Where AI Value Lands

Young-Worker (22–25) New-Job Starts DiD — Exposed vs Unexposed

3 min read

Young-Worker New-Job Starts DiD

Headline result

Per massenkoff-mccrory-labor-market-impacts-2026 Figure 7, CPS panel 2016–2025, workers aged 22–25:

Pooled post-ChatGPT DiD coefficient: −14.3% of baseline hiring rate (SE 7.2). Barely statistically significant (t ≈ 2.0).

The job-finding rate for young workers entering top-quartile-exposed occupations has fallen ~14% relative to the rate for young workers entering zero-exposure occupations, post-ChatGPT.

Methodology

  • Outcome: monthly job start rate among 22–25-year-olds (defined as reporting a job they did not have the previous month).
  • Treatment group: entry into top-quartile-exposed occupations.
  • Control group: entry into zero-exposure occupations.
  • Identification: difference-in-differences around ChatGPT release (Nov 2022).
  • CPS panel dimension used to track month-to-month transitions.

Trend shape (Figure 7 top panel)

  • 2016–2019: job-start rate for exposed occupations ~1.2–1.5% / month; unexposed ~2.0–2.5% / month. Exposed lower at baseline.
  • 2020–2021: COVID-era volatility, large swings in both series.
  • 2022: pre-treatment normalisation; exposed series ~1.3%, unexposed ~2.5%.
  • 2024: visible divergence. Exposed series decays toward ~0.8–1.0%; unexposed series remains stable at ~2.0%.
  • Mid-2025: convergence in the most recent data points (within noise band).

Corroboration

Brynjolfsson, Chandar & Chen (2025) — using ADP payroll data, different methodology — find a 6–16% employment drop for 22–25-year-olds in exposed occupations. The wide range reflects multiple counterfactual specifications:

  • The 6 pp drop is vs a flat-employment counterfactual.
  • The 16 pp drop is from a within-firm design comparing similar workers in the same firm with different occupations.
  • The decrease is attributed to slowed hiring, not increased separations.

Two independent methodologies — Anthropic’s CPS-DiD and Brynjolfsson’s ADP-within-firm — find the same effect direction and rough magnitude.

What the framework cannot rule out

The source explicitly lists alternative explanations:

  • Workers stay at current jobs rather than transition to new exposed roles.
  • Workers take different (non-exposed) jobs.
  • Workers return to school to retrain.
  • Mismeasurement. Survey-based job-transition data is more vulnerable to error than separations (Fujita, Moscarini & Postel-Vinay 2024).
  • Sector-specific hiring slowdowns (tech in particular has slowed post-2022 for non-AI reasons: interest rates, over-hiring correction).

Why this matters to Where Value Lands

  • First detectable AI labor signal in the wiki. Everywhere else in this source — unemployment, wages, employment levels — the signal is null. This is the one.
  • Validates the “canary in the coal mine” framing. AI displacement may show up as lost entry slots before it appears as unemployment.
  • Reframes the detection problem. Labor-market dashboards focused on unemployment will systematically miss AI displacement. Hiring-rate panels for under-25s are the leading indicator.
  • Implication for capital-labor-divergence: the displacement falls on the entry-level cohort first. Capital owners and incumbent senior workers benefit; new graduates bear the cost.

Graph View

Backlinks