Robotization in Turkey boosted local employment between 2014–2021, driven mainly by automotive manufacturing expansion; but incumbent manufacturing workers saw reduced workdays and little transition out of the sector, so net job gains arose from new hiring rather than more work for existing employees.

Main Finding

Using linked employer–employee data for Turkey (2014–2021) and industry-level robot stocks, the authors find that greater local exposure to robots is associated with positive district-level employment growth (instrumented shift-share/Bartik design). The aggregate employment gains are concentrated in manufacturing and driven largely by the automotive sector. At the worker-level, incumbent manufacturing workers in more-exposed industries experience fewer cumulative workdays at their original plants and do not tend to exit manufacturing; aggregate employment gains accrue via firm expansion and entry of new workers rather than via increased intensive-margin employment of incumbents. Earnings rise for workers who remain at their original firm, suggesting complementarity for retained workers.

Key Points

• Main empirical pattern: predicted robot exposure → positive and statistically significant increase in district employment (weighted 2SLS coef ≈ 0.035–0.046 on change in log employment) and a small positive effect on employment-to-population ratio (weighted coef ≈ 0.008). Wage effects are small/mixed (unweighted positive; weighted estimates not significant).
• Sectoral driver: effects concentrated in manufacturing, with the automotive industry contributing disproportionately to exposure and cross-regional variation.
• Worker-level margins:
  • Incumbent manufacturing cohort (2014) tracked through 2021: cumulative workdays fall for incumbents in more-exposed industries, mainly due to reduced days at their original plant.
  • Incumbents are unlikely to transition out of manufacturing; instead, employment growth reflects firm expansion and new hires.
  • Those who stay at their original plant tend to experience rising earnings, consistent with robots complementing some retained workers (especially when incumbents shift occupations within the same firm).
• Identification strategy: shift-share exposure index constructed from 2014 (or 2010 for instrument) pre-period industry shares × industry robot growth; instrument Turkish industry robot growth with robot growth in eight leading EU countries (Germany, Spain, Finland, France, Denmark, Italy, UK, Sweden) using pre-period (2010) Turkish industry shares (Bartik/shift-share 2SLS). Kleibergen–Paap statistics indicate strong instruments.
• Theoretical framing: results are consistent with the “distance-to-frontier” idea — economies farther from the global productivity frontier may see automation raise productivity and scale, increasing labor demand; authors note Turkey’s intermediate position as a plausible explanation (they do not directly test the distance-to-frontier mechanism).
• Robustness/auxiliary findings: excluding automotive reduces mean and variance of exposure (automotive shifts level but not cross-regional pattern); district weighting matters (population-weighted estimates larger and more significant than unweighted); pre-trend checks and controls (demographics, region fixed effects, manufacturing share, net exports vis-à-vis China/Eastern Europe) included.

Data & Methods

• Data
  • Employer–employee linked administrative data: Enterprise Information System (EIS), Turkish Social Security Administration (workers and firm identifiers), 2014–2021 (baseline 2014 because of detailed occupation codes).
  • Robot stocks: International Federation of Robotics (IFR), firm-level robot sales aggregated to 17 IFR industries (2005–2021 coverage used for robots 2014–2021).
  • Instrument construction: EU KLEMS country–industry employment data to aggregate/align industry definitions for EU instrument.
• Exposure construction
  • District-level robot exposure: ∆robots_TR_i = Σ_j (ℓ_ij,2014) × (robot_TR_j,2021 − robot_TR_j,2014) / emp_TR_j,2014 × 1000, where ℓ_ij,2014 are district industry shares.
  • Instrument (predicted exposure): same shift-share using robot growth in the eight EU robot-leader countries and Turkish pre-period shares (2010) to isolate exogenous international robot trends.
• Estimation
  • District-level: long differences (2014→2021) estimated by 2SLS with population-weighting in preferred specifications; controls include demographics, region FE, manufacturing employment share, net exports; standard errors clustered by province.
  • Worker-level: regressions of cumulative log workdays and wages for the 2014 manufacturing cohort on industry-level robot growth, controlling for individual, firm, industry, and region characteristics; examine mobility (plant/industry/occupation) to decompose intensive vs extensive margins.
• Validity checks
  • Instrument strength: Kleibergen–Paap F-statistics large (reported >200).
  • Spatial patterns and robustness: maps and analyses excluding automotive, pre-trend checks, alternative control sets.

Implications for AI Economics

• Heterogeneity by development stage and sector matters:
  • Automation’s net labor-market effect is not uniform; in middle-income/developing contexts like Turkey, automation can be associated with net employment gains via scale/expansion effects in manufacturing, especially when robot adoption is still relatively low compared with frontier countries.
  • Sectoral composition (e.g., large automotive clusters) can dominate local exposure patterns and aggregate outcomes.
• Mechanisms to represent in models:
  • Include distance-to-frontier / productivity-gap channels that allow automation to raise output and labor demand in non-frontier economies.
  • Distinguish intensive-margin (more hours for incumbents) vs extensive-margin (firm expansion and new hiring) responses; this paper finds expansion/new entry is the main channel in Turkey.
  • Model heterogenous worker outcomes: complementarity for retained workers (wage gains) vs reduced employment-days for many incumbents.
• Policy implications relevant to AI adoption:
  • Policies should recognize that automation can both create and displace jobs simultaneously — support for labor-market entry (training, job-matching) may be as important as displacement mitigation.
  • Because incumbents who retain jobs can benefit (higher earnings), but many incumbents see fewer workdays, redistributive and upskilling policies should target those at risk of losing intensive-margin work.
  • Local industrial policy matters: regions with concentrated automated sectors may experience different labor dynamics; spatially targeted workforce programs and labor mobility facilitation are important.
• For empirical work on AI:
  • Use shift-share instruments carefully (pre-period shares, foreign technological trends) but continue to probe assumptions (no differential pre-trends, exogeneity of foreign robot shocks).
  • Complement district/region-level analyses with worker-level panel tracking to decompose channels (intensive vs extensive margins, within-firm occupational changes).
  • When studying AI (beyond physical robots), consider analogues of robot exposure (e.g., software AI adoption by industry) and test whether distance-to-frontier and scale/complementarity channels generalize.

Limitations to keep in mind (as noted by the authors) - The paper does not directly test the distance-to-frontier mechanism with industry-level comparable productivity gaps to a frontier economy. - Exposure is constructed from pre-period industry shares and national industry-level robot growth (no direct district-level robot adoption data). - Instrument validity relies on foreign robot growth being exogenous to Turkish regional shocks conditional on controls; residual confounding cannot be fully ruled out.