AI Technology Application and Employee Responsibility

The paper investigates how the application of AI technology within firms affects employee responsibility, an essential component of corporate social responsibility. Grounded in behavioral science and agency theory, the authors argue that AI reshapes supervision and incentive mechanisms, potentially reducing firms’ willingness to take responsibility for employees. They pose key research questions: (1) What is the effect of AI technology application on employee responsibility? (2) Do product market competition and government ownership moderate this relationship? (3) Does supervision cost mediate the relationship? (4) How do AI application and employee responsibility differentially affect productivity and innovation? The study focuses on Chinese listed manufacturing firms from 2011–2020 to provide firm-level empirical evidence and to reconcile mixed findings in the literature regarding AI’s workplace impacts.

Prior research on AI has predominantly examined employment and productivity, yielding replacement vs. augmentation views and mixed implications for productivity. Some studies highlight AI’s potential to automate tasks and fuel innovation, while others caution about persistent productivity slowdowns due to inequality, learning costs, and limited disruption rates. From a social and legal perspective, AI in workplaces raises concerns over surveillance, transparency, privacy, and discrimination. Behavioral science emphasizes employees as “social people,” advocating people-oriented management and incentives to enhance motivation, loyalty, and performance. Agency theory frames the manager–employee relationship as a principal–agent problem addressed via supervision and incentives. Employee-related CSR is shown to improve satisfaction, loyalty, innovation, and productivity, functioning as an effective incentive tool. The authors suggest that AI lowers supervision costs and shifts management focus, potentially reducing the need for people-oriented incentives (employee responsibility). They hypothesize: H1, AI application is negatively related to employee responsibility; H2, this negative relationship weakens with greater product market competition; H3, the negative relationship is stronger in government-controlled firms than privately controlled firms.

• Sample: Chinese manufacturing firms listed on Shenzhen or Shanghai stock exchanges, 2011–2020. Final unbalanced panel after merging sources and excluding missing key variables: 2,395 firms; total observations ≈ 14,267 for main regressions.
• Data sources: CSMAR (China Stock Market & Accounting Research) for financial, ownership, governance, and other firm data; Hexun.com for CSR sub-scores including employee responsibility.
• Measures: • Dependent variable (ER): Employee responsibility score from Hexun.com, covering performance (e.g., per capita income), safety (inspections, training), and care (welfare/caring programs) dimensions. • Key independent variable (AI application): Text mining of annual reports (Python crawler + JavaPDFBox). Keyword frequency for AI-related terms mapped to AI’s four capabilities (perception, comprehension, action, learning), including terms like artificial intelligence, business intelligence, machine learning, deep learning, biometrics, face/speech recognition, autonomous driving, NLP, etc. Excluded negations and non-firm contexts. Summed counts, took natural log of total frequency. Robustness: binary measure AI_D = 1 if any AI term appears (≥1), else 0. • Moderators: Product market competition measured by Herfindahl-Hirschman Index (HHI) based on industry sales; higher HHI = less competition. Alternative: HHI_A based on total assets. Government ownership (SOE) dummy = 1 if ultimately government-owned, else 0. • Controls: Firm characteristics (size, age, leverage, ROA, diversification per Berry index, R&D intensity) and corporate governance (ownership concentration, board size, independent director ratio, CEO duality, executive shareholding, TMT functional heterogeneity per Blau-type index, age/tenure averages and diversities, education average and diversity, female proportion), plus industry (CSRC 2-digit under C) and year fixed effects.
• Model: Multivariable linear regressions with fixed effects; Driscoll–Kraay standard errors to address heteroskedasticity and cross-sectional dependence. Baseline: ER = α + β1 AI + β2 AI×HHI + β3 AI×SOE + Controls + Industry FE + Year FE + ε.
• Robustness: Address endogeneity with one-period lags (LAI, LHHI, LSOE); Propensity score matching (logit; one-to-one NN matching and also caliper/radius/kernel) defining treated firms by above-average AI intensity within year–industry; alternative measures for AI (AI_D) and competition (HHI_A).
• Mediation analysis: Supervision cost (MC) proxied by ln(management expenses per employee). Tested paths: AI → MC; MC → ER; AI → ER controlling for MC; Sobel test.
• Further analyses: Compared effects of AI and ER on production efficiency and innovation. Production efficiency measured by firm-level TFP via three estimators (EOLS, EOP, ELP). Innovation output (PA) = ln(patent applications); Innovation efficiency (PE) = PA / ln(R&D expenditure).

• Main effect (H1): AI application is negatively associated with employee responsibility. • Baseline regression: AI → ER β ≈ -0.089, p < 0.05. • Correlation ER–AI = -0.019 (p < 0.01).
• Moderation by product market competition (H2): Interaction AI × HHI negative and significant. • β ≈ -1.233, p < 0.01 (baseline). Interpretation: as competition increases (lower HHI), the negative AI–ER relationship weakens. • Robustness with HHI_A: AI × HHI_A β ≈ -1.660, p < 0.05.
• Moderation by ownership (H3): Interaction AI × SOE negative and significant. • β ≈ -0.451, p < 0.01. Negative effect stronger in government-controlled firms.
• Mediation by supervision cost (MC): Partial mediation supported. • AI → ER: β ≈ -0.089, p < 0.05 (without mediator); with mediator, β ≈ -0.067, p < 0.05. • AI → MC: β ≈ -0.034, p < 0.05 (AI reduces supervision cost). • MC → ER: β ≈ 0.656, p < 0.01. • Sobel Z = -4.392, p < 0.01.
• Robustness checks: • Lagged models: LAI → ER β ≈ -0.099, p < 0.1; interactions with LHHI (β ≈ -0.799, p < 0.05) and LSOE (β ≈ -0.511, p < 0.01) hold. • PSM (1-to-1 NN): AI negative main effect and significant interactions with HHI and SOE persist. • Alternative AI measure (AI_D): Main negative effect (β ≈ -0.165, p < 0.01); interactions with HHI (β ≈ -1.128, p < 0.1) and SOE (β ≈ -0.756, p < 0.01) consistent.
• Comparative effects on performance: • Production efficiency (TFP): Both ER and AI positively associated.
  • ER: ELP β ≈ 0.004 (p < 0.05); EOLS β ≈ 0.004 (p < 0.05); EOP β ≈ 0.005 (p < 0.01).
  • AI: ELP β ≈ 0.014 (p < 0.05); EOLS β ≈ 0.009 (p < 0.1); EOP β ≈ 0.011 (p < 0.05). • Innovation performance:
  • ER → Innovation output (PA): β ≈ 0.019, p < 0.05; → Innovation efficiency (PE): β ≈ 0.001, p < 0.05.
  • AI → PA: β ≈ -0.103, ns; → PE: β ≈ -0.002, ns.

The findings indicate that AI application, by lowering supervision costs and shifting managerial focus toward data-driven monitoring, reduces firms’ reliance on people-oriented incentive mechanisms like employee responsibility. This aligns with agency theory: when monitoring becomes cheaper and more comprehensive, principals substitute supervision for intrinsic/extrinsic incentives. Yet, in competitive product markets where creativity, discretion, and emotion-based tasks matter, firms benefit from maintaining higher employee responsibility, weakening the negative AI–ER link. Ownership structure matters: government-controlled firms, with more routine tasks and politically oriented objectives, exhibit a stronger negative AI–ER relationship. Importantly, while AI improves production efficiency, it does not enhance innovation output or efficiency; conversely, employee responsibility supports both productivity and innovation. Thus, in the AI era, human-centric practices remain critical for innovation and sustained competitive advantage.

This study contributes firm-level evidence that AI technology application is negatively related to employee responsibility, with supervision cost partially mediating this relationship. The negative effect is attenuated by greater product market competition and is stronger in government-controlled firms. While both AI and employee responsibility improve production efficiency, only employee responsibility significantly boosts innovation output and efficiency. The results underscore the continued importance of people-oriented management alongside AI adoption. Practical implications include the need for managers to balance automated supervision with policies that sustain employee well-being, trust, and motivation, and for policymakers to update protections around surveillance, privacy, and fairness in AI-enabled workplaces. Future research should refine AI measurement to capture heterogeneity in AI types, scope, and adoption timing, and examine causal mechanisms using designs that better address endogeneity and dynamic effects.

• AI application is measured via annual report keyword frequency, which may not capture heterogeneity in AI types, intensity, deployment scope, or timing across firms.
• The sample focuses on Chinese listed manufacturing firms; generalizability to other sectors or countries may be limited.
• Despite robustness checks (lags, PSM, alternative measures), residual endogeneity and omitted variables may remain.
• Employee responsibility relies on third-party index data; measurement error or disclosure bias is possible.