After ChatGPT's release, clients on Upwork put less weight on workers' human-capital signals and more weight on price in AI-exposed job categories, with demand shifting toward lower-priced freelancers—evidence consistent with a commoditization effect of generative AI.

Main Finding

Generative AI (proxied by the release of ChatGPT) reduced the market importance of observable human-capital signals and increased the importance of price in an online labor market. In more AI-exposed job categories on Upwork, human-capital information became less predictive of hires while price became more predictive — consistent with an AI-driven commoditization of labor. The strongest effects appear to grow over time through 2026Q1.

Key Points

• Data: Balanced worker-quarter panel of 49,610 Upwork workers (2021Q1–2026Q1), covering 2.26 million contracts.
• Outcome: Quarterly demand = log(1 + number of contracts).
• Treatment: Continuous occupational AI exposure from Eloundou et al. (2024, “E1” scores), mapped to Upwork subcategories and rescaled so 1 = most exposed subcategory in sample.
• Representation: Worker profiles encoded via high-dimensional text embeddings from a pre-trained language model (avoids hand-coding features).
• Predictive importance: For each worker, the authors predict demand from profile information and price, and compute Shapley-based importance (Shapley regression values) for blocks of signals (self-presentation, credentials, reputation) and price.
• Identification: Difference-in-differences around the release of ChatGPT (Nov 2022) with exposure as the continuous treatment.
• Core empirical findings:
  • In the most AI-exposed job categories, the combined importance of human-capital signals falls by ≈7.8% (relative to an unexposed category) after ChatGPT; price importance rises by ≈1.1%.
  • Effects increase over time and are largest near the end of the sample (2026Q1).
  • The demand premium for workers with strong human-capital signals declines more in high-AI-exposure categories.
  • Demand reallocates toward lower-priced workers more strongly in AI-exposed categories.
• Interpretation: Results are consistent with labor commoditization via AI (AI standardizes outputs, compresses differences across workers, makes clients more price-sensitive). The authors acknowledge alternative mechanisms (e.g., changes in signaling or screening) but present two supporting tests (declining premium for high-human-capital workers; shift toward lower-priced hires) that bolster the commoditization interpretation.

Data & Methods

• Source: Upwork worker profiles (March 2026 snapshot) and contract histories.
• Sample construction:
  • Start: 201,857 searchable profiles → restrict to workers with ≥1 contract opened in 2022 → require a recorded primary subcategory → final N = 49,610 workers.
  • Panel: 21 quarters (2021Q1–2026Q1), 1,041,810 worker-quarter observations.
• Profile variables organized into four blocks:
  • Self-presentation: title, description, skill tags, lengths, country, availability.
  • Credentials: education, institutions, employment history, portfolio items.
  • Reputation: prior contract titles, written feedback, badges, prior job counts and scores.
  • Price: posted hourly rate (and log).
• Text handling: All profile text encoded with dense embeddings from a pre-trained language model to capture high-dimensional human-capital signals without manual feature engineering.
• Predictive model and importance metrics:
  • Predict worker-quarter demand from profile embeddings/variables and price.
  • Use Shapley regression values to quantify each block’s contribution to the predicted demand for each worker.
• Empirical strategy:
  • Difference-in-differences (pre/post ChatGPT) using continuous AI-exposure by subcategory as treatment intensity.
  • Main comparisons assess how post-shock changes in Shapley importance (and other outcomes like demand premia and reallocation by price) vary with exposure.
• Robustness / caveats highlighted by authors:
  • Profile data are observed as a single March 2026 snapshot; many credential fields are persistent but some self-presentation fields may have changed over time.
  • Do not observe job postings, proposals, or cover letters; cannot observe ordering in search results or recommendation algorithms.
  • AI exposure assigned from March 2026 subcategory mapping (LLM-assisted mapping of platform subcategories to occupational exposure scores); potential measurement error if workers switched subcategories in response to AI.
  • Posted hourly rate observed in 2026; authors argue cross-sectional price differences dominate within-worker time variation.

Implications for AI Economics

• Evidence for commoditization: The paper provides empirical support that generative AI can alter the weights markets place on human capital signals, making workers across skill levels more substitutable in AI-exposed tasks.
• Labor market effects:
  • Reduced premium to observable human-capital attributes may weaken returns to investments in those attributes for workers in exposed occupations.
  • Increased price sensitivity and reallocation toward lower-priced workers could compress wage dispersion (and raise welfare concerns for higher-skill workers) while potentially lowering hiring costs for clients.
• Platform design and policy:
  • Marketplaces may need interventions (better signal-preservation, richer credential verification, redesigned ranking/search algorithms, or platform-provided skill certifications) to prevent a pure race-to-the-bottom on price and to maintain informative signals about genuine worker quality.
  • Policymakers and educators should consider that AI exposure can change incentives for human-capital accumulation and may increase the importance of AI-complementary skills or tasks that remain hard to automate.
• Open questions / directions:
  • Micro-level channels: measure worker-level AI adoption/use, or obtain proposal/cover-letter data to better separate screening vs. productivity-standardization mechanisms.
  • Long-run dynamics: will markets re-equilibrate (e.g., new signals emerge, platforms adapt) or will commoditization persist?
  • Welfare distribution: quantify effects on earnings, inequality, and job turnover across worker skill groups and occupations.

Overall, by combining text embeddings, predictive importance measures, and a continuous diff-in-diff around ChatGPT, the paper documents a shift away from human-capital signals and toward price in AI-exposed online work — consistent with an AI-driven commoditization of certain types of labor.