Automation and labour market inequalities: a comparison between cities and non-cities

The study investigates how automation—measured via robot adoption—affects labour market participation and the composition of employment by skill level across urban versus non-urban areas in Italy. Motivated by renewed concerns that recent technologies can substitute for both routine and some cognitive tasks, potentially decoupling economic and employment growth and worsening income distribution, the paper asks whether displacement and compositional effects differ by territorial context. While prior work documents displacement and wage-share effects of robots across countries, evidence on spatial heterogeneity between cities and non-cities is limited. The authors argue that, although automation displaces jobs everywhere, urban labour markets—given their diversified sectoral mix and higher presence of advanced, non-routine occupations—may experience a reorientation towards high-skilled, higher-paid jobs. This could either reflect upskilling of displaced workers (an equalising mechanism) or the exit of low-skilled workers with an expansion of elite jobs (an inequality-amplifying mechanism). Italy provides a suitable setting due to its varied urban–industrial structure and persistent macro-regional divides in technological intensity. The study focuses on 2009–2019 to capture post-2008 dynamics while excluding COVID-19 years.

The paper builds on a large literature documenting the labour market impacts of automation and robots. Studies for the US, EU, Japan, France, Germany and Italy have found robot-driven displacement and, in some cases, compression of low-skilled workers’ wage shares (e.g., Acemoglu & Restrepo 2020; De Vries et al. 2020; Dauth et al. 2021; Aghion et al. 2019; Caselli et al. 2021). Much prior work uses spatially aggregated data (e.g., commuting zones) without explicitly contrasting outcomes in urban vs non-urban settings, leaving open whether urban labour markets are sheltered or differently affected. A few works examine urban environments but not the differential city–non-city comparison. The authors position their contribution as unpacking the city dimension, linking robot adoption to participation and skill composition to assess implications for inequality across territories.

Data: The core data come from the Italian Labour Force Survey (Rilevazione delle Forze Lavoro, RFL), 2009–2019, aggregated to NUTS3. Dependent variables are (i) employment-to-population ratio and (ii) employment shares by skill group (low-, mid-, high-skill). Skill groups use ISCO 1-digit: low-skill = ISCO 8 (plant/machine operators, assemblers) and ISCO 9 (elementary); high-skill = ISCO 1 (managers), ISCO 2 (professionals), ISCO 3 (technicians and associate professionals); mid-skill = remaining categories. Controls (from RFL) include gender, education, sectoral employment shares, median age, female employment, tertiary share, and composition of skill shares. Robot adoption: National IFR robot data (2004–2017) are apportioned to NUTS3 using three weights: NUTS3 shares of (i) manufacturing employment, (ii) blue-collar occupations (ISCO 8), and (iii) households with broadband. Regional robot stock is converted to robot density by dividing by manufacturing employment (robots per manufacturing worker) and smoothed as a 3-year moving average. Robot density growth is the average annual compound growth over the previous 5 years. To control for region size, density measures are used. City definition: Cities are identified by population thresholds at NUTS3 using an administrative approach. Three dummy variables are defined: top 25% most populous NUTS3, top 20%, and top 10% (≈ ≥1 million inhabitants). Non-city denotes the complement set. Econometric approach: Two models are estimated.

• Model 1 (Tables 1–2): Employment-to-population ratio_rt = α·Automation_{r,t−1} + β·City_r + γ·(Automation_{r,t−1} × City_r) + θ·X_{r,t−1} + year FE + NUTS2 FE + random effects (panel). Independent variables are lagged to mitigate endogeneity. Random-effects choice is supported by a Hausman test (χ²=8.99, p=0.98). Robust standard errors are used. R² ranges 0.75–0.78.
• Model 2 (Table 3 and Supplementary Table 1): Employment share_{r,s,t} = α·Automation_{r,t−1} + β·City_r + θ·Z_{r,s,t−1} + φ·X_{r,t−1} + year FE + NUTS2 FE, estimated in pooled panel with robust SE. Z includes group-specific demographics (median age, female employment rate, tertiary share, skill shares). X includes sectoral structure controls (agriculture, manufacturing, public/private services). Marginal effects by city group are computed to compare urban vs non-urban responses. Time frame: 2009–2019 (to avoid COVID-19 confounding and because NUTS3 microdata are consistently available from 2009).

• Automation displaces labour across all territories:
  • Robot density has a negative and significant association with the employment-to-population ratio: coefficient ≈ −0.005 (SE 0.001), p<0.01 across specifications.
  • Growth of robot density is also negative and significant: coefficients ≈ −0.136 to −0.139 (SE ≈ 0.041–0.042), p<0.01.
  • City dummies are positive and significant, indicating higher participation in more populous NUTS3: p75 ≈ +0.033 (SE 0.011), p80 ≈ +0.037 (SE 0.013), p90 ≈ +0.058 (SE 0.014), all p<0.01. Interaction terms between automation and city are not significant, implying similar displacement magnitudes across space.
  • Marginal effects (Table 2) confirm comparability between city and non-city:
    • Robot density: −0.005*** for both city and non-city groups (top 25%, 20%, 10%).
    • Robot density growth: negative and significant across groups, e.g., top 25% cities −0.109** (SE 0.055), non-city −0.139*** (SE 0.042); top 10% cities −0.160*** (SE 0.059), non-city −0.136*** (SE 0.041).
• Skill composition effects:
  • Low-skill employment shares decline with robotisation in both urban and non-urban areas:
    • Example marginal effects (Table 3):
      • Top 25% cities: −0.005 (SE 0.001) for robot density; −0.568 (SE 0.167) for growth.
      • Non-city: −0.004 (SE 0.001) for robot density; −0.340 (SE 0.108) for growth.
      • Similar negative results for top 20% and top 10% city thresholds.
  • Mid-skill employment shares show no significant response (ns) to robotisation across all specifications and city definitions.
  • High-skill employment shares increase with robotisation in urban areas but are neutral in non-urban areas:
    • Top 25% cities: +0.003 (SE 0.001) for robot density; +0.278 (SE 0.126) for growth.
    • Top 20% cities: +0.004 (SE 0.001) for robot density; +0.282 (SE 0.141) for growth.
    • Top 10% cities: +0.003 (SE 0.001) for robot density; +0.336 (SE 0.165) for growth.
    • Non-city: high-skill effects are not significant.
• Controls and fit:
  • Female employment share positively associated with participation (e.g., ≈ +0.055, p<0.01).
  • Agriculture and public services shares negatively associated with participation (e.g., agriculture ≈ −0.129 to −0.221; public services ≈ −0.469 to −0.658; many at p<0.01).
  • R² between 0.75 and 0.78 in employment-to-population models.

Findings show that robot adoption reduces labour market participation uniformly across urban and non-urban regions, countering the notion that cities are sheltered from displacement. However, urban areas exhibit a distinct compositional adjustment: while low-skilled workers are displaced in both contexts, only cities experience a concurrent rise in high-skilled employment shares. Mid-skilled shares remain unaffected. The absence of significant automation–city interactions in participation models, combined with negative marginal effects in both settings, confirms broad displacement. The increase in high-skill shares in cities, alongside shrinking low-skill shares and overall participation declines, suggests that urban labour markets are reorienting toward elite, higher-paid occupations rather than broadly upskilling displaced workers. This pattern raises concerns about widening within-city inequalities between high-skilled insiders and low-skilled outsiders. In non-urban regions, displacement is not matched by reinstatement or upskilling dynamics, implying exits from the labour force among low-skilled workers. Overall, automation acts as a labour-saving force everywhere, with cities transforming toward higher-skill employment structures and non-cities facing net losses among low-skilled workers, with important implications for spatial inequality and social cohesion.

The paper contributes by providing spatially disaggregated evidence that automation via robotisation reduces labour market participation across the urban–non-urban spectrum and reshapes employment composition differently by territory. Cities experience rising high-skill shares alongside low-skill displacement, pointing to an evolution toward elite job markets and a likely amplification of within-city inequalities. Non-urban areas face displacement of low-skilled workers without compensating high-skill gains, increasing risks of labour force exit. Future research directions include: (i) extending analysis to technologies that perform cognitive and non-routine tasks beyond industrial robots, and (ii) examining spatial spillovers of automation-induced displacement on nearby labour markets. The results call for policies that address displacement among low-skilled workers and mitigate rising urban inequalities, while supporting inclusive skill development and labour market transitions.

• Potential endogeneity: Although regressors are lagged to mitigate bias, endogeneity may persist (e.g., regions adopting robots may have unobserved trends affecting employment). Random-effects specification is chosen given time-invariant city dummies; identification relies on within- and between-region variation with year and NUTS2 fixed effects.
• Measurement of robot adoption: IFR data are apportioned to NUTS3 using weights (manufacturing employment, ISCO 8 share, broadband). This introduces measurement error; robot data end in 2017 and use 3-year moving averages and 5-year growth windows.
• City definition: Administrative NUTS3 population thresholds (top 25%, 20%, 10%) may not perfectly capture functional urban areas; results, however, are robust across thresholds.
• Aggregation level: NUTS3 aggregation may mask within-region heterogeneity; comparisons with studies at finer geographies (e.g., local labour markets) may yield different findings.
• Model for skill shares uses pooled panel (not unit fixed effects by skill group), which may leave residual unobserved heterogeneity.
• Time scope: 2009–2019 excludes COVID-19 years and newer waves of AI adoption beyond robots; external validity to later periods may be limited.