Generative-AI skills raise IT application rates but widen the gender gap: on Upwork men — especially those with high GenAI proficiency — increase applications more than women, suggesting GenAI may exacerbate rather than equalize entry into tech roles.

Main Finding

Preliminary field-experimental evidence from Upwork indicates that possessing GenAI skills increases IT job application rates for both men and women, but the increase is substantially larger for men—especially those with high GenAI proficiency. In short: GenAI skills appear to widen, not close, the gender gap in IT job applications in this freelance setting.

Key Points

• Context: GenAI might either democratize access to technical work (by lowering coding barriers) or create new, complementary skill demands (prompt engineering, evaluation, human-AI collaboration) that perpetuate existing advantages for men.
• Experiment overview:
  • Stage 1: Measured existing GenAI proficiency among freelancers.
  • Stage 2: Randomly invited those freelancers to apply to either an IT job (data analyst) or a comparable non-IT job (market analyst).
• Sample and response:
  • 400 users sampled; 372 invitations sent; 75 respondents; 73 passed attention checks (Stage 1).
  • 72 invitations sent in Stage 2 (one account deactivated).
• GenAI proficiency measured on three dimensions:
• Subscription-based usage (paid GenAI subscriptions).
• Experience-based proficiency (strategies and success improving GenAI outputs).
• Knowledge-based understanding (tokens, prompt engineering, hallucination).
• Manipulation checks and pretests:
  • Pretests for survey clarity (n=50) and job-type classification (n=120).
  • Job-type classification: IT job mean = 6.80 vs. non-IT = 2.26 on a 1–7 IT-relatedness scale (p < 0.001).
• Outcome:
  • Dependent variable: binary Apply indicator (submitted application within 7 days).
  • Finding: GenAI skills raised the probability of applying to IT jobs for both genders, but the magnitude was larger for men; high GenAI proficiency magnified this gender gap.
• Design strengths: field setting with real economic incentives, private invitations to mitigate algorithmic selection, real application behavior rather than stated intentions.
• Limitations noted by authors (and evident from the design): small sample size, natural (non-manipulated) variation in GenAI skill, possible selection bias (survey responders), and limited external validity beyond freelance marketplaces.

Data & Methods

• Platform: Upwork (large freelance marketplace, identity verification present).
• Two-stage field experiment:
  • Stage 1 (measurement): Private survey invitation framed as labor market research; GenAI items embedded to avoid demand effects. Attention checks used; deceptive framing for experimental integrity (purpose disclosed after acceptance).
  • Stage 2 (outcome): Participants randomized to receive a private invitation to either an IT job (data analyst) or a matched non-IT job (market analyst). Participants saw only their assigned job.
• Sample sizes:
  • Invitations sent Stage 1: 372; respondents: 75; valid responses: 73.
  • Stage 2 invitations successfully sent: 72.
• GenAI proficiency variables: subscription status, self-reported experience/proficiency in improving GenAI outputs, and knowledge of GenAI concepts.
• Dependent variable: Apply (1 if applied within 7 days, 0 otherwise). Non-response and explicit rejections treated as non-application.
• Analytic approach: Compare application rates across job type, gender, and measured GenAI proficiency; assess interaction effects (gender × GenAI skill). (Detailed statistical specifications and estimates not included in excerpt; main inference drawn as described in abstract.)

Implications for AI Economics

• Policy and workforce development:
  • GenAI is not a guaranteed equalizer; interventions should actively target gendered disparities in GenAI adoption, confidence, and advanced complementary skills (prompt engineering, evaluation).
  • Training programs should prioritize not only basic GenAI access but also advanced, confidence-building, hands-on opportunities that counteract identity/stereotype effects.
  • Employers and platforms should monitor whether GenAI-related hiring criteria amplify preexisting demographic advantages and consider countermeasures (e.g., blind screening, standardized task-based assessments).
• Research implications:
  • Need for larger, more diverse samples and experimental manipulations (e.g., randomized GenAI training or tool access) to establish causality for mechanisms (confidence, skill complementarity, identity).
  • Study both demand-side (employer selection, platform algorithms) and supply-side (applicant behavior) channels to understand aggregate labor-market impacts of GenAI.
  • Examine longer-term outcomes (hiring, earnings, career progression) to see if initial application differences translate into sustained labor-market inequality.
• Broader economic considerations:
  • Technological advances can reallocate comparative advantages; without targeted policy, they may exacerbate existing inequality rather than reduce it.
  • Models of labor market equilibrium in the GenAI era should incorporate heterogeneous adoption, complementarities between traditional technical skills and GenAI-savvy skills, and gender-differentiated behavioral responses to new tools.

Suggested next research steps (based on this study): - Run larger-scale field experiments that randomly assign GenAI training or tooling to isolate causal pathways. - Collect richer mediators (confidence, identity, time spent using GenAI on tasks) and follow-up on hiring outcomes. - Test interventions (mentorship, confidence-building, role-model exposure) designed to mitigate the widening effect observed here.