AI tools narrow the wage gap between junior and senior freelancers in coding, technical writing and design, compressing entry-level skill premiums and boosting the importance of supervision; simulations imply firms will need to redesign promotion paths as human intuition becomes more substitutable with algorithmic output.

Main Finding

Generative AI (LLMs and multimodal systems) has an asymmetric effect on high-skilled freelance cognitive labor: it compresses wages and lowers returns for entry-level (novice) tasks while increasing returns for expert practitioners. Quasi-experimental DiD estimates on major freelance platforms show novice hourly wages fell ≈20%, mid-level wages fell ≈6.7%, and expert wages rose ≈25% after widespread GenAI adoption — consistent with AI substituting for low-skill cognitive tasks (high elasticity of substitution) and complementing high-skill tasks (low elasticity).

Key Points

• Context: Unlike prior automation waves (routine/manual), generative AI targets non-routine cognitive tasks (coding, technical writing, graphic design), changing the distribution of demand across skill levels.
• Data sources: Large international freelance platforms (Upwork, Fiverr, Freelancer.com), Jan 2021–Dec 2024, pooled at the worker-task-month level; sample size ~10k–25k contracts.
• Identification: Difference-in-differences (DiD) exploiting temporal cross-sectional variation in platform/market exposure to GenAI; treatment = AI-exposed tasks × post-adoption period.
• Main quantitative results:
  • Novice: −20% in hourly wage (post-AI)
  • Mid-level: −6.7%
  • Expert: +25%
• Interpretation: “U-shaped” or polarized effect — AI substitutes for commoditized cognitive work (novices) and complements strategic/high-level work (experts).
• Elasticity of substitution: Estimated/simulated framework indicates σ > 1 for novice tasks (easy substitution by AI) and σ < 1 for expert tasks (complementarity).
• Robustness: Placebo (pre-2021) tests, alternative specifications, and platform subgroup analyses produced consistent patterns.
• Limitations acknowledged by the author: imperfect measurement of individual AI tool use and intensity, unobserved heterogeneity, platform selection biases, and limits to generalizing beyond freelance/online markets.

Data & Methods

• Data:
  • Platforms: Upwork, Fiverr, Freelancer.com.
  • Period: January 2021 – December 2024 (pre- and post-widespread GenAI adoption in late 2022).
  • Unit: individual job contracts aggregated to worker-task-month.
  • Sample: stratified sample of high-skill cognitive contracts (software development, technical writing, graphic design).
  • Skill stratification: novice / mid-level / expert based on years on platform, ratings, projects, income.
• Outcome: hourly earnings or normalized contract value per hour.
• Regressors & controls:
  • Main: AI exposure indicator (task-level susceptibility to GenAI), post-period dummy, interaction (treatment).
  • Controls: worker experience, education proxies, task complexity, geography, client budget, platform/client fixed effects.
• Identification strategy:
  • Difference-in-differences to estimate average treatment effect of GenAI adoption on wages across skill groups.
  • Elasticity of substitution modeled/simulated in a two-input production-like specification (human effort × AI output) to infer substitutability vs complementarity by skill level.
• Inference & validation:
  • Statistical tests at p < 0.05, confidence intervals reported.
  • Robustness: placebo pre-periods, alternative model specs, subgroup/platform analyses.

Implications for AI Economics

• Labor-market structure:
  • Short-to-medium term: wage compression at the bottom of the cognitive-skill distribution and premium growth at the top → increased within-occupation polarization.
  • Entry-level roles that previously served as on-ramp training/apprenticeship may shrink in availability and pay, potentially disrupting career ladders and human capital accumulation.
• Firm & platform strategy:
  • Organizations should redesign promotion/training pipelines (fewer pure-execution junior roles; more supervised AI-integration and human-in-the-loop oversight roles).
  • Value shifts toward supervision, integration, strategy, and interdisciplinary synthesis where humans complement AI.
  • Platforms may see intensified competition on commoditized tasks and concentration of high-value work among top-rated experts.
• Policy:
  • Need for targeted retraining/reskilling focused on AI supervision, prompt engineering, strategic synthesis, and cross-disciplinary coordination.
  • Consider safety nets or transition policies for displaced entry-level cognitive workers; monitor labor market churn and earnings inequality.
  • Regulate/encourage transparent disclosure of AI use on platforms (to preserve hiring signals) and guard against monopsony/market-power effects that amplify wage pressure.
• Measurement & research agenda:
  • Importance of task-level measures of AI exposure and intensity (beyond binary indicators).
  • Extend analysis to employer-side/firm-level data, full-time employment markets, cross-country heterogeneity, and long-run labor supply responses.
  • Track whether reduced entry-level opportunities translate into long-term mobility losses or whether new forms of human capital emerge around AI coordination.
• Economic theory:
  • Empirical support for heterogeneous elasticity of substitution across tasks/skill levels — models of technological change should allow task- and skill-specific substitution/complementarity rather than uniform SBTC/RBTC assumptions.

Caveat: results are derived from online freelance platforms and a specific post-2022 adoption window; generalization to salaried employment, other countries, and longer horizons requires further study.