Generative AI in India appears to accentuate a skills divide: graduates in AI-exposed occupations earn materially higher wages while lower-educated workers face wage declines and greater informalisation; aggregate unemployment fell before large-scale GenAI diffusion, so there is no clear evidence of net job loss attributable to AI.

Title: Long-Term Effects of Artificial Intelligence on Employment and Wages in India Author: Kowser Ali Jan (Annamalai University) — DOI: https://doi.org/10.21203/rs.3.rs-10648911/v1

Main Finding

AI diffusion in India is associated with pronounced skill polarisation rather than aggregate job destruction: graduates in high generative-AI–exposure occupations capture a significant wage premium in the post-GenAI period, while workers below graduate level face a wage penalty and higher likelihood of informalisation. National aggregate trends (PLFS, EPFO) show improving headline labour-market indicators, but these trends largely predate large-scale generative-AI diffusion and cannot by themselves be attributed to AI.

Key Points

• Dual-track empirical strategy:
  • Track A: calibrated simulated PLFS-style microdata panel (two waves) used to run a full inferential-statistics battery (normality diagnostics, t-tests, ANOVA, correlations, chi-square, multiple regression, DiD, triple-difference, logistic regression).
  • Track B: analysis of official aggregates (PLFS unemployment, worker-population, labour-force participation 2017–18 to 2023–24; EPFO payroll releases 2024–25) to triangulate simulation findings against real series.
• Main quantitative results (selected):
  • Graduate wage premium in high AI-exposure occupations (post-GenAI triple-interaction coefficient = 0.907, p < .001).
  • Wage penalty and greater informalisation for non-graduates (DiD coefficient = −0.126, p < .001).
  • Significant association between AI exposure and informal employment (χ²(2) = 234.90, p < .001; Cramér's V = 0.217).
  • PLFS aggregates: unemployment fell from 6.0% (2017–18) to 3.2% (2023–24); trend slope = −0.51 percentage points/year (p < .001) — but this decline began before large-scale GenAI diffusion.
• Contextual evidence and calibration:
  • Calibration and treatment effects were anchored to international occupational AI-exposure indices (Felten et al., Eloundou et al., ILO) and India-specific industry reports (NASSCOM) and studies (Mishra 2026). NASSCOM and EPFO aggregates show continued formal-sector additions and rapid growth in premium AI roles even as overall sectoral employment growth is modest.
• Data-access transparency: unit-level PLFS/EPFO/ESIC microdata were not accessible in the study environment; the paper explicitly builds and documents a calibrated simulated microdata panel to enable inferential analysis while checking conclusions against official aggregates.

Data & Methods

• Data sources:
  • Intended microdata: Periodic Labour Force Survey (PLFS) unit-level variables (occupation, education, gender, location, earnings, employment type) — not directly accessed.
  • Administrative aggregates: EPFO monthly payroll releases, PLFS published aggregate series (unemployment rate, labour-force participation, worker-population ratio).
  • External indices and reports for calibration: Felten et al. (AI exposure), Eloundou et al. (task automation), ILO refined GenAI index, NASSCOM, Mishra (2026).
• Simulated microdata (Track A):
  • Two-wave panel structured to reproduce PLFS marginal distributions of occupation, education, gender, rural/urban, wage levels, and informality shares.
  • An explicit, literature-anchored AI-exposure treatment effect was embedded to investigate augmentation vs automation channels and skill-polarisation.
• Statistical methods:
  • Descriptive stats and diagnostics (normality).
  • Group comparisons: independent-samples t-tests, one-way ANOVA with Tukey post-hoc.
  • Associations: Pearson/Spearman correlations, chi-square tests (with effect sizes).
  • Modeling: multiple linear regression; difference-in-differences (DiD); triple-difference (to capture graduate × AI-exposure × post-GenAI effects); logistic regression for informalisation probability.
• Robustness and triangulation:
  • Results from the simulated panel were compared with observed PLFS/EPFO aggregates (Track B) and literature findings to assess consistency and external plausibility.
• Limitations & caveats:
  • The microdata analysis relies on a calibrated simulation because unit-level microdata access was unavailable; simulated treatment effects were embedded deliberately and thus causal claims from Track A depend on the calibration assumptions.
  • Attribution of aggregate changes to AI is limited because pre-GenAI trends account for much recent movement in headline series.

Implications for AI Economics

• Mechanism emphasis: The paper supports a task-based, augmentation-versus-automation framing for AI’s labour effects — generative AI amplifies returns to occupations and tasks where it complements skilled labour and displaces (or reduces opportunities for) lower-skilled tasks.
• Distributional outcomes: AI is likely to increase wage inequality and formality divides — graduates in exposed occupations gain, while less-educated workers face wage pressure and higher informalisation risk. This has implications for inequality along education, caste, and gender lines (echoing Mishra 2026 findings).
• Policy priorities:
  • Targeted reskilling/upskilling: focus on moving mid- and lower-skill workers into AI-complementary tasks (not only generic digital literacy but task-specific augmentation skills).
  • Social protection and inclusion: expand safety nets and portable benefits to informal workers who face higher displacement/informalisation risk; consider improving access to formal-sector pathways.
  • Labour-market monitoring: collect and publish richer, timely microdata on occupation-task composition and AI exposure disaggregated by education, caste, gender, region, and formality to detect distributional impacts early.
  • Differentiated industrial policy: support sectors and firm types where AI augments labour to create high-quality jobs, while creating transition support for firms/occupations where automation is concentrated.
• Research implications:
  • Need for unit-level, longitudinal microdata access in India to test and refine causal estimates without heavy reliance on simulated panels.
  • Further work to disentangle AI-specific effects from pre-existing trends, to separate automation-type versus augmentation-type deployments, and to assess long-run general-equilibrium effects (demand creation, new task creation).
• Bottom line: Policymakers should not equate AI diffusion with mass aggregate unemployment in India, but should prepare for a pronounced re-shaping of the wage and formality structure — requiring targeted skills, protection, and data interventions to manage distributional risks.

If you want, I can extract or format the key reported statistics and coefficients into a short table (simulated effects, test statistics, p-values, and effect sizes) for quick reference.