-educated share of an occupation and it attenuates to −0.091 — roughly halving. Run the same test on American Community Survey data for 2022–2024 and the gap collapses to −0.022, an order of magnitude smaller, leading the authors to warn against extrapolating ADP-scale magnitudes to the whole economy. They also flag divergent pre-trends that predate ChatGPT, and note their compensation measure excludes bonuses, commissions and equity — precisely the pay components that dominate in the most exposed, highest-income occupations.

Two independent datasets point the other way entirely. On 1 June 2026, Natalia Emanuel, Emma Harrington and Amanda Pallais estimated that remote work explains 64% of the rise in unemployment among young college graduates between 2017–19 and 2022–24, noting the uptick predates AI diffusion and survives controls for occupational AI exposure. And on 30 June 2026, Ramp's economics lab, pairing spending data from over 21,000 firms with Revelio Labs workforce records, found that high-intensity AI adopters grew entry-level headcount 12% in the two years after adoption and raised entry-level workforce share by 1.15 points against controls. Pulling against both, a systematic review of 94 studies published in 2026 found entry- and mid-level posting declines of 14–41% in software and content roles — evidence it reads as displacement rather than transformation.

The honest range is therefore wide enough to be policy-relevant on its own. Treat the ACS estimate as the floor and the ADP baseline as the ceiling and they differ by a factor of eight for the same cohort in the same years. Any intervention sized off the ceiling is eight times too large if the floor is right. That is not a rounding error; it is the entire question.

V.A cost with no invoice

Trace the channels and the asymmetry becomes clear. On productivity, the evidence is weak in exactly the place the thesis needs it strong: Gartner's 20%-realising-value figure sits awkwardly beside 95% deployment, and Ramp found effects only above roughly $30 per employee per month of AI spend, appearing 6–12 months after adoption. On labour, the effect is not displacement but a change in the shape of demand — fewer codified-knowledge entry roles, more tacit-knowledge senior ones, with women facing higher average exposure. On prices, nothing moved: the Stanford paper finds adjustment through employment levels, not base pay, so the usual signal that labour has become scarce or abundant is simply absent.

The market-structure channel is the one nobody is measuring. The Stanford panel deliberately excludes firm entry and exit, and the authors observe that AI-native startups look leaner, more technical and more senior. If the frontier of new firm formation is being built without a junior tier, the training function that used to be a byproduct of employment has quietly been removed from the economy — and no firm has an incentive to restore it, because the returns to training a 23-year-old accrue mostly to that worker's next employer.

This is judgment, not measurement, and it should be read as such: a hiring freeze imposes no cost on this year's income statement and leaves no trace in this year's unemployment rate. Its cost is a cohort that reaches its early thirties without the tacit knowledge the same paper says AI complements. If that is what is happening, the liability lands around 2032–2035 — comfortably outside every forecast horizon currently used to justify AI capital expenditure.

VI.Why the aggregate data will be last to know

The July 2026 Employment Situation, released 7 August, showed nonfarm payrolls down 23,000 and unemployment steady at 4.1%. Nothing in that print distinguishes an AI-driven entry-level freeze from ordinary cyclical softness, and a hiring-side adjustment concentrated in one age band inside particular occupations is close to invisible at that resolution. Meanwhile recent-graduate unemployment ran at 5.7% in June 2026 against 4.1% for all workers — a spread consistent with the thesis and equally consistent with the remote-work explanation.

Firm surveys are the better instrument, and they say something specific. Writing on 5 August 2026, New York Fed research director Kartik Athreya reported that AI use among district service firms rose from 25% to 40% between the August 2024 and August 2025 survey waves, with 44% expected within six months; manufacturers went from 16% to 26%, heading for 33%. Yet those same firms report very few AI-driven layoffs, intend to retrain rather than fire, and anticipate cutting hiring plans — especially for college-educated workers. Adoption is climbing steeply while the labour-market effect stays confined to the hiring margin. That is the pattern to watch, and it is not the one the displacement debate is set up to detect.

VII.How it plays out: three scenarios

Rough two-to-five-year horizons. Probabilities are our analytical judgment, not estimates drawn from any source.