AI diffusion appears linked to reduced work intentions among low-educated pre-retirement workers in high-adoption regions, while men and the highly educated are more willing to keep working; household financial pressure further suppresses employment intent, implying the need for targeted retraining and social supports.

Main Finding

In a cross-sectional survey of 889 pre-retirement individuals in Beijing, Guangzhou, and Lanzhou, greater exposure to AI and delayed-retirement policy contexts is associated with heterogeneous effects on middle-aged workers’ willingness to remain in or re-enter employment. Men, more-educated respondents, and those in technical occupations show significantly higher employment intentions; by contrast, lower-educated workers (especially in eastern and northern regions) and those facing greater household economic pressure report lower willingness to work. The authors conclude that AI’s labor-market impacts vary by demographic, occupation, and region and that targeted vocational training and social supports are needed to preserve employability among middle-aged workers.

Key Points

• Sample and outcome

  • N = 889 valid responses from individuals aged ~50+ (females near statutory 50, males near 60).
  • Outcome: employment intention (willingness to remain in work, seek re-employment, do part-time work, or continue full-time employment).

• Strong, statistically significant associations (multivariate regression)

  • Gender (male): β = 0.48, SE = 0.11, p < 0.01 → men report higher employment intention.
  • Education (bachelor’s+): β = 0.52, SE = 0.10, p < 0.01 → higher education increases willingness.
  • Occupation (technical): β = 0.32, SE = 0.09, p < 0.01 → technical workers more willing to continue working.
  • Family economic pressure: β = -0.39, SE = 0.08, p < 0.01 → greater pressure associated with lower employment intention.

• Regional heterogeneity

  • Eastern and northern regions: lower-educated respondents show significantly lower employment intentions (β = -0.45, p < 0.05).
  • Eastern region: stronger perceived risk of AI-related job replacement (β = 0.40, p < 0.05) but also higher willingness to take digital skills training (β = 0.58, p < 0.01).
  • Overall: regions with heavier AI adoption present both greater displacement pressure and stronger adaptive responses among some groups.

• AI-related measures and interactions

  • AI awareness and usage frequency were included as explanatory variables; interaction terms tested moderation of gender and education effects by region.
  • Evidence indicates AI adoption amplifies pressures on low-educated middle-aged workers; more-educated and technical workers feel relatively better positioned to adapt.

• Recommendations (from authors)

  • Targeted vocational/digital training tailored to middle-aged workers.
  • Age-sensitive, flexible work arrangements and stronger organizational support.
  • Psychological counseling, career guidance, and social-welfare measures to reduce retraining burdens and encourage employer-provided transition supports.

Data & Methods

• Design: Cross-sectional, site-based field survey (community service centers and senior activity venues) with trained staff supervising questionnaire completion.
• Locations: Beijing, Guangzhou, Lanzhou — chosen for geographic and economic variation (north, south, west).
• Participants: 1,000 approached; 889 valid responses after exclusions (notably exclusion of some female respondents affected by flexible retirement mechanisms).
• Instrument: Pilot-tested structured questionnaire covering demographics, employment history, retirement attitudes, AI awareness/use, occupation type, family economic pressure, and health controls. Reliability checks (e.g., Cronbach’s alpha where applicable) were performed.
• Analysis: Descriptive statistics, correlations, and multivariate regression models estimated in Stata. Main covariates: gender, age, education, region, occupation type, family economic pressure, AI awareness/use, health status. Interaction terms analyzed regional moderation of gender/education effects.
• Ethics: Written informed consent, anonymized data, voluntary participation.
• Limitations (noted or inferable)
  • Cross-sectional design — no causal inference.
  • Non-random, site-based sampling across only three cities limits national representativeness.
  • Self-reported measures may introduce response bias.
  • Exclusion criteria for some female respondents could bias gender comparisons.
  • Possible omitted variables (e.g., firm-level policies, detailed income measures).

Implications for AI Economics

• Heterogeneous labor-supply response to AI: AI adoption affects different worker subgroups unevenly — more-educated and technical middle-aged workers are more likely to remain active, while low-educated workers face higher displacement risk and lower reemployment willingness. Models of AI-driven labor-market adjustment should incorporate heterogeneity by education, occupation, age, and region.
• Retraining investment targeting: Public and private retraining investments should be targeted to middle-aged, lower-educated workers in high-AI-adoption regions to avoid widening inequality and to preserve labor force participation under delayed-retirement regimes. Cost–benefit analyses of retraining programs should account for differential take-up and effectiveness across age cohorts.
• Interaction of retirement policy and technology: Delayed-retirement policies may not increase labor supply uniformly when technological change increases skill demands; policy design should account for complementarities (training, flexible work) and potential disincentives for strained households. Macro labor-supply forecasts under delayed retirement should include mechanisms for skill obsolescence and retraining constraints.
• Regional policy heterogeneity: Regional heterogeneity in AI adoption implies that national-level policy tools (e.g., subsidies, tax incentives for retraining) may need regional targeting. Economic geography of AI adoption will matter for local employment dynamics and welfare outcomes.
• Employer incentives and firm-level responses: Findings suggest employers in AI-intensive regions could play a large role (provision of on-the-job retraining, phased retirement, job redesign). Incentivizing firms to invest in age-inclusive upskilling may be efficient compared with ex-post welfare transfers.
• Welfare and program design: Household economic pressure reduces willingness to work, contrary to a simple income-compensation hypothesis. Safety-net design and short-term income support could increase middle-aged workers’ capacity to undertake retraining and remain employed.
• Research directions: Need for longitudinal and quasi-experimental studies to estimate causal effects of AI adoption and retraining programs; microdata linking worker outcomes to firm-level AI investments; modeling long-term fiscal impacts of delayed retirement combined with differential labor-market reallocation driven by AI.

Potential next steps for researchers or policymakers: run randomized or quasi-experimental upskilling interventions targeted at lower-educated middle-aged workers in high-AI regions; collect longitudinal panel data to track employment transitions and earnings; and estimate welfare-maximizing mixes of retraining subsidies, employer incentives, and retirement-path flexibility.