Debates about generative AI and creative work often collapse into a single substitution story. If systems can draft, compose, and edit at near zero marginal cost, should artists be expected to face rapid earnings compression? Or is the first order effect a reorganization of production, with new complementarities, new bottlenecks, and slower moving distributional consequences?
In a new paper published in the Journal of Cultural Economics, I study the late 2022 diffusion of modern generative AI as a shock to task content asking the following: did artistic occupations that are more exposed to large language model capabilities experience different changes in earnings, employment, and hours than less exposed artistic occupations?
My analysis uses an occupation level index of “LLM exposure” from Eloundou et al. (2024) that approximates the share of tasks in an occupation that large language models can plausibly perform or materially assist. The object is not an automation rate. It is a ranking of how strongly generative AI could plausibly reshape task bundles across occupations.
The paper triangulates across datasets that cover different segments of creative labor markets.
- Establishment based wage and employment data that measure payroll jobs cleanly.
- Household survey microdata that better capture self employment and freelance earnings, and permit demographic controls.
- Panel survey evidence that measures AI use directly, links usage to worker outcomes, and records how AI is being used on the job.
My empirical strategy leverages a difference in differences framework interacting post period indicators with occupational exposure. The identifying assumption is that absent generative AI diffusion, higher exposure and lower exposure artistic occupations would have moved similarly.
First, there are no clear short run earnings collapse, through the early post period. Across the wage evidence, more exposed artistic occupations do not exhibit a sharp relative earnings penalty in the immediate post shock window. Point estimates are generally near zero and in some specifications modestly positive. In this sense, early adjustment does not look like a sudden broad devaluation of artistic labor in pay data, at least within the observable horizon.
Second, employment is a more ambiguous margin and plausibly adjusts earlier than wages. These results are less stable across specifications. Some estimates suggest weaker employment growth in more exposed artistic occupations after the shock, but uncertainty is material and the pattern is not uniform. Crucially, this is the margin where early disruption would plausibly appear, since staffing, contracting, and hiring can adjust before wage schedules reflect a new equilibrium, so the absence of a strong negative effect is an important result on its own.
Third, hours respond more clearly than earnings in the microdata. A recurring pattern is that exposure is more strongly associated with changes in hours than with changes in earnings. One interpretation is work intensification and workflow change: more iteration, more coordination, and more experimentation, without immediate translation into higher average pay. Another is a lag between productivity possibilities and the institutional and market structure conditions that determine who captures returns, often referred to as the productivity J-curve.
The panel evidence helps distinguish generalized diffusion from artist specific production changes. First, generative AI is used disproportionately for upstream creative support Reported usage among artists concentrates in ideation, drafting, learning, collaboration, and automation of basic tasks. That usage profile is more consistent with acceleration of early stage work and iteration than with direct replacement of the final creative product. Second, worker experience is mixed and sensitive to definition of “artist”. In pooled cross sections, frequent AI use is associated with higher job satisfaction and higher burnout, a pattern often consistent with selection into more demanding roles or workplaces. When controlling for stable individual differences using fixed effects, much of the baseline relationship attenuates, suggesting that worker and job composition explains a sizable share of the raw association.
The results are more consistent with task reallocation than immediate substitution. Early impacts appear to operate through changing what creative workers do, how much they do, and how work is organized, rather than through a rapid wage shock. There is also the possibility that early adoption can raise measured activity and labor input, including hours, as production reorganizes and experimentation intensifies. Distributional consequences may arrive later and depend heavily on market structure and institutions: platform governance, licensing regimes, attribution norms, and the credibility of provenance. Capability does not mechanically map into creator rents.
The near term evidence does not support a general earnings collapse narrative for artists. That should not be read as reassurance that displacement risks are absent. It suggests that near term impacts may be concentrated in employment dynamics, contracting, and bargaining rather than immediate wage erosion. Moreover, adjustment is already visible as reorganization of creative work. That is consistent with creative production becoming more iterative and output dense, with the possibility of rising time requirements and coordination demands before pay adjusts.
The medium run distribution of gains is likely to hinge on governance and rights infrastructure. If audiences value human made work, credible provenance and enforceable signaling mechanisms become economically important, not merely cultural preferences.
References
Eloundou, T., Manning, S., Mishkin, P., and Rock, D. (2024). Gpts are gpts: Labor market impact potential of llms. Science, 384(6702):1306–1308.
About the article
Makridis, C.A. The labor market effect of generative artificial intelligence on artists. J Cult Econ (2026). https://doi.org/10.1007/s10824-026-09575-3
About the author
Christos A. Makridis is an Associate Research Professor at the W. P. Carey School of Business
About the image
Photo de adrianna geosur Unsplash
