Is Using AI Copyright Infringement? The Law, Explained

Is AI copyright infringement? It depends on which part of the process you mean. Copying protected work to train a model and generating an output that resembles someone else's work are two different legal questions, and U.S. courts have started answering them differently. As of mid-2026, several federal rulings have found that training on lawfully acquired copyrighted works can qualify as fair use—but using pirated copies to build that training set, or generating outputs that closely mimic protected characters or text, can still create real liability.

• What does "copyright infringement" mean when applied to AI?
• Training-data copying vs. output copying: two different legal questions
• The fair-use arguments AI companies actually make
• How U.S. courts have actually ruled so far
• Can AI-generated output itself be copyrighted?
• Famous AI copyright lawsuit examples
• FAQ

What does "copyright infringement" mean when applied to AI?

Copyright infringement means copying, reproducing, distributing, or making a "derivative work" from someone else's protected expression without permission or a legal exception like fair use. Applied to AI, that basic rule gets stretched across two different moments in a system's life.

The first moment is training: a company ingests millions of books, articles, images, or songs into a dataset to teach a model statistical patterns. The second moment is generation: a user types a prompt and the model produces new text, images, or audio in response. Both moments can raise copyright questions, but different ones—which is why courts have started splitting them apart instead of issuing one blanket answer.

Training-data copying vs. output copying: two different legal questions

Training-data copying asks whether copying protected works into a training set—without a license—is itself infringement, or whether it counts as fair use. Output copying asks the more traditional question: is this specific generated image, song, or block of text "substantially similar" to a specific protected work?

Courts are treating these as separate inquiries. In Getty Images v. Stability AI, the UK's High Court held in November 2025 that Stability AI's model weights are not a "copy" of Getty's photos in the legal sense—the model contains statistically trained parameters, not stored reproductions—while still finding Stability liable for limited trademark infringement because some outputs reproduced Getty's watermark. Meanwhile, Disney, Universal, and Warner Bros. v. Midjourney, filed in June 2025, is built almost entirely around the output side: the studios allege Midjourney's tool generates recognizable copies of copyrighted characters on request, not just that it trained on their catalogs.

That split matters for anyone searching "is ai art copyright infringement" or "is ai scraping copyright infringement": scraping and training are being litigated under fair use doctrine, while a specific AI-made image that closely reproduces a specific protected character or photo looks more like ordinary infringement law that predates AI entirely.

The fair-use arguments AI companies actually make

U.S. fair use is a four-factor balancing test, and AI companies lean hardest on the first factor: purpose and character of the use. Their core argument is that training is "transformative"—the model isn't republishing a book, it's learning statistical patterns from it to produce something functionally different. Judge William Alsup used almost exactly that language in Bartz v. Anthropic, calling the training use "exceedingly transformative."

The other three factors get argued too: the nature of the copyrighted work, the amount used (companies argue the whole work often has to be ingested to learn from it, even though only patterns are retained), and market effect (companies argue a model doesn't substitute for buying the original book or photo). That last factor is the shakiest in practice—in Kadrey v. Meta, Judge Vince Chhabria ruled for Meta on the record presented, but wrote separately that a "market dilution" argument might have won if the plaintiffs had actually argued and supported it with evidence.

How U.S. courts have actually ruled so far (as of mid-2026)

No single ruling has settled "is AI training copyright infringement" for the industry as a whole—these are individual cases decided on individual facts. But several concrete decisions exist:

• Bartz v. Anthropic (N.D. Cal., June 2025): Judge Alsup held that training on legally purchased and digitized books was fair use, but that acquiring and retaining millions of books from pirate shadow libraries was not. Anthropic settled the resulting class claims for $1.5 billion in September 2025.
• Kadrey v. Meta (N.D. Cal., June 2025): Summary judgment for Meta on fair use, but narrowly—the court said this specific group of authors failed to present evidence of market harm, not that AI training is broadly lawful.
• Thomson Reuters v. ROSS Intelligence (D. Del., February 2025): the first federal ruling to reject an AI fair-use defense. Ross used Thomson Reuters' Westlaw headnotes to train a competing legal-search tool; the court found that use commercial and market-substituting. Now on appeal to the Third Circuit.
• Getty Images v. Stability AI (UK High Court, November 2025): Getty's core copyright claim was dismissed on procedural grounds (the training didn't happen in the UK), so the underlying "is training infringement" question was never decided on the merits there.
• New York Times v. OpenAI and Microsoft: allowed to proceed past a motion to dismiss in March 2025, but as of mid-2026 remains in discovery with no ruling on the merits.

Other major cases—including Disney/Universal v. Midjourney and the consolidated In re OpenAI Copyright Infringement Litigation (MDL No. 3143)—are still pending. The pattern so far: courts are more sympathetic to fair use when the training source was legally acquired and the output doesn't compete directly with the original; less sympathetic when the input was pirated or the output substitutes for the original product.

Can AI-generated output itself be copyrighted?

This one is more settled than the infringement question. The U.S. Copyright Office's position, upheld by the courts, is that copyright requires human authorship. A work generated entirely by AI, with no meaningful human creative control over its expression, cannot be copyrighted.

That rule was tested directly in Thaler v. Perlmutter, where computer scientist Stephen Thaler tried to register an image credited solely to his AI system, "the Creativity Machine." The D.C. Circuit affirmed in March 2025 that an AI cannot be an "author" under the Copyright Act, and the Supreme Court declined to hear the case in March 2026, leaving the human-authorship rule in place.

The nuance people miss: this doesn't mean nothing involving AI can ever be copyrighted. The Copyright Office says what matters is "the extent to which the human had creative control over the work's expression." An image where you select, arrange, and substantially modify AI-generated elements may still be protected for your human contribution—it's a fully autonomous, prompt-and-done AI output that falls outside copyright.

Famous AI copyright lawsuit examples

People searching "copyright lawsuit examples," "famous copyright lawsuit cases," or "biggest copyright lawsuit" are usually looking for names they can check. Here are five real, well-documented ones—see /ai-lawsuits/ for the full, continuously updated case-by-case tracker with filing dates and current status:

• Bartz v. Anthropic—authors' class action over pirated training books; ended in a $1.5 billion settlement, the largest publicly known copyright payout in an AI case so far.
• New York Times Co. v. Microsoft and OpenAI—alleges ChatGPT was trained on and can reproduce Times journalism; ongoing, no merits ruling yet.
• Thomson Reuters v. ROSS Intelligence—the first case where a court rejected an AI fair-use defense outright; now on appeal.
• Getty Images v. Stability AI—the first major UK ruling on AI and copyright; Getty's core claim was dismissed on procedural grounds, with a narrow trademark win.
• Disney, Universal, and Warner Bros. v. Midjourney—studios suing over AI-generated copyrighted characters, a case built around outputs rather than training data.

For the artist-consent side of this fight—opt-out tools, style imitation, what creators can do now—read AI art theft.

FAQ

Is generative AI copyright infringement?

Not automatically. Courts so far have found that training on lawfully acquired material can qualify as fair use, while training on pirated copies, or generating outputs that closely reproduce a specific protected work, is much more likely to be found infringing. It's decided case by case.

Is AI training copyright infringement?

Sometimes. Federal courts found training on legally obtained books to be fair use in Bartz v. Anthropic and Kadrey v. Meta. A court found training a directly competing commercial product on proprietary content was not fair use in Thomson Reuters v. ROSS Intelligence. Acquisition method and competitive market effect both matter.

Can AI be copyrighted—can an AI system own a copyright?

No. U.S. copyright law requires a human author. The Copyright Office and the D.C. Circuit (in Thaler v. Perlmutter) confirmed an AI system cannot be listed as an author, and the Supreme Court declined to disturb that rule in 2026.

Is AI scraping copyright infringement?

Scraping to build a training set raises the same fair-use questions as training generally, but courts look hard at how the data was obtained. Downloading from pirate "shadow libraries" weighed heavily against Anthropic in Bartz, even though the training use itself was found transformative.

Can AI detect copyright infringement?

AI tools can flag likely matches—similar phrasing, image similarity, audio fingerprinting—but they don't make legal determinations. Whether a match is actually infringement still depends on human legal analysis of access, substantial similarity, and any fair-use defense.

Conclusion: so, is it infringement?

"Is AI copyright infringement" doesn't have one answer because it isn't one question. Training on legally acquired material has, in several 2025 rulings, been found to be fair use; training on pirated material, or generating an output that substitutes for a specific protected work, has not fared as well in court. The Copyright Office's human-authorship rule is much more settled: fully AI-generated output isn't copyrightable, though human-directed edits and arrangements can be.

None of this is finished. Major cases—against OpenAI, Midjourney, and others—are still working through discovery and appeals. Track how each one is actually resolved, case by case, at /ai-lawsuits/, and if the artist-consent side of this issue is what brought you here, read AI art theft next.