Did ChatGPT's release reduce freelance writing work and pay?

The paper's central claim is that freelancers in the occupations most exposed to generative AI lost ground after ChatGPT's release, on one online labor market. The conclusion gives the platform-level figures directly: the release "leads to a 2% drop in the number of jobs on the platform, and a 5.2% drop in monthly earnings" on Upwork, which matches employers with freelancers for small and medium tasks. The abstract states the pattern more broadly: freelancers in highly affected occupations "suffer from the introduction of generative AI, experiencing reductions in both employment and earnings." The authors also report "similar effects" when they study the release of other image-based generative models.

The identification is a difference-in-differences design over the full employment histories of freelancers on the platform, comparing how employment changed in affected occupations against the rest around the release. The excerpt's reason for focusing on writing-related work is that the model's "current rather than future capabilities" bite there first, and that the tasks are easy to hand over: a client "can easily copy and paste paragraphs into the GPT prompt and immediately evaluate its output." From this the authors conclude that "in the short term generative AI reduces overall demand for knowledge workers of all types." That conclusion is wider than the data, which track freelancers on one platform and do not measure overall demand directly.

Against the wider literature, the result lands on the substitution side of a question the excerpt calls "theoretically ambiguous": AI can substitute for labor or complement it by raising productivity. The field experiments the paper cites measure performance under access to AI tools, so they are a different outcome from the employment and earnings measured here, and the excerpt does not present the two as in conflict. The closest neighbor in the library, Does AI turn freelance work into validation instead of creation?, is a position argument about how freelancers build skill. This excerpt measures jobs and earnings for the same kind of worker, which is a different and observed quantity. The scenario model in Does AI growth inevitably shift wealth away from workers? projects a long-run decline for knowledge workers; this paper offers a measured short-run earnings drop of the sort that projection describes. The education-gap experiment in Can AI narrow the education performance gap? measures performance, so it neither confirms nor undercuts these figures.

What the excerpt does not establish is the long run. The authors say "the long-term implications may be significantly different," and they allow that the negative effects could either grow as the technology spreads or shrink as AI "better complement the performance of human workers." The excerpt also omits the sample size, any measure of occupational exposure beyond the writing-related focus, and the standard errors around the 2% and 5.2% figures. Its introduction breaks off during the description of the design. It does not reach the welfare question, which the authors place "beyond the scope of this paper." The implication is narrow: these are short-run estimates for online freelancers in affected occupations. They support a claim about that market, not about aggregate unemployment or wages.