Generative AI exposure in Brazil associates with higher pay and longer contracted hours, not immediate job losses. Occupations most exposed to GenAI saw significant wage and hours increases relative to synthetic counterparts in the early post-2022 period, while employment levels remained unchanged in the short run.

Generatiiviselle tekoälylle altistumisen työmarkkinavaikutukset Brasiliassa

Rosberg, Santeri · January 01, 2026 · Aaltodoc (Aalto University)

openalex quasi_experimental medium evidence 7/10 relevance Summary only summary available; pdf_status=paywall Source PDF

Structured author observations

Linked only from stored provider relations; the raw author line above is never matched by name.

OpenAlex

Latest observation: August 24, 2026

Rosberg, Santeri provider ID

Using RAIS administrative data and a synthetic difference-in-differences design, the thesis finds that highly GenAI-exposed occupations in Brazil experienced no short-run employment decline but did see significantly larger increases in mean monthly wages and contracted weekly hours compared with synthetic controls through 2024.

Citation observations

Cumulative provider counts captured on specific dates; providers are never combined.

0 cumulative citations

OpenAlex · Observed August 24, 2026

View corpus context

This thesis examines whether occupations with high exposure to generative artificial intelligence (GenAI) experienced different labor market changes than low-exposure occupations in Brazil after the spread of GenAI tools in 2022. The aim is to add to the still limited early empirical evidence on the impact of AI in emerging economies. The study focuses on three outcome variables: employment at year-end, mean monthly wages, and mean contracted hours per week. The analysis uses Brazilian RAIS vínculo administrative microdata for the years 2018, 2021, 2022, 2023, and 2024. Occupational exposure was defined using the 2025 International Labour Organization GenAI exposure index for ISCO-08 occupations, which was then mapped to Brazilian CBO2002 occupation codes. The main estimation method was a synthetic difference-in-differences (SDiD) approach, with comparisons to simple difference-in-differences (DiD) calculations. The SDiD results were statistically significant for two of the three main outcomes –wages and hours. No statistically significant employment effect for highly exposed occupations was found in the short run. However, mean monthly wages and mean contracted hours increased significantly more in highly exposed occupations than in their synthetic counterfactual. The findings suggest that in the observed period, GenAI exposure in Brazil was more closely linked to complementarity and within-job adjustment than to immediate occupational displacement. The results should be interpreted with caution because time period under study (both the pre- and the post-treatment period are still short) and the analysis is for a subset of regions in Brazil.

Summary

Main Finding

Occupations in Brazil with high exposure to generative AI (GenAI) saw no statistically significant short-run employment losses, but experienced significantly larger increases in mean monthly wages and mean contracted hours per week relative to a synthetic counterfactual. This pattern is consistent with complementarity and within‑job adjustment rather than immediate occupational displacement in the period observed (post‑GenAI spread beginning 2022).

Key Points

Outcomes analyzed: year‑end employment (occupation-level), mean monthly wages, and mean contracted hours per week.
Data span: administrative RAIS vínculo microdata for 2018, 2021 (pre), and 2022–2024 (post).
Exposure measure: 2025 ILO GenAI exposure index (ISCO‑08) mapped to Brazilian CBO2002 occupation codes.
Main estimation: synthetic difference‑in‑differences (SDiD) to construct counterfactual trends for low‑exposure occupations; simple DiD used for comparison/robustness checks.
Results (SDiD):
- No statistically significant change in employment levels for highly exposed occupations in the short run.
- Statistically significant increases in mean monthly wages for highly exposed occupations relative to synthetic controls.
- Statistically significant increases in mean contracted hours per week for highly exposed occupations relative to synthetic controls.
Interpretation: early evidence points to task complementarity and within‑occupation adjustments (e.g., upskilling, changing task mix, increased intensity) rather than rapid displacement of entire occupations.
Caution: findings are short‑run and localized — both pre‑ and post‑treatment windows are relatively short and the sample covers a subset of Brazilian regions.

Data & Methods

Data source: RAIS vínculo administrative employer–employee microdata (detailed panel on formal employment spells, wages, contracted hours).
Time periods: 2018 and 2021 used as pre‑treatment observations; 2022–2024 as post‑treatment following the spread of GenAI tools.
Exposure coding: used ILO 2025 GenAI exposure index defined on ISCO‑08 occupations; mapped to Brazil’s CBO2002 occupational classification to assign exposure levels to observed jobs.
Estimation strategy:
- Synthetic difference‑in‑differences (SDiD) is the primary method: constructs a weighted synthetic control from low‑exposure occupations to estimate counterfactual trends for high‑exposure occupations and then compares post‑treatment deviations.
- Simple difference‑in‑differences (DiD) calculations were performed for comparison/robustness.
Robustness and limitations in methods:
- SDiD helps relax parallel‑trends assumptions by building a synthetic control, but short pre‑treatment series limits the precision of trend estimation.
- Potential measurement error in mapping exposure indices across occupational taxonomies.
- Analysis limited to a subset of regions; possible local shocks and spillovers not fully captured.

Implications for AI Economics

Early-stage complementarity: In an emerging economy context, GenAI adoption may initially raise task intensity and pay within exposed occupations rather than causing immediate occupational displacement.
Short vs. long run: Absence of short‑run employment losses does not preclude longer‑term reallocation or job quality changes; monitoring over a longer horizon is essential.
Policy priorities:
- Support for within‑job skill upgrading (training, on‑the‑job learning) to capture complementarities and mitigate inequality risks.
- Strengthen labor market monitoring using administrative data to detect emerging displacement or compositional shifts across regions/sectors.
- Consider regional heterogeneity when designing interventions, as impacts may vary across local labor markets and formal/informal sectors.
Research needs:
- Longer panels post‑2024 to assess medium‑ and long‑run effects, including occupational switching and sectoral reallocation.
- Granular task‑level and firm‑level analyses to identify mechanisms (task changes, productivity, hiring/firing margins).
- Examination of heterogeneity by skill level, firm size, and informality to inform targeted policy responses.

Assessment

Paper Typequasi_experimental Evidence Strengthmedium — Uses rich administrative (RAIS) data and a modern quasi-experimental estimator (SDiD), and finds statistically significant effects on wages and hours; but the pre- and post-treatment windows are short, exposure is based on an international index mapped to local occupation codes (risking misclassification), analysis covers a subset of regions and the formal sector only, and potential unobserved time-varying confounders at occupation-region level remain. Methods Rigormedium — Methodologically appropriate choice (SDiD) and robustness checks against DiD indicate careful design; however, limited pre-treatment observations, possible measurement error in mapping exposure indices to local codes, aggregation at the occupation level, and limited discussion (in the summary) of parallel-trends diagnostics, heterogeneous treatment timing, or alternative controls reduce the overall rigor. SampleBrazilian RAIS vínculo administrative microdata for selected regions, covering years 2018, 2021, 2022, 2023 and 2024; unit of analysis aggregated at the occupation (CBO2002) level with outcomes: year-end employment, mean monthly wages, and mean contracted hours per week; occupational GenAI exposure assigned via ILO 2025 GenAI exposure index mapped from ISCO-08 to CBO2002. Themeslabor_markets human_ai_collab IdentificationSynthetic difference-in-differences (SDiD) comparing occupations classified as high GenAI exposure (using the ILO 2025 GenAI exposure index mapped from ISCO-08 to Brazil's CBO2002) to a synthetic control constructed from lower-exposure occupations, using RAIS administrative microdata over 2018, 2021–2024; results are compared to conventional DiD estimates. Identification relies on SDiD's construction of counterfactual trends and the assumption that the synthetic control captures the untreated counterfactual trajectory. GeneralizabilityFindings pertain to Brazil's formal-sector employment recorded in RAIS and may not generalize to the informal sector., Analysis covers a subset of Brazilian regions, limiting national representativeness., Short pre- and post-treatment windows (few years) capture only early, short-run effects and may miss medium/long-run dynamics., GenAI exposure measure is an external index mapped across occupational coding systems, which may misclassify local task content., Occupational-level aggregation can mask within-occupation heterogeneity (firm-, worker-, and task-level variation).

Claims (8)

Claim	Direction	Outcome	Confidence & Evidence	Details
Highly GenAI-exposed occupations experienced significantly larger increases in mean monthly wages than their synthetic counterfactual after the spread of GenAI tools in 2022. Wages	positive	mean monthly wages	Reading fidelity high Study strength medium	not reported 0.48
Highly GenAI-exposed occupations experienced significantly larger increases in mean contracted hours per week than their synthetic counterfactual after the spread of GenAI tools in 2022. Employment	positive	mean contracted hours per week	Reading fidelity high Study strength medium	not reported 0.48
No statistically significant effect on employment (year-end employment) for highly GenAI-exposed occupations was found in the short run. Employment	null_result	employment at year-end	Reading fidelity high Study strength medium	not reported 0.48
The observed pattern (no short-run employment decline but larger increases in wages and hours in high-exposure occupations) suggests GenAI exposure in Brazil during the study period was more closely linked to complementarity and within-job adjustment than to immediate occupational displacement. Job Displacement	positive	within-job adjustment / complementarity versus occupational displacement (interpreted)	Reading fidelity high Study strength speculative	not reported 0.08
Occupational GenAI exposure was defined using the 2025 International Labour Organization (ILO) GenAI exposure index for ISCO-08 occupations, which was mapped to Brazilian CBO2002 occupation codes. Other	mixed	occupational GenAI exposure index / classification	Reading fidelity high Study strength high	not reported 0.8
The main estimation method used was synthetic difference-in-differences (SDiD), with comparisons to simple difference-in-differences (DiD) calculations. Other	mixed	estimation approach (SDiD vs DiD)	Reading fidelity high Study strength high	not reported 0.8
The analysis covers Brazilian RAIS vínculo administrative microdata for the years 2018, 2021, 2022, 2023, and 2024 and is for a subset of regions in Brazil. Other	mixed	data coverage / geographic scope	Reading fidelity high Study strength high	not reported 0.8
The results should be interpreted with caution because both the pre-treatment and post-treatment periods are still short. Other	mixed	temporal identification / length of pre- and post-treatment periods	Reading fidelity high Study strength medium	not reported 0.48