Research update · 14 September 2026
The 19% early-career gap got bigger. The causal fight did too.
A revised Stanford payroll analysis gives the strongest headline yet for young workers in AI-exposed U.S. occupations. A new cross-country paper makes the central limit harder to ignore: remote work can produce much of the same junior-hiring pattern.
What the payroll study measured
Stanford researchers use ADP payroll data covering millions of U.S. workers through June 2026. They compare employment trends by age and occupational AI exposure. For ages 22–25, workers in highly exposed occupations were about 19% below the employment level they would have reached if they had kept pace with similarly aged workers in less-exposed occupations. Experienced workers showed no comparable gap.
Why it is a real warning
The gap grew from 15% in the authors’ July 2025 data vintage to 19% in June 2026. In levels, employment in the two most exposed occupational quintiles for ages 22–25 fell about 11% from November 2022 to June 2026 while the three least exposed groups grew about 10%. The authors say the adjustment appears mostly in employment rather than base pay and is concentrated where observed AI use is more automating than complementary.
Why it is not a causal AI result
The authors explicitly call the pattern descriptive. Occupational exposure is not employer adoption. Some differential trends appear before widespread generative-AI use; the gap shrinks after accounting for education; estimates that control for overall firm hiring are more sensitive; and the ADP analysis sample differs from national benchmarks. Their base-pay measure also excludes bonuses and equity.
The remote-work challenge
Warwick researchers use 243 million new hires and 407 million online postings across the U.S., U.K., Canada and Australia from 2017 to 2025. Separately, a two-standard-deviation increase in either GenAI or remote-work exposure predicts a lower junior share of hires and limited-experience ads. Jointly, remote work remains predictive while the GenAI coefficient sharply attenuates and is often not statistically distinguishable from zero.
That does not prove remote work caused the Stanford or Census patterns. The datasets, countries, outcomes and exposure measures differ. It does show why “the gap appeared after ChatGPT” is not enough to name the cause.
What changes in the cumulative answer
The U.S. entry signal is no longer a single paper: Census industry-state hiring, Census graduate records and ADP payroll data all merit close attention. The causal answer is less settled, not more. A Federal Reserve analysis of broad U.S. postings still finds little distinct AI-driven decline, which limits any leap to an economy-wide job-collapse verdict.
What would change this conclusion
- Firm-level studies that observe both generative-AI adoption and remote-work adoption.
- Comparable tracking of entry hires, employment, pay and later career progression.
- Later data showing whether the gap narrows, persists or spreads beyond entry routes.