Generated by Codex with GPT 5.6 Sol XHigh

Knowing How the Image Was Made

People often discuss AI art as though dislike of it were mainly aesthetic: perhaps viewers find it bland, derivative or emotionally hollow. “The Ethics of AI Art,” by neuroaesthetics researcher Ionela Bara, points to a different source of resistance. In her experiments, learning more about how an image generator works changed people’s moral judgments without changing how much they liked the resulting pictures.

The distinction matters because AI art arrives with two stories attached. One is the visible image. The other is the largely hidden process behind it: models trained on enormous collections of pictures and text, prompts that may name a living or historical artist, and businesses that can turn the output into money or prestige. A viewer can enjoy the first while becoming uneasy about the second.

Bara begins with a 2025 Christie’s auction in New York, where AI-generated works sold for more than \$700,000. The commercial success drew protests from more than 6,000 artists, who argued that the underlying models had been trained on copyrighted work without consent. Similar worries have spread through writing and other creative fields. The conflict is therefore not simply over whether machines can produce attractive objects. It is over who supplied the creative labor, who receives credit and who is allowed to profit.

Information Changes the Moral Frame

Bara and her colleagues ran three experiments with 100 participants apiece. In the first, people evaluated 20 landscapes and 20 portraits generated with DALL-E 3 from detailed prompts based on the work of Spanish Impressionist Joaquín Sorolla. Half saw the images without additional context. The others read a short explanation of how a text-to-image system is trained on labeled art and then generates images from prompts that can include an artist’s name and style.

Participants who received that explanation judged uses of the images as less morally acceptable, particularly when the imagined use involved earning money, gaining prestige or presenting the output as conventional art. Yet the explanation did not reduce the images’ aesthetic appeal. What people learned altered their view of the transaction, not necessarily their experience of the picture.

A second experiment tested whether conventional signals of legitimacy could counter that discomfort. Participants were told that some works had been exhibited, praised or sold. Such cues often produce an authority effect: people infer value from expert approval or public success. Here, however, acclaim did not restore the moral acceptability of AI art for participants who understood more about its production. Institutional validation could not erase concern about the process.

The final experiment looked for a faster, more automatic bias. Participants rapidly paired AI- or human-made Impressionist images with labels such as “good” and “bad” in a go/no-go association task. The researchers found no strong reflexive tendency to classify either source as inherently better or worse. Bara interprets that result alongside the earlier experiments: moral resistance to AI art appears to be shaped by information and reflection rather than a settled gut reaction.

Transparency Is Useful, but Not Neutral

The experiments support a focused conclusion. A label saying “AI-generated” does not fully explain how people judge an image. Their response depends on what they understand about training data, prompting, artistic imitation, ownership and reward. Greater technical literacy can therefore make an audience more critical even when it leaves aesthetic enjoyment intact.

That creates a real tension for artists and institutions using generative tools. Disclosing the model, data sources, prompts and human contributions may invite sharper scrutiny, but it can also establish credibility and give audiences enough information to make a considered judgment. Opacity may avoid immediate criticism while leaving the underlying questions unresolved.

The findings are suggestive rather than universal. Each experiment involved 100 people, and the article provides little detail about participant demographics, effect sizes or whether reactions would differ across artistic styles, models or cultures. The studies measured judgments about moral acceptability; they did not settle the legal or philosophical question of whether a particular training practice is ethical. Nor does the absence of a strong automatic bias mean that people begin from a perfectly neutral position.

The durable insight is that aesthetic value and moral acceptance are separate judgments. A person can find an image beautiful while objecting to the system that produced it. As generative tools become ordinary, the ethics of AI art will depend not only on what appears on the canvas but also on whether the creative process is visible enough for viewers to decide what they are endorsing.