Women in the UK lag men in generative-AI adoption largely because they worry more about societal risks—not because of access or skills; boosting optimism about AI could raise young women's GenAI use from about 13% to 33%, markedly narrowing the gender gap.

Main Finding

Women in the UK adopt generative AI (GenAI) substantially less than men largely because they perceive greater societal risks from AI (mental health, privacy, climate, labour‑market impacts). Those risk perceptions explain a meaningful share of variance in adoption (9–18% of feature importance in predictive models) and are a stronger predictor of use for women than conventional factors like digital literacy or education — especially among younger adults. Shifting societal optimism about AI increases GenAI uptake among young women (from ~13% to ~33% in the synthetic‑twin analysis), substantially narrowing the gender gap.

Key Points

• Sample and baseline gap:
  • Data: UK Public Attitudes to Data and AI Tracker, 2023–2024; N ≈ 8,000 (nationally representative).
  • Frequent personal GenAI use: women 14.7% vs men 20.0% → baseline gender gap = 5.3 percentage points.
• Role of risk perceptions:
  • Composite risk index (average of four binary items: concerns about mental health, climate, data privacy, labour market) has mean ≈ 0.20.
  • Risk perceptions explain 9–18% of total feature importance in random‑forest models and rank among the top predictors for women across age groups.
  • Concerns depress women’s personal GenAI use more than men’s (e.g., mental‑health concern associated with a 16.8 pp gender gap).
• Intersectionality:
  • Largest disparities among younger, digitally fluent respondents who also hold high societal‑risk concerns: gender gaps in personal use can exceed 45 percentage points (e.g., mental‑health concern + high digital literacy).
  • Work‑use gaps are generally smaller than personal‑use gaps, but specific intersections (privacy × digital fluency, mental‑health concern) still show large gaps.
• Predictive modeling:
  • Gender‑specific random forests stratified by age (18–35, 36–50, 51+) produced stable performance (AUC ≈ 0.63–0.67).
  • For young women, risk perceptions ranked 2nd in importance; for young men, they ranked much lower.
• Synthetic‑twin (panel) results:
  • Synthetic matches between survey waves used to simulate two “interventions”: improved digital literacy and increased societal AI optimism (100 iterations).
  • Digital‑literacy gains raise adoption for both sexes but tend to widen the gender gap among young adults (example: young men 19%→43% vs young women 17%→29%).
  • Increased societal optimism raises uptake more among women, narrowing the gap (young women 13%→33%; young men 21%→35%).
• Limitations:
  • Self‑reported adoption and literacy (may understate / reflect confidence).
  • Synthetic‑twin design is not a randomized experiment — causal claims are suggestive.
  • UK‑specific context; international patterns may differ.

Data & Methods

• Data source: UK Department for Science, Innovation & Technology — Public Attitudes to Data and AI Tracker, Waves 3–4 (2023–2024), nationally representative sample ≈ 8,000 respondents.
• Key measures:
  • Outcome: frequent (≥weekly) personal and work use of GenAI (self‑reported).
  • Predictors: digital literacy (ordinal/self‑reported), education, occupation, age, and a composite AI risk index (four binary items: concerns about mental health, climate, data privacy, labour‑market impacts).
• Analyses:
  • Descriptive cross‑tabs and intersectional heatmaps to document subgroup gaps.
  • Gender‑specific random‑forest classifiers stratified by age groups to estimate feature importance (non‑parametric, captures nonlinearities). Reported AUC ≈ 0.63–0.67.
  • Synthetic‑twin panel design: match Wave 3 respondents to Wave 4 look‑alikes on age, gender, education, occupation; run 100‑iteration simulations to estimate how increases in digital literacy or societal optimism associate with changes in GenAI use between waves.
• Supplementary checks: intersectional tables (e.g., Table 1) showing ordered gender gaps across covariates.

Implications for AI Economics

• Distributional outcomes and inequality:
  • Differential GenAI adoption by gender risks amplifying productivity, visibility, and earnings gaps as AI becomes embedded in work. Early male‑skewed adoption can compound into persistent labor‑market advantages.
  • If models and tools are shaped primarily by male usage patterns, downstream product bias and poorer performance for underrepresented prompting styles/tasks may reduce women’s returns to AI use — a feedback loop increasing inequality.
• Skill formation and labour markets:
  • Capability (digital‑skills) interventions alone may not close the gap and can unintentionally widen it among young adults because men appear to leverage improved skills into greater adoption more readily.
  • Policies and training should combine skills development with efforts that address value‑based concerns (e.g., demonstrations of safeguards, transparent governance) to achieve equitable uptake.
• Model training and externalities:
  • Gender‑skewed user data could bias future GenAI systems, affecting quality and fairness. Economists and practitioners should account for user heterogeneity when estimating adoption externalities and when designing sampling/feedback loops for model improvement.
• Policy interventions and regulation:
  • Two complementary approaches:
  • Improve underlying technologies and governance to address legitimate societal concerns (lower carbon models, stronger bias mitigation, privacy protections, supply‑chain transparency, wellbeing safeguards). This reduces objective harms and makes perception shifts more credible.
  • Targeted communication and interventions that change societal risk perceptions (evidence on benefits, credible oversight, case studies of safe use) — these appear particularly effective at increasing women’s adoption without creating a male‑biased uptake.
  • Regulatory design should consider how incentives for safer, more transparent AI affect both objective risk and perceived acceptability, since perception shifts can change behavior meaningfully.
• Research priorities for AI economics:
  • Establish causality: randomized trials or field experiments that manipulate information about societal risks/benefits and measure subsequent adoption and labor outcomes.
  • Cross‑country comparisons and firm‑level studies to see how institutional contexts mediate perception–adoption links.
  • Use objective usage telemetry to validate self‑report findings and to study how adoption translates into productivity and wages by gender.
  • Model the longer‑run general equilibrium effects of gendered adoption on human capital accumulation, wages, and inequality, including feedbacks through model bias and product design.

Summary takeaway: unequal GenAI adoption in the UK is driven less by access or basic skills and more by gendered differences in perceptions of AI’s societal risks. Addressing that gap requires both technological and governance improvements to reduce real harms and credible, targeted interventions to shift perceptions — otherwise early adoption asymmetries risk entrenching future economic inequalities.