What makes AI-involved art worth buying?

What gives an artwork value: how it looks, the work behind it, or the people whose contributions made it possible? AI image generators bring these questions together because a finished image does not necessarily reveal the prompting, selection or other artistic work involved.

An arXiv preprint investigates these judgments through a competition culminating in a physical gallery exhibition. Its authors, in order, are Wisdom Obinna, Hexuan Wang, Gabriel Lima, Cecile Duncan, Kristelia García, Nina Grgić-Hlača, Toni-Lee Sangastiano and Elissa M. Redmiles. Their findings concern hypothetical prices and opinions about fair rewards, not actual sales or legal entitlements.

A competition with an exhibition at stake

The study included 105 creators and 104 peer evaluators recruited from one university community. Most participants were undergraduates; some reported training in art, computer science or law. Five additional expert jurors helped select the exhibition.

Creators entered either an AI-only category or a mixed-media category combining AI with traditional or digital artistic techniques. A custom platform recorded prompts, generated images and session time. Creators also described their process and assessed their own creative input.

Peer evaluators judged the finished images without being told their category or shown process records. Each assessed about six works, rating creativity, technical quality, visual appeal and perceived effort. Both groups answered questions about prices, rewards and control over the images.

Twelve selected works were displayed in a university gallery and online. Participation payments and exhibition prizes gave the competition real consequences, but buying and selling questions remained hypothetical.

Imagined prices followed different signals

Nearly half of evaluators said they would buy at least one artwork they assessed: 36% indicated a price of $100 or less, and 10% indicated more than $100. Among those willing to buy, higher price responses were associated with greater perceived creativity, visual appeal and effort.

For creators, asking prices were positively associated with the total number of words submitted across their prompts. Their creativity and technical-quality assessments were not significantly associated with asking prices in the reported model. Creators with artistic training also reported lower asking prices.

Prompt word count was the researchers’ main behavioral proxy for effort because it correlated with prompt numbers and session duration. It is not a direct measure of artistic skill or all creative labor, particularly work outside the platform, and may partly reflect verbosity.

Importantly, this recorded measure did not significantly correlate with evaluators’ estimates of effort, their artistic ratings or their willingness to pay. What viewers imagined about the process therefore did not reliably track this particular record of it. That does not establish that viewers misunderstood every form of creative contribution.

Who deserves the reward?

Creators overwhelmingly included themselves among deserving prize recipients, and roughly 63% selected themselves alone. Most evaluators also supported rewarding the prompting user. Only a minority in either group supported rewarding artists whose work had been used to train the models.

Disciplinary background was associated with some choices. Evaluators with legal training were more likely to support rewards and rights for training-data artists. Those with computer science training were less likely to support monetary rewards for prompting users and reported lower willingness to pay.

These associations do not show that education caused the differences. Training was self-reported, participants chose their roles, and legal training did not necessarily mean copyright expertise. Some legal-background estimates also had very wide uncertainty intervals.

What the exhibition cannot establish

This was an exploratory, observational study, not an experiment manipulating process information. It included no purely human-made comparison group, so it cannot establish whether AI involvement itself increased or reduced value.

Creator questions emphasized creativity and skills used during production, whereas evaluator questions emphasized qualities of the finished image. That difference prevents straightforward statistical comparisons between their ratings. The single-community sample and hypothetical pricing further limit conclusions about professional artists or real markets.

The authors propose documenting iteration histories, curation and human guidance to make creative contributions more visible. They did not test whether such documentation changes prices or rights judgments. The exhibition instead identifies a question for future experiments: would showing the process change what viewers value, and whom they believe should benefit?