Generative AI lifts workplace productivity through automation and decision support, but gains are uneven and accompanied by worker concerns about skill obsolescence and job security; evidence is concentrated in developed economies, leaving the effects in developing countries poorly understood.

Main Finding

Generative AI (LLMs and tools like ChatGPT and Copilot) produces substantial and robust productivity gains in knowledge-intensive tasks, but these gains are highly heterogeneous and accompanied by mixed employee perceptions and rising job insecurity. The net effects depend on task-technology fit, organizational deployment logic (automation vs. augmentation), worker skill, and institutional context—with a pronounced informational and capability divide between developed and developing economies.

Key Points

• Productivity

  • Consistent empirical evidence of large gains in knowledge work: examples reported include ~40% faster task completion (Noy & Zhang, 2023), ~15% higher customer-service output (Brynjolfsson et al., 2025), and ~55.8% faster coding with Copilot (Peng et al., 2023).
  • Gains are conditional: LLMs improve performance within their capability frontier but can degrade outcomes on tasks beyond that frontier (Dell'Acqua et al., 2023).
  • Mechanisms: task automation, decision support, and knowledge augmentation; strongest effects in language-, synthesis-, and information-heavy tasks.
  • Heterogeneity: effects vary by occupation, individual openness/skill, and implementation design; evidence concentrated in developed, knowledge-intensive sectors.

• Employee perceptions

  • Ambivalent attitudes: many workers report increased efficiency, confidence, and satisfaction when AI is framed/implemented as a complementary resource.
  • Negative reactions cluster around identity threat, reliability/bias concerns, surveillance, and poor change management, producing resistance even when productivity benefits exist.
  • Implementation factors (training, psychological safety, participation) materially affect perceptions.

• Job insecurity and labor effects

  • Job insecurity is a key mediator linking AI adoption to employee wellbeing, commitment, and behavior (e.g., upskilling, turnover intention).
  • LLM exposure is greatest among high-wage, high-education occupations—reversing earlier automation patterns and producing potential middle-skill displacement and labor-market polarization.
  • Reskilling programs exist but are uneven; anticipatory insecurity causes organizational harms even before displacement materializes.

• Developing vs developed economies

  • Literature is geographically skewed toward developed countries; evidence from South Asia, Sub-Saharan Africa, and Latin America is sparse.
  • Developing-economy barriers: weak digital infrastructure, talent shortages, regulatory gaps, limited change-management capacity.
  • AI benefits in developing contexts appear concentrated in large firms and high-skill workers, risking widened inequality (the "Generative AI Divide").

• Theoretical framing

  • Review draws on Task-Technology Fit (task-specific gains), Job Demands-Resources (AI as both resource and demand), and the Automation–Augmentation paradox (trade-offs between short-term substitution gains and long-term engagement/adaptability).

Data & Methods

• Review type: Systematic literature review following PRISMA; narrative synthesis (no meta-analysis due to heterogeneous designs/measures).
• Timeframe and scope: Studies published January 2020–December 2025; focus on generative AI, LLMs, ChatGPT, Copilot in workplace contexts.
• Search and selection:
  • Databases searched: Google Scholar and Dimensions.ai; plus hand-searching of references (8 additional records).
  • Initial hits: 3,252 records (1,940 Google Scholar; 1,312 Dimensions.ai); after deduplication 2,892 unique records.
  • Screening: title and abstract screening reduced pool to full-text review; final included studies = 40.
• Inclusion criteria: studies addressing generative AI in workplace settings reporting on productivity, perceptions, job insecurity, displacement or skill change; empirical, review, or conceptual studies; English; full text accessible.
• Evidence types: randomized controlled trials, field experiments, controlled experiments, surveys, qualitative studies, systematic reviews, conceptual papers, country analyses.
• Quality assessment: three-tier framework applied by study design (internal validity for experiments, representativeness and measurement validity for surveys, analytical rigor for qualitative/conceptual). Greater weight given to peer-reviewed experiments and quasi-experiments when resolving conflicts.
• Limitations of methods noted: reliance on Google Scholar/Dimensions (Scopus/Web of Science not accessed), geographic bias in primary studies, and heterogeneity precluding pooled effect estimates.

Implications for AI Economics

• For empirical modeling and measurement

  • Treat productivity gains as conditional and task-specific rather than uniform: incorporate task-technology fit measures and heterogeneity across worker skill distributions.
  • Include job insecurity and psychological responses as mediating variables when estimating labor supply, effort, and turnover costs.
  • Expand outcome sets beyond output/time (quality, autonomy, innovation, wellbeing, reskilling behavior).
  • Prioritize causal designs (field RCTs, quasi-experiments) and representative longitudinal surveys to capture short- and medium-run dynamics.

• For labor-market and distributional analysis

  • Expect potential polarization: rising demand at top-skilled and low-skilled tasks with displacement risk for middle-skilled administrative roles; model heterogeneous wage and employment effects accordingly.
  • Incorporate firm-size and sectoral heterogeneity—large firms and frontier firms may capture disproportionate gains, amplifying within-country inequality.
  • Evaluate redistribution and social-insurance policies (training subsidies, wage insurance, active labor-market policies) in welfare and policy-counterfactual models.

• For cross-country and development economics

  • The "Generative AI Divide" implies different aggregate productivity and distributional outcomes across countries; models should account for infrastructure, human capital, regulatory capacity, and adoption asymmetries.
  • Research and policy should prioritize capacity-building: digital infrastructure, training ecosystems, regulation that addresses bias and surveillance, and support for SMEs to adopt augmentation-focused AI.

• For firm strategy and policy prescriptions

  • Augmentation-first deployment (design AI to complement workers) is likelier to yield sustainable productivity and preserve human capital—models of firm behavior should include managerial choice over automation vs. augmentation.
  • Policies encouraging transparent change management, worker participation, and subsidized upskilling can reduce insecurity-driven frictions and improve realized productivity.
  • Regulatory and governance frameworks need to be context-sensitive: universal rules may misfire where institutional capacity and labor-market structures differ.

• Research gaps / priorities

  • Comparative empirical work across developing economies and sectors beyond knowledge-intensive settings.
  • Longer-term longitudinal studies on career trajectories, wages, and firm-level investment dynamics.
  • Meta-analytic synthesis once standardized effect measures accumulate.
  • Studies measuring how specific implementation choices (incentives, monitoring, governance) mediate productivity vs. wellbeing trade-offs.

Summary conclusion: Generative AI can raise aggregate and individual productivity substantially, but economic impacts depend critically on tasks, institutions, and deployment choices. AI economics research and policy should shift from asking whether generative AI matters to asking how, for whom, and under what governance regimes the gains are realized and equitably distributed.