Does a strong track record protect freelancers from AI?

The paper asks whether a freelancer's standing protects them from the release of ChatGPT, and the answer it reports is no. The authors "do not find evidence that high-quality service, measured by their past performance and employment, moderates the adverse effects on employment." They add that "we find suggestive evidence that top freelancers are disproportionately affected by AI." The conclusion repeats the point: offering high-quality service "does not mitigate the negative effect of AI on freelancers." The claim is that in this market a track record does not shield freelancers, and that the best-established ones may be hit harder.

The discussion reads this through the skill-biased technological change model of Card and DiNardo (2002). If the labor supply on Upwork "does not change dramatically over the short run," then a relative effect on the two worker types "would translate to changes in relative marginal product" after ChatGPT. On that reading, high-quality workers being hurt disproportionately is a shift in their relative productivity. The authors present this only as an interpretation of "Tables 3 and A4" as "suggestive evidence," and the excerpt does not reproduce the footnote behind the supply assumption.

The paper sets this against experimental work it cites: Noy and Zhang (2023), Brynjolfsson et al. (2023) and Peng et al. (2023) find that LLM adoption "differentially benefits low-ability workers relative to high-ability ones." The Upwork pattern runs the other way on employment, since the stronger freelancers are the ones the authors suggest lose more. The nearest library parallel is Can AI narrow the education performance gap?, where AI narrows a gap in an experiment and lower-education users keep part of the gain without it. That result is about performance and this one is about employment, so they are a contrast rather than a refutation. The moderator tested here is also narrower than the claim in What makes accountable judgment scarce when AI cognition is cheap?: the excerpt tests past performance and employment history, not judgment, and finds no buffer. The sibling note from this source gives the aggregate effect that this heterogeneity result sits beside.

What the excerpt does not establish is the size of the heterogeneity effect. The top-freelancer result is labeled "suggestive," and the excerpt gives no sample size, no definition of "top," no estimates or standard errors for Tables 3 and A4, and no evidence on whether the pattern holds beyond the short run. A null moderator is a failure to find moderation, not proof that quality is irrelevant. The implication is that the non-buffering finding is the paper's firmer heterogeneity claim, while the disproportionate-harm reading should be held as a hypothesis the authors themselves call suggestive.