The Entry-Level AI Gap Widens to 19% at Stanford • The Weighted Average

Employment for workers aged 22–25 in the most AI-exposed occupations now sits about 19% below where it would be had it tracked their less-exposed peers, according to the revised Canaries in the Coal Mine paper from the Stanford Digital Economy Lab, built on ADP payroll records covering millions of US workers through June 2026. The same measure read 15% at the July 2025 data vintage. That is a four-point widening across eleven months of data — roughly 0.36 points a month, a rate that has not slowed since the authors first flagged it.

A gap that widens through the adjustment, not around it

The mechanism matters more than the headline. Stanford’s release memo on the August 2026 revision reports no evidence of widespread, economy-wide displacement; experienced workers show no comparable gap; and the adjustment operates primarily through reduced hiring rather than increased separations. Pay is barely moving — the authors find adjustment “showing up primarily in employment rather than base pay.” Firms are not cutting the people they have. They are declining to open the door.

The levels underneath the ratio are blunter than the ratio. A footnote in the same memo records that employment of 22–25-year-olds in the two most exposed occupation quintiles fell about 11% between November 2022 and June 2026, while the same age group in the three least-exposed quintiles grew about 10% — a 21-point spread between two cohorts of the same age over the same 43 months. The Canaries Dashboard shows where it concentrates: software developers and customer service representatives, both in the highest exposure group, post large declines for early-career workers and expansion for older ones.

The new revision adds a mechanism worth naming. Occupations built on codified knowledge — formal, documented, teachable from textbooks and written procedures — show the entry-level declines, while occupations resting on tacit knowledge acquired through practice and mentorship show employment gains, particularly for experienced workers. That distinction is the operational content of the finding. Generative models reproduce what has already been written down; the parts of a job that were never written down are the parts still hiring.

What an employer should actually do about it

The honest caveats come from the authors themselves, and they are substantial. These are descriptive patterns, not causal estimates. The gaps shrink when the researchers control for education. Some divergent trends predate generative AI. The estimates are larger in the ADP sample than in national survey benchmarks, and improvements to the data pipeline left the results “more sensitive to specification choices.” Anyone quoting 19% as a settled fact about AI is quoting more than the paper says.

What survives the caveats is enough to act on. The divergence has continued widening well after interest rates peaked; it persists when technology firms and computer occupations are excluded; it concentrates in occupations where observed usage is automating rather than complementary; and US Census Bureau administrative research shows broadly consistent raw patterns by age and industry. Four independent legs is more support than most labor-market claims carry.

One more finding deserves a line in any workforce plan. The authors report that women face greater AI exposure on average, a heterogeneity they intend to keep monitoring. A firm whose exposed entry-level roles skew female inherits a composition problem alongside a pipeline problem, and the two compound: reduced hiring in a skewed cohort changes who is available for promotion years later, long after the hiring freeze that caused it has been forgotten.

For an operator, the decision this changes is a hiring-plan decision, and it points the opposite way from the obvious one. If the tasks disappearing are the codified ones and the value remaining is tacit, then a firm that stops hiring juniors is not saving money — it is declining to manufacture the mid-career judgment it will need to buy at a premium in four years. Ars Technica’s report on the study quotes lead author Erik Brynjolfsson warning of a labor market that “keeps its overall employment level while quietly closing the on-ramp for people starting their careers.” The on-ramp is an asset on somebody’s balance sheet, even if no accounting standard records it.

Three concrete moves follow. Audit which of your entry-level roles are codified: if the job description could be executed from the written procedure alone, it is the exposed kind, and the answer is to redesign the role around review, judgment and escalation rather than to delete it. Instrument the ladder, not the headcount — track internal promotion rates from junior to mid-level, because a firm can hold headcount flat while its pipeline empties. Rebuild apprenticeship deliberately, since tacit knowledge transfers through mentorship and repeated exposure, which is precisely what shrinks when juniors stop arriving.

The archive’s earlier reading of the 6.7-point wage-growth gap in AI-exposed jobs found employers adjusting careers rather than cutting jobs; the Stanford revision says the adjustment now shows up in hiring instead of pay. Both point at the same managerial blind spot. It also sits directly beneath today’s lead on the $13 billion bidding for Hugging Face: the capital is repricing the layer that distributes automation while the labor cost of that automation lands on people who never signed the contract. Stanford’s dashboard now updates monthly, so the next revision will settle whether 0.36 points a month is a trend or a plateau.