The more people understand how AI image generators work, the less they think it is ethical to sell or exhibit AI-made art, according to a study published in the journal Cognition.
The research, led by Ionela Bara, a postdoctoral researcher at ETH Zurich’s Social Brain Sciences lab, drew on three experiments with roughly 100 participants each. In the first, Bara showed participants 40 images — 20 landscapes and 20 portraits — generated by DALL-E 3, using prompts built around the style of Impressionist painter Joaquín Sorolla.
Half the participants saw the images with no explanation. The other half first received a detailed description of how AI image generation actually works: training on large datasets, the role of text prompts, the absence of a human hand at the canvas.
Those who got the explanation rated AI art as significantly less morally acceptable — particularly when it came to making money from the work, winning awards, or calling it conventional art. The art itself still looked good to them. Aesthetic appreciation did not shift. The problem, in their view, was not the image. It was what the artist stood to gain from it.
Prestige Fails to Shift Moral Judgments
The second experiment tested whether success could soften those judgments. Bara told participants — again, those who had learned how AI systems work — that a given AI-generated piece had been exhibited, sold, or publicly praised. None of it helped. Prestige did not move the moral needle.
The third experiment got at something more fundamental. Using a go/no-go association task, Bara tested whether people hold automatic, gut-level biases toward seeing AI art as good or bad. The answer was no. Participants who had not been educated about AI made no strong instinctive distinction between AI-generated and human-created Impressionist paintings. The moral resistance, the study concluded, is not wired in. It develops with knowledge.
A Market Under Fire
That finding lands differently in the current climate. Christie’s held an auction in New York selling AI-generated works — surreal portraits, photorealistic images, cartoon-inspired pieces. More than 6,000 artists protested, arguing that the AI models behind those images were trained on copyrighted work without creator consent. Christie’s said the pieces demonstrated “human agency in the age of AI.” The two sides were not debating aesthetics. They were debating ethics — and according to Bara’s research, anyone who understood the technology well enough was likely to land on the artists’ side.
The legal landscape has shifted, too. On March 3, the US Supreme Court declined to hear a challenge to the ruling that AI-generated artwork cannot hold copyright protection. That same week, a UNESCO report found that artists broadly are facing income declines and heightened exposure to IP violations as AI-generated content spreads. A survey of UK novelists found more than half believe generative AI could end their career.
Bara recommended that artists who use AI tools disclose what model they used, what data it was trained on, how they wrote their prompts, and where they made creative decisions along the way. She suggested transparency might invite criticism but could also build credibility — and give audiences, curators, and policymakers the tools to think critically about the technology rather than react to it.
The study did not measure what happens when people learn about AI art over months or years, only the immediate effect of a single explanatory paragraph. Whether repeated exposure deepens the moral objection or dulls it remains, for now, an open question.