Young-worker career paths after AI adoption: does a thinner entry rung become lost earnings?

Research brief completed 16 October 2026. Question: do observed declines in young-worker hiring in AI-exposed work translate into lost employment or earnings, or reallocation to other firms and roles? A good answer links individual entry cohorts and employer AI adoption over time, observes rejected/redirected applicants and exits, checks pretrends and output/demand, and distinguishes lower junior share from headcount and job quality. Earlier young-worker map offered occupational payroll gaps; adoption comparisons found Danish small-firm hiring slowdown but multinational junior-share decline partly from more seniors. These newly checked sources move the individual entry outcome question but still lack identified AI-causal career trajectories.

Four subquestions and original-source checks

1. Linked graduate outcomes? Orr, Tucker and Warren’s September Census working paper joins graduate administrative records to observed employment, pay and sector. For the most exposed decile of majors, regression-adjusted initial employment is about five percentage points lower, and full-quarter earnings among employed graduates 13% lower; roughly half the earnings difference follows movement toward retail/restaurants and other lower-paying sectors. The gap attenuates later but remains in observed windows. Dodini and Smith’s Texas Fed authored analysis independently sees 1.7-point less first-year Texas employment and 5% lower conditional wages per 10-point increase in exposure (different contrast: do not directly compare magnitudes); more exposed graduates also re-enter school. Neither records each employer’s actual AI rollout or all rejected applications. The Texas boundary specifically misses out-of-state jobs.
2. Population alternative work and unemployment? Fairlie and Wu’s September national CPS paper finds no statistically detected overall recent-college-grad unemployment spike in summer 2026 versus prior summers and older/less-educated comparison groups, even adding those who want a job but are not officially unemployed. This does not undo a concentrated exposed-major decline or low-quality fallback employment. Monthly survey slices are modest, occupation for unemployed people is often missing, and it does not track a graduating cohort’s applied-for versus accepted roles. Search found no directly linked applicant-level AI rollout → rejected application → later annual earnings dataset.
3. Does output/demand offset youth losses? Humlum and Vestergaard’s March-revised Danish adopter study links reported chatbot use to payroll and finds no worker or workplace earnings/hours effect above about 2% over its first two years, despite task reorganization and mobility. Later Danish adopter and multinational data find lower small-firm hiring and higher senior headcount at big adopters, respectively. Endpoints/timing differ; neither measures nonadopter loss versus adopter market-share gain, verified prices/output and young applicants jointly. No defensible aggregate demand-offset or lifetime cost can be calculated from these sources.
4. Competing cause and identification? Lambert and Schindler’s August original and short essay compare remote-work and AI exposure across four countries: both measures correlate strongly, and AI association with falling junior hire share attenuates when remote feasibility/adoption is included. Remote-first mentoring costs plausibly reduce junior hiring without an AI-substitution mechanism. Their posting-based remote adoption is itself selected, so this does not prove AI had no effect; they acknowledge studies where AI association survives remote controls. Tech demand and major composition remain rival explanations for Census exposure-by-cohort designs. National CPS null and linked exposed-major harm do not conflict: populations, definitions and endpoints differ.

What changed: a narrower individual harm claim now has teeth: among covered cohorts, AI-exposed majors entered worse jobs and some shifted to lower-paying sectors. The strong rebuttal is not “nothing happened” but “exposure and the ChatGPT date do not identify what adopting firms caused”; remote work, sector demand and other shocks may help explain it, while national mass graduate unemployment was not observed through summer 2026. What is missing: matched graduate applicants across in- and out-of-state firms, timestamped verified AI use and remote-work changes, vacancies and offers, quarter-by-quarter total earnings including zeroes, and employer output/prices. A credible design compares similar employers adopting at different times with pretrend checks, covers nonadopter market losses and tracks AI versus training-policy interventions, with confidentiality controls. Then estimate both graduate distributional cost and all-employer net effects, not just the exposed-sector headcount.

Search process and feed: one query per linked-outcomes, applicant paths, output-offset and competing-shocks question; narrowed searches to originals at Census, Dallas Fed, NBER/IZA and the authors’ LSE/Warwick essay/paper. Verified the Census employment/pay/sector figures against its original abstract/full working paper and the national null against the NBER and IZA original; authors of the remote-work alternative explicitly concede AI may add a second shock. Further broad labor search unlikely to solve missing verified adopter and lifetime path tonight. The authored LSE short original is a weekday hold if Dru asks about causality; Census 2026 long paper is a weekend hold if he asks about harm not visible in unemployment. The lone government-cyber post has not been opened; do not pile an unrelated second link on it tonight.