AI Art, Copyright, and Ethics: Who Owns What an Algorithm Creates?

AI art copyright comes down to two questions that people often blur together. Who owns an image a model generates? And was it lawful, or fair, to train that model on millions of artworks without asking? US authorities currently say a purely machine-made image has no copyright owner, while courts in the US, UK, and Germany have reached different answers on training. This article explains the legal tests, the cases that matter as of October 2026, and the ethical arguments the law has not settled. It is education, not legal advice.

Two questions, not one

The first question is about outputs. When you type a prompt and a model returns an image, does anyone hold a copyright in it, and if so, who? This is the ownership question, and it decides whether a business can stop a competitor from copying its AI-generated logo, book illustrations, or product shots.

The second question is about inputs. Image generators learn from huge collections of pictures, many gathered from the public web. Did copying those pictures into a training set infringe the rights of the people who made them? This is the training question, and it decides whether an AI company owes artists permission or payment.

The two questions can have opposite answers. An image can be unprotectable (no one owns it) and still have been made by a model whose training was unlawful. Or a model can be lawfully trained and still produce an output that is too close to a specific protected work. Keeping them apart makes most of the news coverage easier to read. For the technology behind the training step, see AI Image Generators Explained.

Who owns what the machine makes

The US rule: no human author, no copyright

US copyright law protects works of human authorship. The US Copyright Office has repeated that position across several decisions and a formal report series. Part 2 of its report on AI, released January 29, 2025, addresses whether outputs created with generative AI can be copyrighted. As summarized by law firms that reviewed it, the Office concluded that "given current generally available technology, prompts alone do not provide sufficient human control to make users of an AI system the authors of the output." It also concluded that existing copyright principles are adequate to handle these questions.

The same report leaves room for protection. Human-authored material that is perceptible in an AI output can be protected, as can a creative selection or arrangement of AI-generated material, and creative modifications a person makes to an output. The line is drawn around what the person contributed, not around the tool they used.

Zarya of the Dawn: the split registration

The clearest early example is a comic book called Zarya of the Dawn. In a February 21, 2023 decision, the Office concluded that its author, Kristina Kashtanova, was the author of the text and of the selection, coordination, and arrangement of the book's words and images. The images generated by Midjourney were not the product of human authorship, so they were excluded from the registration. The Office reasoned that Midjourney's specific output cannot be predicted by the user, which distinguishes it from tools whose results an artist controls.

The practical lesson is that a work can be partly protected. A graphic novelist who writes the story and lays out the pages owns that creative work, while the raw generated panels sit outside copyright.

Thaler and the "pure AI" case

The strictest test case involved no human creativity at all. Stephen Thaler applied to register "A Recent Entrance to Paradise," a picture generated entirely by his AI system, naming the AI as sole author. In 2025, the US Court of Appeals for the D.C. Circuit affirmed the Copyright Office's refusal, holding that the Copyright Act requires a human being as the author in the first instance. The Supreme Court declined to hear the case on March 2, 2026, which leaves the appeals court decision standing. A denial of review is not a ruling on the merits, so it does not by itself answer the harder questions about mixed human and AI work.

Where the line is still blurry

How much human input is enough has no bright-line test. A pending case, Allen v. Perlmutter, challenges the Office's refusal to register an artwork that its creator says took more than 600 refining prompts. Law firm commentary on the Thaler denial recommends that creators document their contributions: drafts, selections, edits, and compositing work. That advice follows from the Office's logic, since protection attaches to human choices that can be identified.

How other countries compare

Copyright systems are national, so the answer changes at the border. The table summarizes the positions described in this section.

Jurisdiction	Position on AI-generated images	Key source
United States	No copyright for purely AI-generated output; human selection, arrangement, and modification can be protected	Copyright Office Part 2 (Jan 29, 2025); Thaler v. Perlmutter (D.C. Cir., 2025)
China	An AI-assisted image can be protected if it reflects original human intellectual input	Beijing Internet Court, Li v. Liu (Nov 27, 2023)
European Union	Training-data transparency and copyright-compliance duties for general-purpose model providers; no single ownership rule for outputs described here	EU AI Act obligations (from Aug 2, 2025)
United Kingdom	Court found Stable Diffusion's model weights were not infringing copies of Getty's images (a training question, not an ownership one)	Getty Images v. Stability AI [2025] EWHC 2863 (Ch)

In Li v. Liu, the Beijing Internet Court treated an image made with open-source image-generation software as a protected work of fine art. The court said AI has no free will or legal personality, so the creative input is attributed to the human who configures and directs the tool. A summary of the judgment notes the image was not drawn from stock images or preset templates, which supported the finding of human intellectual input. The case is a first-instance decision, and it shows why the amount of documented human involvement matters so much.

For businesses, the uncertainty is real. A company that commissions AI-generated marketing images in the US may be unable to stop others from copying the raw outputs. That is a reason many creative teams now treat generated images as raw material and add human design work on top.

Was training on scraped art lawful?

The fair use argument

The central US defense is fair use, a doctrine that permits some copying without permission, judged on four factors: the purpose of the use, the nature of the work, the amount copied, and the effect on the market. AI companies argue that training is transformative because the model extracts statistical patterns and does not republish the images. Artists and rights holders respond that training copies whole works at huge scale, and that the resulting model can substitute for commissioned art.

The most relevant Supreme Court precedent is not about AI at all. In Andy Warhol Foundation v. Goldsmith, decided May 18, 2023, by a 7 to 2 vote, the Court considered only whether the first fair use factor favored the Foundation's 2016 licensing of Warhol's "Orange Prince" image to Condé Nast. It held the factor did not: the licensed image and the original photograph both served as portraits of Prince illustrating magazine stories, and the use was commercial. The Court expressly limited its ruling to that licensing and took no view on Warhol's original creation of the works. The decision matters for AI because it moved the analysis away from asking only whether a use adds "new meaning" and toward asking whether it serves the same purpose as the original in the same market.

What the Copyright Office said about training

On May 9, 2025, the Office released a pre-publication Part 3 of its AI report, on generative AI training. It did not announce a blanket rule. According to law firm summaries, it said that various uses of copyrighted works in training are likely to be transformative, but whether they are fair depends on which works were used, from what source, for what purpose, and with what controls on outputs. It distinguished research-oriented uses from commercial uses that generate expressive content competing with the originals.

Two points in the report bear directly on art. First, it discussed "market dilution," the idea that a flood of AI-generated works can harm a category of creators without copying any one of them, while acknowledging that this was uncharted territory and that no court had yet relied on it to deny fair use. Second, it said stylistic imitation made possible by training on an original work may affect the creator's market. The report also concluded that, for now, licensing markets should be allowed to develop without government intervention. The Office said a final version would follow with no substantive changes expected, so treat the pre-publication text as the Office's current thinking, not as binding law.

The US court record so far

No US court has issued a final ruling on whether training an image generator on scraped artwork is fair use. Decisions in book-related cases give hints, though they involve text and are not binding in image cases.

Case	Court and date	What happened
Andersen v. Stability AI	N.D. Cal., filed Jan 13, 2023	Artists sued Stability AI, Midjourney, and DeviantArt; Runway AI is now also a defendant. In late 2023 Judge William Orrick let Andersen's training claim against Stability AI proceed while dismissing other claims with leave to amend. A February 2026 report says the summary judgment hearing moved to February 17, 2027.
Disney v. Midjourney	C.D. Cal., filed June 2025	Disney and Universal allege direct and secondary infringement, citing characters such as Darth Vader and Homer Simpson. Reports describe it as the first suit by major studios against a generative AI company.
Bartz v. Anthropic	N.D. Cal., ruling June 2025	Judge William Alsup held training on lawfully obtained books could be fair use but that keeping pirated copies was not protected. Anthropic agreed in September 2025 to a $1.5 billion class settlement covering hundreds of thousands of works.
Kadrey v. Meta	N.D. Cal., ruling June 25, 2025	Judge Vince Chhabria granted Meta summary judgment on the training claims but said the plaintiffs had failed to present evidence of market dilution and that the ruling does not make such training lawful in general.

Andersen is the closest US case to the art world, and it is still in the pretrial stage. The Disney suit is notable because the plaintiffs focus on outputs that reproduce recognizable characters, a theory that does not depend on how training is judged.

The book cases point in two directions. Judge Alsup separated lawful acquisition of training data from piracy, and Judge Chhabria stressed that the market effect of AI output could be decisive in a better-argued case. Artists should read both rulings as signals about which arguments courts find persuasive, not as final answers about images.

Outside the US: three different results

United Kingdom: Getty v. Stability AI

Getty Images sued Stability AI in the High Court of England and Wales. In its judgment of November 4, 2025, the court held that Stable Diffusion's model weights are "articles" under UK law, including intangible ones, but not "infringing copies," because the evidence showed they never contained or stored copies of Getty's works. Training took place in the US, and the court rejected Getty's attempt to reach the model through UK import provisions. Getty did succeed in part on a narrow trade mark claim, based on some generated images that carried its watermark. Commentators describe the ruling as narrow, since the case narrowed significantly before trial and the court did not decide the training-stage or output-stage copyright questions. Permission to appeal on the secondary infringement question has been granted, so this is not the last word.

Germany: the LAION case

A photographer, Robert Kneschke, sued the nonprofit LAION, which compiled the dataset used to train Stable Diffusion. The Hamburg Regional Court dismissed his claim on September 27, 2024, relying on the scientific research exception in German law. On December 10, 2025, the Higher Regional Court of Hamburg rejected his appeal, applying the general text and data mining exception and holding that an opt-out written in natural language did not meet the statutory requirement of being machine-readable. The court allowed a further appeal to the Federal Court of Justice. For artists, the practical point is that in the EU framework a rights reservation has to be expressed in a form machines can read, not just stated in plain language.

European Union: transparency duties

The EU AI Act imposes obligations on providers of general-purpose AI models. According to the European Commission, providers must implement a policy to comply with EU copyright law, using state-of-the-art technologies to identify and respect rights reservations, and must publish a sufficiently detailed summary of the content used for training. The obligations apply from August 2, 2025. The Commission says it will enforce full compliance, including through fines, from August 2, 2026, and models placed on the market before August 2025 have until August 2, 2027. This is a transparency regime, not a payment regime, but it gives artists a way to learn whether their work was used.

Does the model contain copies?

A recurring argument is that a diffusion model "stores no images." It holds billions of numerical weights, not a gallery. The Getty ruling accepted that factual point for Stable Diffusion's weights. But the argument has limits, and the Copyright Office's Part 3 report addresses them. If a model can reproduce identical or nearly identical copies of a work without that work being supplied in the prompt, the Office said, the weights could implicate the reproduction right, and weights that retain substantial protectable expression could implicate the derivative work right.

A German court has applied similar thinking to text models. In GEMA v. OpenAI, decided by the Munich District Court on November 11, 2025, memorisation of song lyrics within the language models met the requirements of a reproduction, even though the works existed only as probability values across parameters. That is a different medium and a different legal system, but it shows that "the model stores nothing" is not a universal defense. Whether image models memorize specific artworks is an empirical question, and it is one reason the output-side lawsuits, such as the Disney case, may matter as much as the training-side ones. For how training works at a conceptual level, see How AI Models Are Trained: The Full Pipeline Explained, and for how GANs and diffusion models learn to make images, see What Is Generative AI Art?

The style question: legal versus ethical

When someone prompts a model for art "in the style of" a living illustrator, has anything wrong happened? In US law, copyright protects specific expressions, not a general approach or look, so a style alone is generally not protected. The Copyright Office's Part 3 report nonetheless says stylistic imitation made possible by training may affect a creator's market, which is why style is part of the market dilution debate and not a closed question.

The ethical side is sharper. In September 2022, MIT Technology Review reported that Lexica, a search engine for Stable Diffusion images and prompts, counted about 93,000 uses of the name of Polish fantasy illustrator Greg Rutkowski as a prompt, while Michelangelo, Picasso, and da Vinci had about 2,000 or fewer each. Many of his works had been scraped from ArtStation. Rutkowski did not blame the people using his name, calling it "a cool experiment," but said it was "starting to look like a threat to our careers." At that time, the article reported, artists could not opt in or have their work removed from the dataset.

That example captures the gap between legal and ethical analysis. Nothing in the account establishes that a specific law was broken, yet an artist's name became a free input to a commercial system without his involvement. How much weight to give that harm is the real disagreement. For a technical look at how a model can imitate a look without copying a picture, see Neural Style Transfer: How AI Learns to Remix Artistic Style.

Models built on licensed data: the alternative path

Not every system follows the scrape-first approach. Adobe states that the first commercially released Firefly model was trained on Adobe Stock images, openly licensed content, and public domain content whose copyright has expired. Adobe also describes a Firefly bonus payment for eligible Adobe Stock contributors whose content was used in training. The terms are Adobe's own, and independent verification of how much contributors receive is limited, so the numbers deserve scrutiny. But the structure shows that a licensed training pipeline is possible when a company already holds a licensing relationship with its contributors.

The trade-off is cost and coverage. A licensed library is smaller than the open web and requires payment and administration. Whether that produces worse models, or simply different ones, is contested. The existence of such models weakens the argument that scraping was the only way to build a competitive generator, though it does not settle whether scraping was unlawful.

What fairer systems could look like

Proposals fall into a few families, and each has a known weakness.

• Opt-in licensing. Artists grant permission, usually for payment. Strength: clear consent. Weakness: most artists never see the offer, so the dataset shrinks and rare styles may be underrepresented.
• Opt-out registries and technical signals. Rights holders flag works they do not want used. Strength: low friction for developers. Weakness: it puts the burden on the artist, and the Hamburg ruling shows that vague wording may not count.
• Collective licensing. A collecting body negotiates and distributes payments, the way music licensing organizations do. The Copyright Office's Part 3 report discusses voluntary, collective, and compulsory licensing, noting that collective licensing could lower transaction costs if safeguards against anticompetitive conduct exist.
• Revenue sharing by influence. Payment proportional to how much a work shaped an output. Strength: intuitive fairness. Weakness: measuring influence across millions of training images is technically unsolved.

Technical tools offer partial protection. Glaze, built by the SAND Lab at the University of Chicago, adds changes to artwork that look unchanged to people but appear to AI models as a very different style, aiming to block style mimicry. Its developers say it is "not a panacea," that it is less effective against styles already in base models, and that future algorithms may overcome it. Provenance standards are a different tool. C2PA's Content Credentials attach metadata showing a piece of content's origin and edits, which helps with labeling and attribution but does not by itself stop training.

What people often get wrong

"AI-generated art is automatically public domain." In the US, the unprotected portion can be freely copied, but a human-authored arrangement, text, or edit around it can still be protected, as Zarya of the Dawn shows.

"Courts have ruled that training is fair use." No US court has issued a final ruling for image generators. The book-case decisions are narrow and fact-specific, and Judge Chhabria wrote that his ruling does not establish that such training is lawful in general.

"The model stores nothing, so there is no infringement." The Getty court accepted this for Stable Diffusion's weights, but the Copyright Office and a Munich court both recognize that memorized material can matter.

"Style imitation is illegal." It is generally not infringement by itself, though it is part of the market harm debate.

"One country's ruling applies everywhere." The UK, Germany, China, and the US have applied different rules, and the EU adds transparency duties on top.

What this means

The law is converging on a simple principle and diverging on the hard parts. The simple principle is that copyright rewards human creative choices: US and Chinese courts differ in how they count those choices, but both look for them. The hard parts are the training question and the market effect, where courts have so far split on facts, evidence, and jurisdiction.

For artists, the pattern in the cases is that specific, provable harm gets traction: reproduced characters, memorized works, pirated sources, watermarks. General claims that training is unfair have been harder to turn into legal wins. For AI companies, the direction of travel is toward documented data sources, which the EU already requires and which US litigation rewards. For everyone else, the useful habit is to ask which of the two questions, ownership or training, a headline is actually about. Related background is in Rise of AI Art: Creativity, Copyright Controversy, and Artist Backlash in 2026 and AI Ethics: The Key Debates Shaping How AI Gets Built.

Key takeaways

• US authorities hold that purely AI-generated works lack human authorship and are not copyrightable, while human selection, arrangement, and modification can be protected (Copyright Office Part 2, January 29, 2025; Thaler v. Perlmutter).
• Whether training on scraped art is fair use is unresolved for image generators; the main US artist case, Andersen v. Stability AI, has a summary judgment hearing set for February 17, 2027, according to a February 2026 report.
• UK and German courts have so far favored AI developers on specific points, while EU rules add transparency and rights-reservation duties.
• Style is generally not copyrightable, but the Copyright Office says stylistic imitation may harm a creator's market.
• Licensed-data models such as Adobe Firefly show that an alternative to scraping exists.

FAQ

Q: Can AI art be copyrighted? A: In the US, a purely AI-generated image cannot, because copyright requires human authorship. Human contributions such as creative selection, arrangement, or modification of AI material can be protected, but only those human-authored parts. China's Beijing Internet Court protected an AI-assisted image in 2023 where it found original human input.

Q: Is it legal to train AI on art scraped from the internet? A: It depends on the jurisdiction and facts, and many questions are open. US courts have not issued a final ruling on image generators. The UK High Court found Stable Diffusion's weights were not infringing copies, and a German court upheld a dataset-building exception for LAION. None of these is a blanket approval of scraping.

Q: Can an AI "steal" an artist's style? A: Style is generally not protected by copyright, so imitation alone is usually not infringement. The Copyright Office has said, however, that stylistic imitation enabled by training may affect an artist's market, so the issue is part of the legal debate and the ethical one.

Q: What can artists do to limit use of their work? A: Options include style-protection tools such as Glaze, which has stated limits, and machine-readable rights reservations that EU courts and the AI Act's copyright-policy duties refer to. Provenance standards such as C2PA help label content but do not prevent training. Artists with specific disputes should consult a qualified attorney.

Q: Are AI companies compensating artists? A: Some are, within limited programs. Adobe describes a Firefly bonus for eligible Adobe Stock contributors, and Anthropic agreed to a $1.5 billion settlement over pirated books in 2025, though that case involved authors, not visual artists. Most image-generator training to date has not involved payment to individual artists.

Q: What happens if an AI image copies a famous character? A: That is an output-side problem and does not depend on how training is judged. Disney and Universal sued Midjourney in June 2025 on exactly that theory, alleging direct and secondary infringement. The case is pending.

Sources

• US Copyright Office, Copyright and Artificial Intelligence (Parts 1 to 3): copyright.gov/ai
• Mayer Brown, "Supreme Court Denies Review in AI Authorship Case" (March 2026): mayerbrown.com
• Andy Warhol Foundation v. Goldsmith, No. 21-869 (2023), Cornell Legal Information Institute: law.cornell.edu
• Sidley Austin, highlights of the Copyright Office's Part III report: datamatters.sidley.com
• AIPPI, "Getty Images v Stability AI: UK High Court Finds No Secondary Copyright Infringement": aippi.org
• Bird & Bird, "Higher Regional Court Hamburg Confirms AI Training was Permitted (Kneschke v. LAION)": twobirds.com
• European Commission, guidelines on obligations for general-purpose AI providers: digital-strategy.ec.europa.eu
• MIT Technology Review, on Greg Rutkowski and Stable Diffusion (September 16, 2022): technologyreview.com
• Glaze project, University of Chicago SAND Lab: glaze.cs.uchicago.edu
• C2PA, Content Credentials: c2pa.org

---

Read next:

• Generative AI Art: How Diffusion Models Actually Work
• Text-to-Image AI: DALL-E, Midjourney, and Stable Diffusion Compared
• AI Music Composition: How Neural Networks Learn to Make Music