When the Machine Paints, Who Made the Art?
AI image generation has disrupted the economic and aesthetic foundations of visual art. The philosophical questions it raises have not been resolved — and probably cannot be.
In September 2022, Jason Allen's Théâtre D'Opéra Spatial — an image generated using Midjourney's AI image generation software — won first place in the digital arts category of the Colorado State Fair's fine arts competition. Allen had used the software to generate hundreds of variations, selected the most compelling, made some adjustments, and entered it under the name "Jason Allen via Midjourney." The judges didn't know the winning image was AI-generated.
The controversy that followed had two distinct components. The economic argument: AI art generators are trained on enormous datasets of human-created art, without the artists' consent or compensation, and they produce outputs that compete with those artists' work in the market. The aesthetic argument: AI-generated images are not art because they are not the product of intentional expression by a conscious agent who has something to say.
Both arguments are real. Both are also simpler than the underlying problem.
The copyright training question
The legal question about AI art — whether training on copyrighted artworks without license constitutes infringement — is being litigated in several parallel cases. Getty Images has sued Stability AI for using its photo library as training data. A class action brought by artists including Kelly McKernan, Karla Ortiz, and Sarah Andersen against Stability AI, Midjourney, and DeviantArt is proceeding in federal court.
The legal outcome of these cases will turn on whether training data use constitutes fair use under US copyright law — a four-factor balancing test that has not been applied to this fact pattern before. The preliminary indications from court rulings and scholarly analysis suggest that training on publicly available data may qualify as fair use under existing doctrine; the output generation question (whether outputs that resemble specific artists' styles constitute infringement) is more uncertain.
Metaculus forecasts a 58 percent probability that a major US federal court will issue a ruling finding that training generative AI on copyrighted material without license constitutes copyright infringement before 2028. If that ruling holds, the business model of every major AI image generation company is in jeopardy; they would face liability to enormous classes of copyright holders.
The economic disruption to human artists has already occurred regardless of how the legal question resolves. Stock photography and illustration — previously stable income sources for working visual artists — have been significantly disrupted by AI image generation. Shutterstock and Getty have integrated AI generation into their own platforms; smaller stock photography operators have experienced dramatic revenue declines. The rate at which advertising, media, and entertainment companies are using AI image generation to replace commissioned illustration is difficult to measure precisely but is clearly increasing.
The aesthetic question
The dismissal of AI art as "not real art" relies on a conception of artistic creation as the expression of a conscious, intentional agent — a view that has deep roots in the Romantic tradition and that most working artists, critics, and consumers share intuitively.
But the intuition runs into problems at the edges. Photography was dismissed as mechanical reproduction when it emerged, incapable of genuine artistic expression; this turned out to be wrong. Algorithmic composition — using chance operations or generative processes to produce musical material — has been a legitimate artistic practice since John Cage's experiments in the 1950s. Found art, readymades, and appropriation art have complicated the authorship question for the better part of a century.
What is genuinely different about AI art generators is the degree to which the creative labor is delegated to the system rather than performed by the artist. When Jason Allen generated Théâtre D'Opéra Spatial, he exercised curatorial judgment — selecting from generated options, making adjustments, deciding what to enter. But the generative work — the production of visual forms, the application of aesthetic choices — was performed by the model. How much of the credit for the aesthetic qualities of the output belongs to Allen, and how much to the system he used, and how much to the artists whose work trained the system?
These questions do not have clean answers, which is why the aesthetic debate about AI art tends to cycle rather than resolve. The relevant insight may be that authorship has always been more complicated than the Romantic model suggests — that art is always partly a product of tools, traditions, and training — and that AI art forces an unusually direct confrontation with that complication.
What professional artists are actually doing
The response of professional visual artists to AI disruption has been heterogeneous. Some have incorporated AI tools into their practice — using generation as a rapid ideation layer, a reference generator, or an element in composite workflows. Some have actively refused to use AI tools and have organized collective resistance to AI companies' use of their work as training data. Many have done both, in different contexts and for different purposes.
The artists most affected are those whose work has historically been defined by technical skill in production — the ability to render a specific style, produce a certain quality of digital illustration, or deliver reliable commercial visual outputs. These are the skills that AI generation most directly displaces. The artists least affected are those whose value is most bound up with conceptual development, client relationship, and the integration of their practice with their broader creative identity.
This is a familiar pattern in technological disruption: the parts of creative work that are most rule-based and reproducible are most susceptible to automation; the parts that are most contextual and relationship-dependent are most durable. What is less familiar is the speed of the disruption and the breadth of its application.
Kalshi was trading a contract on whether a fully AI-generated work — one produced entirely by AI with no human modification of outputs — will be awarded a major international art prize before 2028 at 34 percent. The framing of the contest will determine the outcome as much as the quality of the work; judges who have aesthetic criteria for "AI art" will evaluate it differently than those applying traditional criteria without modification.
Leila Farahani is a contributing writer at The Auguro covering culture, institutions, and the politics of representation.