AI's effect on income inequality is conditional, not universal: across 20 OECD countries (2015–2019) no single factor drives changes in distribution; instead, 14 distinct combinations of AI and socioeconomic conditions produce different outcomes for the Gini, top 1% and bottom 50% income shares.

Main Finding

Using dynamic fuzzy-set QCA on a balanced panel of 20 OECD countries (2015–2019), the study finds that artificial intelligence (AI) is neither uniformly equalizing nor uniformly inequality‑widening. No single condition (AI indicator or socioeconomic factor) is necessary for a given income-distribution outcome. Instead, AI’s impact on inequality depends on its configuration with socioeconomic conditions: the authors identify 14 sufficient conditional configurations (grouped into 7 AI‑centered pathway types) that produce different outcomes for overall inequality (Gini), the top 1% share, and the bottom 50% share. Policy conclusions emphasize balanced, context‑sensitive strategies that align AI development with social and economic fundamentals to mitigate inequality risks.

Key Points

• Sample & scope: 20 OECD countries, 2015–2019 (balanced panel). Outcomes studied: Gini coefficient (overall inequality), top 1% income share (TOP), bottom 50% income share (BOT).
• Core AI condition variables: AISR — AI scientific research (AI publications per capita); AIAD — AI application development (relative contributions to public AI projects); AII — AI venture investment (VC in AI startups).
• Edge (socioeconomic) condition variables: URBAN (urbanization), OPEN (KOF globalization index), HUM (human capital per capita), TECH (ICT capital compensation share).
• Method: dynamic fuzzy‑set QCA (fsQCA) applied to panel data, using between‑group (BECONS), within‑group (WICONS), and pooled consistency measures; calibration used 95% / 50% / 5% quantile anchors for fuzzy membership.
• Main empirical results:
  • No individual condition met QCA thresholds to be deemed necessary for high/low inequality.
  • 14 sufficiency configurations were found and grouped into seven AI‑centered pathways showing how combinations of AI activity and socioeconomic context lead to divergent inequality outcomes.
  • AI’s effects are heterogeneous across outcome measures (overall Gini vs top and bottom income shares): some configurations associate AI presence with rising top shares or rising overall inequality, while others link AI plus enabling socio‑economic conditions to improvements at the bottom or lower overall inequality.
• Limitations noted by authors: balanced panel requirement constrained the sample period to 2015–2019 (pre‑COVID & pre‑most recent AI waves), so results capture trends up to 2019; configurations are descriptive/exploratory (fsQCA identifies pathways rather than causal magnitudes).

Data & Methods

• Data sources:
  • Income distribution: World Inequality Database (pre‑tax Gini, top 1%, bottom 50%).
  • Socioeconomic controls: World Bank (urbanization), KOF (globalization/opening index), Penn World Table (human capital), Total Economy Database (ICT capital share).
  • AI indicators: OECD.AI database (AI publications per capita; GitHub‑based measure of public AI project contributions; Preqin‑derived VC investments in AI startups).
• Sample selection: retained 20 OECD countries with complete data across all variables and years (2015–2019), resulting in N=20, T=5.
• Calibration: direct fuzzy‑set calibration using objective quantile anchors (95%, 50%, 5%) for full membership, crossover, and full non‑membership.
• Analytical approach: dynamic fsQCA on panel data to assess necessity and sufficiency:
  • Necessity analysis via adjusted‑distance thresholds (stricter 0.1 cutoff used).
  • Sufficiency analysis to extract solution configurations (14 configurations → 7 pathway types).
  • Use of BECONS (between groups), WICONS (within groups), and pooled consistency to evaluate stability across cases and over time.

Implications for AI Economics

• Heterogeneity and configuration dependence: Economists and policymakers should avoid simplistic, global claims that “AI increases inequality” or “AI reduces inequality.” The distributional impact of AI depends crucially on co‑existing socioeconomic conditions (human capital, ICT infrastructure, openness, urbanization) and the particular mix of AI research, application development, and venture investment.
• Policy levers matter: To channel AI toward more equitable outcomes, policies should focus on:
  • Complementary investments in human capital and digital/ICT infrastructure so AI complements rather than substitutes broad segments of labor.
  • Inclusive innovation ecosystems (supporting AI application diffusion and absorptive capacity across firms and regions), not only concentrated VC flows to a narrow set of startups.
  • Social and labor policies (training, mobility, redistribution) that address short‑run displacement risks while enabling long‑run creation effects to materialize.
• Targeting income tiers: Since AI’s effects differ for top‑end concentration and bottom‑end welfare, policies should be calibrated: e.g., taxation and capital‑income regulation to limit top concentration when AI‑driven capital gains dominate; active labor market and reskilling programs where AI complements create middle‑skill opportunities.
• Research implications: Configuration methods like dynamic fsQCA are useful complements to regression approaches in AI economics because they uncover multiple sufficient pathways and interactions. Future studies should extend the time window (post‑2019 data), expand country coverage (non‑OECD and lower‑income contexts), and combine QCA with causal inference methods to estimate effect magnitudes.
• Caution on generalization: Findings apply to the sampled OECD context (2015–2019). As AI technologies, markets, and policy responses evolve rapidly, continual monitoring with updated data is needed.

Reference (article): Qian Y., Xu Z., Skare M., Lai K‑h., Qin Y., Gou X. (2026). The role of artificial intelligence in income distribution: a dynamic fsQCA study based on 20 OECD countries from 2015 to 2019. Humanities and Social Sciences Communications. https://doi.org/10.1057/s41599-026-08263-z

If you want, I can (a) extract and summarize the seven pathway types in more detail (based on the paper’s solution tables), or (b) produce a short policy brief drawing concrete recommendations for an OECD country of your choice.