Where AI Value Lands

Unemployment DiD — Exposed vs Unexposed Workers (2016–2025)

3 min read

Unemployment DiD — Exposed vs Unexposed Workers

Headline result

Per massenkoff-mccrory-labor-market-impacts-2026 Figure 6, CPS 2016–2025:

Pooled post-ChatGPT (Nov 2022 – mid 2025) DiD coefficient: +0.0020 (SE 0.0019).

Statistically indistinguishable from zero. The unemployment gap between top-quartile-exposed and zero-exposure workers has not widened since the release of ChatGPT.

Methodology

  • Treatment group: workers in top quartile of observed exposure (per observed-exposure-measure).
  • Control group: workers in zero-exposure occupations (~30% of US workers).
  • Identification: difference-in-differences with ChatGPT release (Nov 2022) as treatment timing.
  • Robustness checks (per Appendix):
    • Vary percentile cutoff from median to 95th. In all cases the post-treatment coefficient is flat or negative (unemployment decreasing in the exposed group relative to control).
    • Restrict to young (22–25) workers. Coefficient remains flat (the young-worker effect is in hiring, not unemployment — see ai-young-worker-hiring-slowdown).
    • Substitute Department of Labor UI claims for CPS survey-based unemployment. No clear impact on exposed jobs.

Trend shape (Figure 6 top panel)

  • 2016–2019: top-quartile unemployment rate lower than no-exposure (white-collar baseline advantage).
  • 2020 COVID shock: no-exposure (in-person) workers see unemployment spike to ~18%; exposed workers spike to ~10%. The shock affects less exposed workers far more.
  • 2021–2022: rapid convergence as COVID normalises.
  • 2022 onwards: parallel trends. Top-quartile unemployment hovers ~3–4%; no-exposure ~5–6%.
  • Post-2022: no visible divergence; ChatGPT release does not produce a structural break.

Detectability frontier

The 95% confidence interval on the post-ChatGPT coefficient permits detection of differential unemployment increases on the order of 1 percentage point. A scenario in which all workers in the top 10% of exposure were laid off would lift top-quartile unemployment from 3% to 43% and aggregate from 4% to 13% — easily detectable. A “white-collar Great Recession” (top-quartile doubling from 3% to 6%) would also be visible.

What the framework cannot detect:

  • Slow trickle below 1pp differential.
  • Exit from the labor force (workers stop searching).
  • Hiring suppression without separation (which is the channel the young-worker finding suggests is active — see ai-young-worker-hiring-slowdown).
  • Wage suppression without unemployment.
  • Reallocation within an occupation.

Interpretation

What the result supports

  • No detectable AI displacement in unemployment data as of mid-2025. Two-and-a-half years post-ChatGPT, the labor-market dashboard most policy-makers watch shows no signal.
  • The “white-collar Great Recession” scenario has not happened. A common worry — that AI would crush the educated middle class — is not supported by the current data.

What the result does not support

  • It does not rule out AI labor effects. Three channels are not measured: hiring (visible separately in ai-young-worker-hiring-slowdown), labor-force exit, and wage suppression.
  • It is short-horizon. Two-and-a-half years is brief for a labor-market shift driven by capital reallocation and skill retraining.
  • It depends on the exposure measure. If the observed-exposure-measure systematically misclassifies the truly exposed occupations (which the source’s own validation against BLS partially addresses — see ai-exposure-vs-bls-growth), the DiD is testing the wrong cells.

Why this matters to Where Value Lands

  • Adjudicates the “AI is already destroying jobs” claim. It is not, at least not at the resolution unemployment data can detect.
  • Tightens the detectability question. Any future paper claiming AI displacement must clear the ~1pp differential threshold.
  • Reframes the labor-incidence channel. Since unemployment is flat, the labor-side displacement (if any) must show up elsewhere — hiring (active signal per ai-young-worker-hiring-slowdown), labor-force exit (not measured), or wage suppression (visible in high-vs-low-exposure-worker-characteristics cross-section but no time-series here).

Graph View

Backlinks