Population-wide study finds no clear ChatGPT-driven youth job losses in Norway: small negative point estimates for young workers in AI-exposed occupations are statistically insignificant and a backdating test suggests pre-existing trends account for much of the observed change.

Labor Market Consequences of Generative AI: Early Evidence from Norway

Dennis Facius, R. Iacono · Fetched July 13, 2026 · CESifo working papers

semantic_scholar quasi_experimental medium evidence 8/10 relevance Summary only summary available; pdf_status=paywall DOI Source PDF

Using nationwide Norwegian administrative data and multiple quasi-experimental designs around the November 2022 ChatGPT release, the study finds no robust evidence that generative AI displaced early-career workers in highly AI-exposed occupations.

Does Generative AI displace early-career workers? We provide population-wide evidence from Norwegian administrative registers, 2015 through March 2025, exploiting the November 2022 release of ChatGPT as an availability shock. Using the within-firm composition difference-in-differences employed in recent work, supplemented with a synthetic difference-in-differences at the occupation level and a firm-level shift-share design, we find no robust evidence of employment displacement among young workers in highly AI-exposed occupations, nor any robust response across other age cohorts or on incumbent labor-market outcomes. While estimated coefficients for young workers are negative, in line with the existing literature, they are small and statistically insignificant. A backdating exercise on the synthetic difference-in-differences yields larger absolute estimates than the actual treatment date across most age bands. This suggests the apparent post-2022 decline reflects, at least in part, pre-existing secular trends rather than a clean AI-period break.

Summary

Main Finding

Using population-wide Norwegian administrative data (2015–March 2025) and three complementary identification strategies that exploit the November 2022 release of ChatGPT as an availability shock, the paper finds no robust evidence that generative AI displaced early-career (young) workers in highly AI-exposed occupations. Estimated effects for young workers are negative but very small and statistically insignificant, and there is no robust response for other age cohorts or for incumbent labor-market outcomes. Placebo/backdating tests indicate part of the apparent post‑2022 decline reflects pre-existing secular trends rather than a clean causal AI shock.

Key Points

Treatment: November 2022 ChatGPT release is used as an availability shock to generative AI.
Data: Population-wide administrative registers for Norway covering 2015–March 2025.
Primary result: No robust employment displacement of early-career workers in high-AI-exposure occupations.
Coefficients for young workers are negative (consistent in sign with prior literature) but small in magnitude and statistically insignificant.
No robust effects detected for other age cohorts or for incumbent labor-market outcomes (e.g., wages, hours, turnover).
Robustness/backdating: A synthetic difference-in-differences backdating exercise produces larger absolute estimates when using pre-2022 placebo treatment dates, suggesting pre-existing trends partly drive the observed post‑2022 patterns.

Data & Methods

Data source: Nationwide administrative registers (covering employment spells, occupations, and worker demographics) from 2015 through March 2025.
Identification strategies:
- Within-firm composition difference-in-differences (DiD): compares changes in the age composition of employment within firms across occupations with different AI exposure levels (approach similar to recent literature).
- Synthetic difference-in-differences at the occupation level: constructs counterfactual occupation-level trends from weighted combinations of control occupations to account for pre-treatment dynamics.
- Firm-level shift-share design: uses firms’ exposure shares across occupations (occupation exposure × firm occupation shares) to create cross-firm variation in AI exposure.
Outcomes examined: employment on the extensive margin (displacement) and incumbent worker labor-market outcomes (the paper reports no robust effects on these outcomes).
Robustness checks: placebo/backdating exercises (treating earlier dates as treatment) to assess whether post-2022 patterns reflect a discrete break or ongoing secular trends.

Implications for AI Economics

Causal inference caution: Apparent post-2022 declines in young-worker employment may reflect existing trends; researchers should use placebo/backdating and synthetic DiD to guard against misattributing secular trends to AI shocks.
Policy narratives: These results caution against strong short‑run narratives of mass displacement of early-career workers from generative AI availability alone. Short-term labor-market effects may be smaller and slower to materialize than some accounts suggest.
Heterogeneity and adoption: Availability (ChatGPT release) is not the same as firm adoption or task automation. Future work should measure firm-level adoption intensity, task-level replaceability, and the dynamics of adoption when assessing displacement risk.
Longer-run and intensive-margin effects: The study focuses on extensive-margin employment around an availability shock; important open questions remain about longer-run effects, occupational task reallocation, wage dynamics, and complementarities (upskilling, productivity gains).
External validity: Results are for Norway (comprehensive administrative data, specific labor-market institutions); effects may differ in other countries, sectors, or regulatory contexts.
Research best practice: Combine multiple identification designs (within-firm DiD, synthetic DiD, shift-share) and conduct backdating/placebo tests to separate treatment effects from pre-existing trends.

Assessment

Paper Typequasi_experimental Evidence Strengthmedium — Uses population-wide administrative registers and multiple complementary quasi-experimental designs, which provide strong empirical leverage; however, treatment timing is imperfect (diffusion of generative AI is gradual and heterogeneous), exposure measurement is at occupation/firm levels (potential task-level misclassification), and backdating/placebo tests reveal pre-existing secular trends that weaken a clean causal interpretation. Methods Rigorhigh — Applies state-of-the-art identification techniques (within-firm composition DiD, synthetic DiD, shift-share), exploits rich administrative data with firm and occupation identifiers, and performs robustness checks including backdating/placebo exercises, which together indicate careful and sophisticated empirical work. SamplePopulation-wide Norwegian administrative registers covering employment spells, firm identifiers, and occupation codes for workers from 2015 through March 2025; analysis stratifies by age cohorts (focusing on early-career/young workers) and by occupation-level measures of AI exposure. Themeslabor_markets adoption IdentificationExploits the November 2022 public release of ChatGPT as an availability shock and implements three quasi-experimental designs: (1) a within-firm composition difference-in-differences comparing changes in age-cohort shares across firms conditional on occupation exposure; (2) a synthetic difference-in-differences at the occupation level constructing counterfactual occupation-level trends; and (3) a firm-level shift-share design that uses pre-existing occupational exposure to generative AI to weight firm-level treatment. The paper also conducts backdating/placebo tests to probe pre-trends. GeneralizabilityCountry-specific: Norway's small, highly regulated labor market and social safety net may limit applicability to larger, more heterogeneous economies., Shock interpretation: Using ChatGPT's public release as the treatment date may not capture gradual, firm-level adoption or earlier private access to generative AI tools., Occupation-level exposure: Aggregating exposure at the occupation level can mask within-occupation task heterogeneity and firm-specific implementation differences., Short post-period for late effects: Post-treatment window (~Nov 2022–Mar 2025) may be too short to capture longer-run displacement or reallocation effects., Institutional context: Norway's sectoral composition and digital infrastructure may differ from other countries, affecting external validity.

Claims (8)

Claim	Direction	Outcome	Confidence & Evidence	Details
We provide population-wide evidence from Norwegian administrative registers, 2015 through March 2025. Other	null_result	data coverage / sample scope	Reading fidelity high Study strength high	not reported 0.8
We exploit the November 2022 release of ChatGPT as an availability shock for identification. Other	other	identification strategy / treatment timing	Reading fidelity high Study strength medium	not reported 0.48
Identification uses a within-firm composition difference-in-differences design, supplemented with a synthetic difference-in-differences at the occupation level and a firm-level shift-share design. Other	other	research design / estimation methods	Reading fidelity high Study strength medium	not reported 0.48
We find no robust evidence of employment displacement among young workers in highly AI-exposed occupations. Employment	null_result	employment (displacement) among young workers in highly AI-exposed occupations	Reading fidelity high Study strength high	not reported 0.8
We find no robust response across other age cohorts or on incumbent labor-market outcomes. Employment	null_result	employment and other incumbent labor-market outcomes across age cohorts	Reading fidelity high Study strength high	not reported 0.8
Estimated coefficients for young workers are negative, in line with the existing literature, but they are small and statistically insignificant. Employment	negative	estimated effect on young workers' employment (coefficients)	Reading fidelity high Study strength high	not reported 0.8
A backdating exercise on the synthetic difference-in-differences yields larger absolute estimates than the actual treatment date across most age bands. Employment	mixed	estimated employment effects across age bands under backdated treatment timing	Reading fidelity high Study strength medium	not reported 0.48
This suggests the apparent post-2022 decline reflects, at least in part, pre-existing secular trends rather than a clean AI-period break. Employment	null_result	interpretation of trend vs causal break in employment outcomes	Reading fidelity high Study strength medium	not reported 0.48