When economists and tech commentators discuss how white-collar work survives automation, the argument usually retreats to two fallback positions. The first is the supply-side taste moat: machines can generate code, drafts, and layout, but humans maintain an informational monopoly on taste, understanding what other humans want and steering the agent toward productive ends. The second is the demand-side artisanal premium: even if an algorithm produces indistinguishable output, consumers will pay extra for goods carrying human provenance, much like hand-thrown ceramics or single-origin coffee.

Both claims sound plausible because they appeal to human dignity and standard product differentiation. Yet neither thesis had faced an audited, incentive-compatible market test where cash actually changed hands based on consumer willingness to pay.

In a new working paper, John Conlon of Ohio State and Peter Schwardmann of Carnegie Mellon built that clearing mechanism. Across three preregistered studies involving 1,025 participants, they constructed an experimental market around the production and pricing of more than 1,600 playable computer games. In the baseline setup, human producers improved rudimentary arcade games by collaborating with a frontier coding agent across three rounds of prompt revisions and real-time gameplay testing. Against them, the researchers ran a deliberately handicapped autonomous AI: an LLM prompted a single time to make the game creative and engaging, with no test runs, no error correction, and no iterative loops.

The clearing prices settled immediately. Pooling across disclosure groups, consumers paid an average of $0.47 to play the AI-designed games against $0.37 for the human-designed versions. That ten-cent spread represents a 0.31 standard deviation advantage for the unassisted model. When the authors recruited professional creatives (game developers, graphic designers, and commercial digital producers), the gap narrowed slightly, but the AI still cleared at a statistically significant 0.16 standard deviation premium over professional human output. Consumers rated the fully autonomous games higher on fun, creativity, and polish.

The market participants did not see the discount coming. Before production began, amateur participants expected their games to outsell the autonomous AI by nine cents. Professional creators were even more confident, predicting a seventeen-cent premium for their human direction. They only marked down their expectations after reading the text of the AI’s initial design prompt. That spread reveals significant miscalibration. Creative professionals priced their labor as an irreplaceable curation layer, treating their aesthetic judgment as an asset that commanded an automatic pricing spread over machine output.

The artisanal premium fared even worse. In the consumer phase, the authors varied whether buyers were told how each game was made. In blind conditions, players could not tell human games from machine games, assigning a 58 percent probability of human authorship to both pools. When the researchers explicitly labeled the games as human-designed, or informed players that their payments would transfer directly to the human creator as a cash bonus, willingness to pay did not budge. The joint test across disclosure treatments returned a p-value of 0.87. Given the option between human-backed provenance and better mechanics, consumers took the mechanics every time.

Secondary effects on the producers themselves point to a subtler labor dynamic. Watching the autonomous model work actually made human producers better at their tasks; subsequent games they built scored higher on fun and polish. But their subjective pride in the work fell by 0.14 standard deviations, and their willingness to publish their creations on a public registry dropped by 0.13 standard deviations. Learning that an unassisted model can outperform human direction improved output while degrading the perceived meaning of the craft.

In financial markets, holding an asset whose market value sits below its carrying cost forces an inevitable impairment charge. For creative and knowledge work, human taste has been treated as a permanent intangible asset capable of insulating labor from machine competition. Conlon and Schwardmann’s market shows that when cash is on the line, consumer demand routes directly to product utility rather than production origin. If the median buyer assigns zero dollar value to human provenance and discounts human curation against autonomous generation, the human in the loop shifts from a premium feature into an expensive operational friction.