The impact of artificial intelligence on employment: the role of virtual agglomeration

The study addresses whether and how artificial intelligence (AI), proxied by industrial robot adoption, affects employment in China. Motivated by concerns over technological unemployment amid rapid digitalization, it examines four questions: (1) whether automated processing has reduced jobs in China over the past fifteen years; (2) through which mechanisms AI influences employment (labour productivity, capital deepening, and refined division of labour); (3) whether digital-era organizational change via virtual agglomeration is a channel; and (4) if AI affects gender disparities and industry-specific employment differently. Positioning AI as a general-purpose technology within China’s digital transformation, the paper hypothesizes that AI increases employment overall (H1) and does so via productivity gains, capital deepening, and labour division (H2), and that virtual agglomeration is an additional channel (H3).

The literature reflects long-standing debate on technology’s labor-market effects, from job creation and higher productivity (Smith, Schumpeter) to displacement and technological unemployment (Sismondi, Marx). Modern perspectives highlight AI as distinct from prior automation because it can perform cognitive tasks, intensifying concerns about substitution across both low- and high-skill occupations. Evidence is mixed: studies find both positive employment effects via productivity, new tasks, and expanded demand, and negative effects via substitution, especially in manufacturing. Some research suggests short-run displacement but long-run job creation, and heterogeneity across skill, sector, and time horizons. Mechanisms supporting positive employment include FDI, growth, skill upgrading, industrial technological intensity, and reduced information frictions. This paper contributes by: using robot installation density to measure AI; analyzing heterogeneity by gender and industry attributes (finding relative gains for women and labor-intensive industries); and introducing virtual agglomeration as a new mechanism. It also leverages a machine-learning linear regression to assess AI’s contribution to employment growth and employs an instrumental variable (industrial electricity on-grid price trends) for causal identification.

Data: Panel data for 30 Chinese provinces (excluding Hong Kong, Macao, Taiwan, and Xizang) from 2006–2020. Core data on robot installation density from the International Federation of Robotics (IFR) matched to Chinese manufacturing via a shift-share (Bartik) approach aligning IFR’s 14 manufacturing categories to Chinese sectors using 2006 baseline employment shares and the Second National Economic Census. Employment scale (ES) is measured by the number of manufacturing employees in cities and towns. Variables: Explained variable: ES (manufacturing employment). Core explanatory variable: AI (provincial robot installation density, shift-share constructed). Mediators: Labour productivity (LP) computed from national accounts (ΔL = Y − kI), Capital deepening (CD) proxied by net fixed capital stock of industrial enterprises above designated size, Division of labour refinement (DLR) proxied by number of producer-services employees, and Virtual agglomeration (VA) measured as location-entropy-style agglomeration in information transmission/computer/software sectors weighted by regional internet broadband access ports. Control variables include road accessibility, R&D, wage costs, industrial structure, marketization, macro-control, urbanization, capital deepening, division of labour, and labour productivity (see descriptive statistics). Models: Baseline two-way fixed effects (province and year FE) linear regression: ES_it = δ0 + δ1 AI_it + Controls_it + μ_i + τ_t + ε_it. All variables are log-transformed to reduce volatility and heteroscedasticity; Hausman and F-tests support FE specification. Mediation tests estimate AI’s effect on mechanism variables (LP, CD, DLR, VA) under FE to assess channels (H2, H3). Robustness checks include: (1) replacing the dependent variable with manufacturing employment share of total employment; (2) augmenting controls with proxies for potential omitted variables (e.g., real estate prices, population density, human capital, unions/associations); (3) a machine-learning linear regression (gradient descent) to quantify AI’s contribution weight to employment changes. Endogeneity: Two-stage least squares (2SLS) using industrial on-grid electricity prices as an instrumental variable for AI, justified by their impact on firms’ robot adoption costs but not directly on employment demand; first-stage relevance and weak-IV tests reported satisfactory. Heterogeneity analyses by gender (male vs female employment) and by industry type (labour-intensive, capital-intensive, technology-intensive) are conducted within the FE framework. Estimation details: Significance assessed via t-statistics (clustered at province level, implied). Key tables: Descriptive statistics (Table 1). Baseline FE results (Table 2). Robustness and IV results (Table 3). Heterogeneity and mechanism estimates (Tables 4–5).

- Baseline effect: AI positively affects manufacturing employment. The two-way fixed-effects estimate (column (5)) reports a coefficient of 0.989, significant at the 1% level, indicating that increases in robot installation density are associated with higher employment, supporting H1. - Robustness: Across alternative outcome (manufacturing employment share), expanded controls, and ML-based importance weighting, the positive effect persists. Table 3 indicates: Method 1 coefficient 0.535 (t=8.31), Method 2 coefficient 0.978 (t=32.18), and ML importance weight 0.843 (i.e., AI explains 84.3% of the increase in employment scale). All are significant at conventional levels except the weight, which is an importance metric rather than a test statistic. - Endogeneity: 2SLS second-stage estimate shows AI’s positive effect on employment (0.239, t=5.03), alleviating concerns of reverse causality and omitted variables. - Heterogeneity by gender: AI promotes both male and female employment, with a relatively larger coefficient for females (Female: 1.032, t=20.99; Male: 0.966, t=32.47), suggesting AI may relatively improve women’s employment share. - Heterogeneity by industry type: Positive effects are strongest in labour-intensive industries (LI: 0.054, t=3.48) and smaller in capital-intensive (CI: 0.039, t=3.67) and technology-intensive (TI: 0.026, t=4.91) sectors. - Mechanisms (mediation): AI significantly increases capital deepening (CD: 0.052, t=3.59), labour productivity (LP: 0.071, t=3.31), and division of labour refinement (DLR: 0.302, t=4.93), each significant at 1%, supporting H2. AI also increases virtual agglomeration (VA: 0.141, t=2.63, 5% level), supporting H3. These channels contribute to net job creation by expanding scale, lowering costs, and creating new tasks and organizational forms. - Model selection and transformation: Hausman and F tests select fixed effects; variables are log-transformed to handle heteroscedasticity. Overall, results suggest job-creation effects exceed substitution effects over the study period.

Findings indicate that AI adoption via industrial robots in China increases manufacturing employment, countering pure displacement narratives. The positive net effect arises as firms expand output and scale in response to productivity gains, with capital deepening and refined task division generating new, more complex jobs alongside automated routine tasks. Virtual agglomeration emerges as a distinctive digital-era mechanism: cloud platforms and networked production amplify knowledge spillovers, flexible labor pooling, and intermediate input market scale, fostering additional employment beyond geographic clusters. Heterogeneity results suggest AI relatively bolsters women’s employment and benefits labour-intensive sectors more strongly, implying AI-enabled organizational change may reduce some traditional gender and sectoral constraints. Given China’s relatively early stage of robot diffusion and persistent low labor costs, substitution effects are not yet dominant, and firms’ "machine replacement" often responds to labor shortages. Overall, the results address the research questions by confirming AI’s positive role (H1), identifying productivity, capital deepening, and division of labor as key channels (H2), and establishing virtual agglomeration as an additional pathway (H3).

The study provides evidence from 30 Chinese provinces (2006–2020) that AI, proxied by industrial robot density, has a net positive effect on manufacturing employment. Mechanism tests show that AI promotes employment by increasing labor productivity, deepening capital, and refining the division of labor, while virtual agglomeration serves as a new, significant channel in the digital economy. Heterogeneity analyses reveal relatively larger gains for women and in labor-intensive industries. Policy recommendations include: accelerating high-end domestic robot and AI R&D; upgrading digital and intelligent infrastructure (e.g., 5G, cloud, IoT, industrial internet); fostering producer services and building collaborative innovation ecosystems; improving social security and employment monitoring systems to manage potential displacement; expanding unemployment protections and training; and reforming education and vocational training to build AI-complementary, interdisciplinary skills and lifelong learning pathways. Future research should integrate micro-level and case data, leverage quasi-experimental designs, and explore spatial spillovers and cross-country heterogeneity to generalize findings beyond China.

- External validity: Evidence is based on China’s 30 provinces during 2006–2020; conclusions may not generalize across countries or periods with different digital infrastructure, labor markets, or policy regimes. - Data and level of analysis: Macro/provincial panel with relatively coarse granularity limits identification and inference. Microdata (e.g., firm- and worker-level) and field studies are needed, especially post-COVID structural shifts. - Causal identification: While TWFE and IV (2SLS) improve identification, they fall short of strict causal inference. Lack of suitable policy shocks constrains quasi-experimental evaluation; future work could employ DID, RDD, or synthetic control when policy pilots (e.g., digital parks) are available. - Spatial spillovers: Potential cross-regional diffusion and spillover effects from robot adoption are not explicitly modeled. - Scope for generalization: Mechanisms may vary with digital infrastructure, workforce skills, and industrial structure across countries; expanding samples and comparing development stages would clarify heterogeneity.