Inequality Measures

Bottom Line

AI adoption is associated with wider wage and opportunity gaps unless organizations act. Natural experiments at firms and establishments find higher skill premia, job losses in non-AI tasks and junior roles, and higher top-end pay within adopting firms Liang et al. (2026); Aoki et al.; Shen (2026). Within cities, GenAI exposure clusters and coincides with wage stagnation in high-exposure neighborhoods after ChatGPT's release He et al. (2026).

Some tools and policies mitigate inequity, and in energy and carbon some inequality measures fall, but effects differ by place, sector, and group Ye et al. (2026); Fan (2026); Smith et al. (2026).

What This Means in Practice

• If you deploy AI, share productivity gains with workers (for example, profit-sharing or pay floors) and fund targeted upskilling. Otherwise expect larger skill pay gaps, pressure on routine roles, and more gains to capital.
• Track equity metrics continuously. In HR, procurement, and public services, pair simple fairness rules with policies that learn under uncertainty and invest in disadvantaged groups.
• Map local task exposure. Target wage and job supports where AI exposure and wage stagnation co-occur, especially in core districts and for junior, non-AI roles.
• Require distributional audits (who gains and who loses) before launch, and favor models with demonstrated accuracy and fairness in your domain.
• Fund portable, job-relevant digital skills and create on-ramps for under-represented groups with low AI exposure but high potential gains.

What the Research Finds

Labor-market inequality: skill premia, polarization, and who bears the losses

• AI adoption is associated with higher skill premia, wider wage gaps between higher- and lower-skill workers, and a shift from routine cognitive tasks to complex, interpersonal tasks; top-quintile wages rise while middle-quintile wages fall modestly Liang et al. (2026); A. T. D..
• At adopting establishments, productivity and output rise, but employment and total compensation fall, with losses concentrated in non-AI tasks and among non-senior workers Aoki et al..
• Macroeconomic models predict that treating AI as capital shifts income from labor to capital and rents, disproportionately affecting lower- and middle-skill workers Mici et al..
• Within firms, executive pay is higher with AI adoption, likely through financing and productivity channels, widening top-end pay dispersion Shen (2026).

Where and for whom: geographic, demographic, and sectoral divides

• In Beijing, high-exposure neighborhoods saw wage stagnation after ChatGPT's release even as high-skill workers moved in He et al. (2026).
• Across countries, exposure to AI substitution is higher in lower-income economies; women appear more exposed to substitutable tasks than men (descriptive atlas) Garg et al. (2026).
• Households show a generative-AI adoption divide: higher-income and younger households adopt faster and gain more leisure time (a natural experiment using browsing data) Blank et al..
• In India's graduate labor market, Scheduled Caste and Tribe graduates are much less represented in GenAI-exposed occupations within the same districts; because AI-exposed roles pay more, caste earnings gaps may widen Mishra (2026).

Inequality measures in use: what shifts on Gini, Theil, parity, and domain indices

• Policing simulations find persistent city- and year-specific disparities on Disparate Impact Ratio (selection-rate ratio), Demographic Parity Gap (selection-rate difference), and Gini (0 equals equality, 1 equals inequality); GAN-based debiasing (generative adversarial networks) only partly reduces them Barman and Barman (2026).
• Provincial Chinese data link AI development to lower interprovincial carbon inequality on Gini, via energy efficiency, monitoring, and allocation; effects are weaker on Theil (an entropy-based inequality index) Fan (2026).
• A multidimensional energy-justice index finds AI adoption improves overall, recognition, and procedural justice, but can initially worsen distributional justice; gains cluster in advanced eastern provinces Ye et al. (2026).
• New domain measures include a Comfort Equity Index for building control; equity-aware reward shaping improves modeled comfort equity but still leaves disparities for some groups Zaregarizi and Yavari (2026).

Algorithmic gatekeeping, fairness trade-offs, and practical mitigation

• Platform and algorithmic work management can reproduce or amplify inequalities through biased data and opaque rules; reviews also report inclusion benefits when safeguards exist Rouco et al. (2026); Kohli (2026).
• In sequential decisions, uncertainty and selective feedback (only observing outcomes for selected cases) harm under-represented groups more; uncertainty-aware exploration can reduce disparities without lowering primary objectives (theory and simulations) Lee et al. (2026).
• Static group-fairness constraints can worsen long-run disparities, while simple investment-oriented policies can remove them at low cost in stylized models Jabbari and Wang (2026).
• In property taxation, richer models improved both accuracy and fairness across 26 million sales, countering a presumed fairness-accuracy trade-off Smith et al. (2026).

New since 2026-04-06: reinforcement or change in balance

• New natural experiments and other studies add to distributional concerns: firm-level skill premia in China, establishment-level displacement in the U.S., and intra-urban clustering with wage stagnation in Beijing all point to widening gaps without countermeasures Liang et al. (2026); Aoki et al.; He et al. (2026).
• Mitigation looks more actionable: a randomized controlled trial shows simple scaffolding (checklists or prompts) narrows variance in AI-assisted performance for early-career knowledge-worker analogs, and field-scale AI feedback raises participation most for less-embedded teams Idan and Anand (2026); Wang et al. (2026).

What We Still Don't Know

• Long-run, causal distributional paths: we lack multi-year tracking of wage and employment distributions within firms, cities, or sectors after AI adoption using credible designs; most evidence is short-run or cross-sectional Aoki et al.; Liang et al. (2026).
• Under-represented contexts: evidence on AI's inequality effects in LMICs (low- and middle-income countries), informal economies, and across caste and gender at scale is sparse relative to advanced economies Garg et al. (2026); Mishra (2026).
• Rent capture and institutions: direct measurement of how ownership, bargaining, and antitrust interventions shift the split between labor and capital in AI-intensive markets is limited beyond theory and simulation Mici et al..
• Fairness dynamics over time: which fairness interventions reduce disparities long-run in real deployments, not only in simulations, remain to be tested in domains beyond tax assessment Jabbari and Wang (2026); Smith et al. (2026).
• Standardized inequality dashboards: there is no widely adopted, real-time, task-based national system linking AI exposure to distributional outcomes across skill, industry, region, race, and gender, though operational frameworks have been proposed Joshi (2026).