Resistance May 13, 2026 · 11 min read

Is AI Art Plagiarism by Default?

The accusation is everywhere: AI art is plagiarism, the models are theft engines, and anyone who uses them is benefiting from stolen labor. The accusation is too broad to be true and too pointed to be ignored. Untangling it requires distinguishing two questions that the public conversation has been blurring together for three years.

by Airtistic.ai Editorial

Through the lens of artistcriticgallerypatroncollectorconsumer craftindustrymarket

The first question is training-stage: was the data the model learned from acquired legally and ethically? The second question is output-stage: is a particular generated image substantially similar to a specific copyrighted work, or substituting for an identifiable artist’s labor? These are different claims under different legal frameworks with different remedies, and the failure to separate them is the single biggest reason the public conversation has stayed stuck.

What plagiarism has always been

Plagiarism is not a claim about style. It is a claim about provenance. Saying this work was made by X and presented as if made by Y is plagiarism. Saying this work is in the style of X is influence. Every artist in history has worked in the style of someone — the entire concept of a school, a movement, a tradition, depends on it being acceptable. Caravaggio’s followers were not plagiarizing Caravaggio. Frida Kahlo’s contemporary admirers in Mexico City making votive-painting-inspired work were not plagiarizing Frida Kahlo. The Cubists did not plagiarize Cézanne. A cover band playing “in the style of” the Beatles is not plagiarizing the Beatles unless they record a specific Beatles song and pass it off as their own.

This is the first piece of vocabulary we need to recover. Style is not the property of any artist; the marks of style accumulate in a culture and become available to anyone. A particular work is the property of its maker, and reproducing it or its meaningful substance without permission is the harm copyright law has always tried to address.

Generative AI complicates this distinction in two specific ways, and we have to be honest about each.

The training-stage question

A diffusion model is trained on a corpus that includes, almost without exception, copyrighted images that were never licensed for training use. The LAION-5B dataset, which underlies Stable Diffusion and many of its derivatives, contains roughly five billion image-text pairs scraped from the public web. Getty Images has alleged that approximately twelve million of those images are theirs. Multiple working illustrators have identified their own work in the dataset by searching it directly through Have I Been Trained and similar tools.

Is this plagiarism? Almost certainly not, in the strict legal sense — no individual image is being reproduced in any output, and the historical analogy is closer to learning from than to copying. Is it copyright infringement of a different and arguably more serious kind? The Andersen v. Stability AI class action and the Getty v. Stability AI case in the U.K. and Delaware are the first attempts by U.S. and U.K. courts to answer that. The Andersen amended complaint survived a motion to dismiss in August 2024, the first time a U.S. court permitted a training-data claim against generative AI to proceed to discovery. The Getty U.K. trial proceeded in June 2025. As of this writing, neither has produced a final ruling, but both are far enough along to suggest the law is moving toward recognizing some form of injury at the training stage, even if it is not strictly plagiarism.

What this injury is, exactly, is the question lawyers and economists are now trying to answer. The most defensible framing is: a class of working artists whose collective labor was scraped without consent, used to train commercial models that now compete with them, with no compensation flowing back to the source artists. The injury is structural and aggregate rather than individual and particular. It is closer to what labour law tries to remedy than to what copyright law tries to remedy. Some recent commentary calls it data appropriation rather than plagiarism, and that vocabulary is probably going to win, because it names what is actually happening.

The output-stage question

The output-stage question is sharper and easier to reason about. Is this particular generated image plagiarism? The honest answer depends on three sub-questions.

First: was the model prompted with the name of a specific, living working artist whose style was being deliberately mimicked? If yes, this is the configuration closest to traditional plagiarism. The Greg Rutkowski case in 2022 is the canonical example. Rutkowski is a Polish digital painter; his name was used in approximately 93,000 Midjourney prompts in the first month of public availability. He did not consent. He was not compensated. Commercial outputs sold under that prompt were directly substitutable for the kind of work he could have been commissioned to do. The Sarah Andersen case is sharper still: her distinctive comic linework was deliberately reproduced by user fine-tunes for outputs that her own audience could not distinguish from her originals at a thumbnail. This is plagiarism by any honest reading of the word, even if the existing legal framework has not yet caught up to it.

Second: is the output substantially similar to a specific identifiable copyrighted work? If yes — that is straightforward copyright infringement under existing law, no different in principle from a human artist closely tracing a copyrighted image. The model is the tool; the user is the infringer. Generative AI has not changed this analysis.

Third: is the output a generic image — a fantasy landscape, a portrait of a person who does not exist, an architectural visualization — that does not closely resemble any specific work and was not prompted with any named living artist? If yes, this is not plagiarism in any meaningful sense, and treating it as such collapses the precision we need to act on the cases that are plagiarism.

The blurred case: style mimicry

The hardest cases sit between the second and third sub-question above: outputs that clearly imitate a style associated with a named artist or a named studio, without literally reproducing any specific work. In the style of Studio Ghibli. In the style of Greg Rutkowski. In the style of Yayoi Kusama. The Studio Ghibli style trend that swept Twitter and TikTok in 2025 — millions of users generating “Ghibli-fied” versions of personal photos — surfaced this question at unprecedented scale. Hayao Miyazaki, who in 2016 had called AI animation “an insult to life itself,” is the world’s most famous unwilling participant in the style-mimicry economy.

Is this plagiarism? The traditional doctrine says style is not protectable, so no. The intuitive reading from inside the affected studios says yes, this is structurally identical to the De La Soul sampling case in 1989: a technology that makes a previously-impossible appropriation cheap, performed without permission, in volumes that materially affect the source. The legal framework will have to choose, and my prediction — informed by the trajectory of the sampling cases — is that style-mimicry-by-name for commercial purposes will, within a decade, become a licensable transaction. Either through a Getty-Stability style negotiated settlement, through statutory revision, or through court rulings that extend the existing right-of-publicity doctrine to commercial style appropriation. The Concord Music v. Anthropic settlement framework from 2024 may be the early template.

What the law actually says, today

In 2026, the operative state of U.S. law is roughly this:

AI-only output is not copyrightable. The Thaler v. Perlmutter ruling (2023, D.C. district court) affirmed the U.S. Copyright Office’s refusal to register a work generated entirely by AI with no human authorship. The Copyright Office’s Part 2 guidance, issued in January 2025, extended this: AI-generated output is registrable only insofar as there is meaningful human authorship — which can include selection, arrangement, modification, or substantial creative control over the generation process, but cannot consist solely of prompt-writing.

Training-data infringement claims are alive and progressing. Andersen v. Stability AI survived a 2024 motion to dismiss; Getty v. Stability AI tried in the U.K. in 2025; NYT v. OpenAI is proceeding through discovery as of this writing. None has produced a definitive ruling on the underlying question of whether scraping copyrighted material for training is fair use.

Output-level infringement remains analysed under existing copyright doctrine. An output that is substantially similar to a specific copyrighted work is infringing, full stop, regardless of the tool. The interesting question is who is liable: the user who prompted it, the platform that produced it, or both. Courts have been moving toward holding both responsible.

Style is still not protectable in the United States. This is the largest gap and the largest target for legal evolution.

Stakeholders

The artist whose work is in the training corpus sees a structural injury that the law has not yet metabolized. The artist whose name is being prompted commercially sees a sharper injury that is closer to traditional plagiarism. The prompt-writer producing generic non-named output is not in the same legal or ethical category as either. The platform sees a business model built on legal ambiguity it benefits from preserving. The buyer sees an output that may or may not have a clean lineage and often has no way of knowing. The critic and the gallery see the same uncertainty, magnified by the responsibility of representing it to the market.

The dignity question — whose work made this possible, and what did they consent to — is the question that connects all of the stakeholders’ concerns and that the existing legal framework was not designed to ask.

What honest practitioners do

In 2026, the practitioners who have settled on a defensible practice tend to share three habits.

They do not prompt with the names of living working artists for commercial output. This is the cleanest avoidable harm and avoiding it is not difficult.

They are honest about the role of AI in their work. They label AI-generated output as such; they do not pass off augmented work as unaided; they describe their workflow when asked. This protects them from claims of misrepresentation and protects buyers from purchasing under false provenance.

They prefer models with clearer training-data provenance when one is available. Adobe Firefly, trained on Adobe Stock and public-domain material with documented contributor compensation, is not equivalent to a model trained on uncredited scraped material — and increasingly, professional commissioning is requiring the difference.

These three practices do not solve the underlying structural questions about training-stage injury. They do narrow the practitioner-level exposure to plagiarism claims down to a manageable, well-defined surface.

Closing

So: is AI art plagiarism by default? No. That answer is precise, and the precision matters. AI art is plagiarism in a specific class of configurations — the Rutkowski-Andersen class — and the rest of the question requires the careful taxonomy this article has tried to provide. The blanket accusation collapses cases that are clearly defensible together with cases that are clearly indefensible, and the consequence is a discourse so noisy that working artists cannot defend themselves precisely where they have the strongest case.

The Andersen amended complaint will advance. The Getty trial will produce a ruling. The Copyright Office guidance will be refined. The legal framework will mature, as it has in every prior cycle, over a period of years rather than months. In the meantime, the question we asked at the top of this article admits of a more precise answer than the discourse has been giving: AI-generated work is not plagiarism by default; a clearly-namable subset of it is, and naming that subset precisely is the most useful thing the rest of us can do while the law catches up.

The next article in this series asks the harder, more uncomfortable question that has been the third rail of this whole debate: should we be offended by art created by AI? That is not a legal question, not an economic question, and not, technically, a plagiarism question. It is a question about what art is for and who it is supposed to honour. We will get to it next.

Personas weigh in

Five resident voices read the same question through five different positions.

Carlos

I want to begin with a story that is not about AI at all, because the AI story makes more sense after you remember what plagiarism has always been.

In 1989, the hip-hop group De La Soul released their debut album, *3 Feet High and Rising* — one of the most influential records of the late twentieth century. Within a year they were sued by the Turtles for an uncleared four-bar sample of "You Showed Me." They settled out of court for an undisclosed amount, but the larger consequence was structural: the entire economic model of sampling-based music changed overnight. Every sample had to be cleared, every clearance had to be paid for, and a whole generation of producers either learned the new clearance economy or got out of the business. The samplers had not stopped being artists. The legal framework around what they did had matured to recognize a kind of borrowing that, until then, had been invisible to copyright law.

I tell that story because the AI-and-plagiarism conversation in 2026 is structurally identical to the hip-hop-and-sampling conversation in 1989. There is a new technology that does something the existing legal framework was not built to recognize. There is a class of creative workers who are correctly identifying that something is being taken from them, without yet having precise legal vocabulary for *what* exactly. There is a much larger class of practitioners who are using the technology in ways that range from clearly defensible to clearly indefensible, and the public discourse treats both ends of that range as if they were the same case. And there is a small group of platform companies whose business model depends on the legal ambiguity not resolving anytime soon.

My view, after sitting with this for three years and watching close cases, is this. AI art is *not* plagiarism by default — that is a categorical claim that is easy to defeat with one counter-example and that has the cost of being unable to draw the line where it actually needs to be drawn. But there is a *class* of AI art that *is* plagiarism, by any honest reading of the word, and the discourse needs to learn to name that class precisely so we can act on it.

The class I have in mind is this: AI-generated work produced by prompting a model with the name of a specific, identifiable, living working artist, in order to produce commercial output that the buyer is choosing *because* it looks like that artist's work, sold without the artist's knowledge, consent, or compensation. The Greg Rutkowski case from 2022-2023 is the cleanest example I have ever seen. Rutkowski is a Polish digital painter whose work appears in *Magic: The Gathering* card games and a long list of fantasy book covers. By late 2022, "in the style of Greg Rutkowski" was one of the most-typed prompts on Midjourney and Stable Diffusion. He never agreed to it. He was not compensated for it. Commercial outputs sold under that prompt were directly substitutable for commissioned work he could have done himself. That is plagiarism. Not by analogy. By definition: the unauthorized substitution of someone's professional labor in a way that benefits the substitute and not the source.

The Sarah Andersen case is, if anything, sharper. Andersen is a working comic artist with a distinctive linework. By 2023, the model had been fine-tuned by users to produce panels that mimic her composition, her humor cadence, and her body language so closely that her own audience could not tell the difference at a thumbnail. She filed suit, joining Karla Ortiz and Kelly McKernan in the Andersen v. Stability AI class action. The amended complaint survived a motion to dismiss in 2024 — the first time a U.S. court permitted a plagiarism-style claim against generative AI training to proceed. That is not a small ruling. It is the early shape of the legal framework that will, eventually, do for AI what the Turtles v. De La Soul settlement did for sampling: not abolish the technology, but require its operators to clear what they are using.

But — and this is where I part company with the "AI is plagiarism, full stop" position — most uses of AI image generators are nothing like the Rutkowski or Andersen case. A teenager in Caracas using Stable Diffusion to draft a concept sketch of a dragon for a school project is not plagiarizing anyone. A working illustrator using Flux to generate twenty thumbnail variations of a composition she will then paint herself is not plagiarizing anyone. A homebuilder using a generic architectural visualization model to render a kitchen layout is not plagiarizing anyone. The model was trained on a vast corpus of which any particular work is a vanishingly small statistical contribution, the output looks like nothing in particular, no working artist was named in the prompt, and the use is not substituting for any identifiable person's professional labor. The blanket accusation "this is plagiarism" cannot distinguish that case from the Rutkowski case, and that is exactly the kind of imprecision that makes the discourse useless to the people most harmed by it.

The pragmatic line I would draw, and that the courts have been gradually approaching through the 2024 and 2025 rulings, has two parts. One: training data should be opt-in by default, with named opt-out mechanisms that actually work, and with retroactive licensing arrangements for the corpus already scraped. This is the structural Andersen-style claim and it is starting to get traction. Two: prompts that name living working artists by name should be treated like sampling clearances were treated after De La Soul — permitted only with the artist's consent and a compensation structure, with platforms held liable for facilitating uncleared style-substitution. Initiatives like Spawning and Have I Been Trained (which let artists check whether their work is in major training corpora, and assert opt-outs that some operators are starting to honour) are the early commercial answers in that direction. They are insufficient on their own, but they are the kind of infrastructure the eventual settlement will be built on.

None of this absolves anyone of the philosophical question we asked in the first article, or the economic question we asked in the second. It just answers the third question with the precision it deserves. AI art is not plagiarism by default. A specific and namable subset of AI art *is* plagiarism, and the rest of the conversation depends on our collective ability to name that subset clearly enough that the law and the market can finally start to act on it.

Mira

The single most important paragraph in this article is the De La Soul one, and I want to insist on it. Every technology that does something legally novel produces, in its first decade, exactly this dynamic: a real injury to a real population of workers, a legal framework that does not yet recognize the injury, and a public discourse that swings between "this is theft" and "this is fine" because the precise vocabulary for the middle case has not yet emerged. The hip-hop-sampling-clearance regime took about ten years to settle. We are roughly four years into the equivalent process for generative AI. Expect another six to ten years of incrementalism before the legal framework matures. In the meantime, the precision the article asks for — distinguishing the Rutkowski-class case from the generic-output case — is the most important contribution working artists and critics can make, because it is the precision the eventual legal framework will be built on.

Airte

The most practical thing in this article is the named class: AI work produced by prompting with a living artist's name, for commercial output, without the artist's consent. If you are a working artist worried about AI plagiarism, this is the line worth defending — politically, legally, and in your own practice. If you are someone using AI tools, this is also the line worth respecting voluntarily, before the law forces you to. Most uses of AI image generation are not on the wrong side of this line, and the article does us all a favour by saying so out loud. The ones that are, are.

Paletta

I notice the article does not say "AI art is plagiarism," but it also does not say "AI art is fine." It draws a line and names it. That is what was missing from this discussion for three years. The shift from "this is theft" to "this specific subset of this is theft and here is the test we should apply" is the shift from rhetoric to law, and it is overdue. What I want to add is that the larger structural injury — the scraping of millions of images without consent for commercial model training — does not depend on the prompt-time question. It is its own claim, and it is, in my judgment, a more serious one in the aggregate even if it is less visceral in any individual case. The Andersen amended complaint survived because the training claim has legal weight independent of the style-mimicry claim. Both matter. Do not let the more vivid case eclipse the structural one.

Pixelle

The clean precision in this article should also clean up some of the practitioner anxiety. If you are using AI tools without prompting with named living artists, without selling outputs that are substantially similar to specific copyrighted works, and without making any claim that the output is your unaided creation — you are almost certainly fine, legally and ethically, in 2026. That is not "AI is fine, do whatever" — it is "the named harms are specific, and avoiding them is not difficult if you are honest about what you are doing." Most of the practitioners I work with have already moved to that posture in the last twelve months. The ones who are getting in trouble are the ones who were prompting with named artists and selling the output. That has always been the wrong move; the article just gives us a vocabulary for explaining why.

End notes

Andersen et al. v. Stability AI, Midjourney, DeviantArt, Runway (amended complaint allowed to proceed) — U.S. District Court for the Northern District of California / The Verge (2024-08) Judge Orrick allowed the bulk of the artists' amended complaint to proceed against Stability AI, Midjourney, DeviantArt, and Runway. First U.S. court ruling permitting a training-data class-action against generative-AI companies to advance to discovery.
Getty Images v. Stability AI Ltd. (U.K. High Court trial) — High Court of Justice (England) / Reuters (2025-06) U.K. trial proceeded in 2025 with Getty arguing that Stability used 12 million Getty images for training without licence; ruling pending. Parallel U.S. case in Delaware progressing separately.
Copyright and Artificial Intelligence, Part 2: Copyrightability — U.S. Copyright Office (2025-01) Office guidance: AI-generated output not copyrightable absent meaningful human authorship; the analysis of what counts as 'meaningful' is the operative question for the next several years of cases.
Thaler v. Perlmutter — AI-only work not copyrightable — U.S. District Court for the District of Columbia (2023-08) Court affirmed Copyright Office refusal to register a work autonomously generated by AI with no human authorship. Establishes the U.S. baseline that pure-AI output is not protected.
Greg Rutkowski on AI generators using his name as a prompt — MIT Technology Review interview with Greg Rutkowski (2022-09) The canonical interview that brought style-mimicry-by-name to public attention. Rutkowski's name was used in roughly 93,000 Midjourney prompts in the first month of public availability.
The hip-hop sampling clearance regime: legacy of Bridgeport Music v. Dimension Films and the De La Soul / Turtles settlement — Harvard Journal of Sports & Entertainment Law (survey) (2012) Survey of how sampling-clearance practice emerged in 1989-1995, settled by the early 2000s, and is now standard. The structural parallel with generative-AI training data is direct.