AI is widely used by disabled job seekers in China but correlates with lower reported job-search success, likely because the most disadvantaged are more likely to adopt AI; platform accessibility — not AI use per se — consistently predicts better outcomes.

Main Finding

AI tools are widely used by disabled job seekers in China (78.02%) but do not independently improve job-search outcomes. AI use is negatively associated with reported job-search success—likely reflecting a selection effect (more disadvantaged job seekers adopt AI). Platform design and accessibility (ease of use, satisfaction with accessibility) are the strongest, consistent predictors of better outcomes. The paper argues that AI alone cannot overcome structural barriers (platform inaccessibility, employer discrimination, stigma around welfare enterprises) and that preconditions—stable accommodation subsidies, enforced accessibility standards, and fairness audits of screening algorithms—are needed for AI to deliver inclusion gains.

Key Points

• Hypotheses tested
  • H1: AI adoption is negatively associated with job-search success because the most disadvantaged are more likely to use AI. (Observed negative association; interpreted as selection effect.)
  • H2: Platform accessibility (design/usability) is the main predictor of success and satisfaction, more so than AI utilization. (Supported.)
• Sample and response profile
  • Survey of 91 disabled job seekers in China (snowball sampling).
  • Disability composition: 83.52% visually impaired; small shares of ambulatory physical, multiple disabilities, hearing/intellectual impairments.
  • Age: concentrated in 18–44 (≈57%); Education: 41.76% undergraduate or above; bachelor’s largest single group (34.07%).
• Key descriptive results
  • 78.02% reported prior use of AI tools (chatbots, virtual assistants).
  • Major application barriers reported: no employer response (45.05%), automated resume-screening errors (40.66%), difficulty reading/navigating sites (50.55%).
  • Only ~16% reported being satisfied or very satisfied with built-in platform accessibility features (9.89% + 6.59%).
  • High demand for AI features that identify accessible employers (78.02%), match jobs to ability (64.84%), and draft/optimize resumes (64.84%).
  • Respondents’ top AI concerns: excessive standardization (61.54%), algorithmic screening bias (51.65%), impersonal recruitment (50.55%).
  • Willingness to use a specialized AI job-assistance system: 73.63% totally willing; additional 21.98% somewhat willing.
• Regression / statistical findings (OLS; N = 91 unless noted)
  • Model 1 (job-search success): AI user status associated with a statistically significant reduction in success score (β = −0.583, SE = 0.206, t = −2.835, p = 0.006). Mann–Whitney U corroborates difference (U = 444, p = 0.006; Cohen’s d = −0.80).
  • Across three OLS models:
  • Platform ease of use is the only predictor significant in all three models.
  • Satisfaction with platform accessibility is significant in two of three models.
  • Education and age are never significant (all p > 0.10).
  • R² reported for Model 1 = 0.254 (F(6,84) = 4.77, p < 0.001).
• Meta-review context
  • Systematic meta-review of 69 audit/correspondence experiments (1972–2025) confirms substantial disabled callback penalties across countries (examples: Denmark wheelchair penalty ~54.7%; Norway ~48%; UK studies show ~15% penalty in accounting/finance applications). These establish a causal baseline for employer discrimination that AI can replicate if trained on biased data.

Data & Methods

• Mixed-methods design:
  • Quantitative: structured online survey of 91 disabled job seekers in China (snowball sampling within disabled communities), descriptive statistics, and three OLS regression models predicting (a) job-search success, (b) overall satisfaction with online job-search, and (c) satisfaction with job-match quality. Shared independent variables: AI use (binary), platform ease of use (Q18, 1–5), satisfaction with platform accessibility (Q20, reverse-coded), education, and age. Model 1 includes weekly hours searching; Model 2 adds job-match quality; Model 3 uses shared variables only. Ordinal scales treated as continuous. Two respondents missing Q26 reduced some models to N=89.
  • Qualitative/contextual: systematic meta-review of international audit and correspondence experiments on disability hiring discrimination (69 studies).
• Limitations noted by the study (and implied):
  • Non-representative sample: heavy bias toward visually impaired respondents (83.52%) and use of snowball sampling.
  • Small sample size (N = 91) limits power and generalizability.
  • Cross-sectional, self-reported data: causality cannot be established—observed negative AI association may reflect selection bias.
  • Ordinal survey scales treated as continuous in OLS (standard but imperfect assumption).
  • Employer-side algorithm behavior not directly audited—paper relies on user reports and meta-review to infer algorithmic harms.

Implications for AI Economics

• Selection effects complicate measurement of AI impacts
  • Observational associations between AI use and outcomes may reflect selection: disadvantaged workers are both more likely to adopt AI tools and more likely to have poor outcomes. Econometric identification requires experimental or quasi-experimental designs (randomized assistance, instrumental variables, or audits) to separate technology effects from user heterogeneity.
• AI can reproduce and amplify historical discrimination
  • Algorithms trained on biased hiring data risk encoding taste- and statistical-discrimination patterns (Arrow/Phelps framework). This raises the need to audit and correct screening models—especially where automated resume filtering is a common obstacle (40.66% reported).
• Platform design and accessibility are high-return intervention points
  • Platform ease of use is consistently the strongest predictor of positive outcomes—suggesting that UI/UX and accessibility regulation may be more effective (and lower-cost) than pushing AI adoption alone. From an economic-policy perspective, investment and regulation on platform accessibility can reduce transaction costs and increase labor market participation.
• Market failures and policy levers
  • Evidence of employer discrimination, small-firm noncompliance with quota/accommodation rules, and platform inaccessibility point to coordination and information failures. Policy levers include:
  • Enforced accessibility standards for job platforms (regulatory requirement, monitoring).
  • Fairness and performance audits of screening algorithms (transparency, counterfactual testing, mandated appeal/feedback loops).
  • Stable accommodation subsidies and incentives to reduce employer uncertainty and anticipated costs.
  • Public provision or certification of disability-friendly job listings to reduce search frictions.
• Design recommendations for AI interventions
  • Focus AI on assistive, personalized features that users value (accessibility-aware job matching, resume drafting tailored to disability accommodations, employer accessibility tagging) rather than generic automation that risks standardization and depersonalization.
  • Build transparency, explainability, and feedback channels into AI hiring tools to mitigate fears of opaque, biased filtering.
  • Ensure interoperability with employer HR systems for reduced information asymmetry and better matching outcomes.
• Research implications for AI economics
  • Need for causal evaluations (randomized controlled trials or field experiments) of AI-assisted job-search tools targeted to disabled populations to estimate net welfare effects.
  • Employer-side algorithm audits and correspondence experiments that include algorithmic interventions (e.g., reweighted/rescored resumes) to measure direct algorithmic effects on callbacks.
  • Cost–benefit analyses comparing investments in platform accessibility vs. AI enhancements vs. employer subsidies to guide efficient policy prioritization.

Suggested next research steps (brief) - Implement a randomized trial providing a subset of disabled job seekers with an accessibility-optimized AI assistance tool and compare application outcomes and callback rates to controls. - Conduct employer-side audits of mainstream screening algorithms to quantify bias and test mitigation (retraining, debiasing, human-in-the-loop). - Larger, nationally representative surveys stratified by disability type to assess heterogeneity of AI effects across impairment groups.