Generative AI copyright: what the US Copyright Office and courts say

The short version

• Using copyrighted works to train a generative AI model (software that produces new text, images or other content) can infringe copyright, and the usual defence is fair use.
• Fair use turns on four statutory factors, and US courts have split on how they apply to training.
• In 2025, training large language models (the AI systems behind chatbots, which generate text) was held fair use in Bartz v. Anthropic and Kadrey v. Meta, but building a library from pirated copies was not (Bartz), and a non-generative legal-research tool lost in Thomson Reuters v. ROSS.
• Purely AI-generated output is not copyrightable: US copyright needs a human author, and prompts alone are not enough.
• The US Copyright Office’s reports are guidance, not law; its training report is still a pre-publication draft, and dozens of cases are pending.^{1,2,3,4,5,6,7,8}

Does training generative AI on copyrighted works infringe copyright?

It can. The US Copyright Office says building a generative model uses an owner’s exclusive rights at several stages: collecting the works, making copies of them to train on, and sometimes reproducing them in a model’s outputs. Each of those is infringement unless something excuses it, and the main excuse the developers rely on is fair use.²

Training data is the body of existing works a model is shown so it can learn patterns from them. Two questions follow: whether using the data to train is lawful, and who owns what the model then makes. This guide covers both. On the first question, the party that did the copying carries the burden of proving fair use.^2,3

How the law regulates AI in general, and who is liable for what AI says, sits in AI in law. The European Union reaches these questions by a different route; the EU AI Act and how it applies to dispute resolution sits in the EU AI Act in dispute resolution.

What is fair use, and what are the four factors?

Fair use (17 U.S.C. § 107) lets someone use a copyrighted work without permission in some cases. It is the defence at the centre of whether training AI on copyrighted works is fair use. A court weighs four statutory fair-use factors together, and no single one decides.³

• The purpose and character of the use, including whether it is commercial, and whether it is transformative. Transformative use means use that adds a new purpose or meaning rather than standing in for the original; the word is a judicial gloss and is not in the statute.
• The nature of the copyrighted work.
• The amount and substantiality of what was used.
• The effect of the use on the potential market for or value of the work.³

On whether fair use covers AI training, the Copyright Office AI report reads training as often transformative. It wrote that “various uses of copyrighted works in AI training are likely to be transformative,” but that how fair they are will depend on what works were used, from what source, for what purpose, and with what controls on the outputs. It expects the first and fourth factors to carry considerable weight. This is the Copyright Office AI Part 3 training report, and as of 11 November 2026 it is a pre-publication draft, not law.²

How have US courts ruled on AI training so far?

Three district-court orders in 2025 point in different directions. Each is fact-specific, and none binds the whole country.^4,5,6

The Thomson Reuters v. ROSS ruling held that copying Westlaw headnotes (the editorial summaries of the points of law at the head of a case report) to train a non-generative legal-research tool was infringement and not fair use, with the first and fourth factors going to Thomson Reuters. The judge added that “only non-generative AI is before me today.” The Third Circuit affirmed on interlocutory appeal (an appeal taken before the case has ended) on 29 September 2026, and the rest of the case continues below.^4,7

On Bartz v. Anthropic (Anthropic’s Claude models), fair use went the other way on the training itself. The court called training the large language models on the works “exceedingly transformative,” even “spectacularly so,” and a fair use, and held that format-shifting (converting lawfully purchased print books to digital) was fair use. Copying whole books did not defeat fair use, because the court weighed the amount used against that transformative purpose. Building a permanent library from pirated copies was not. The order separates how a company acquires data from what it then does with it.⁵

The Kadrey v. Meta copyright order (Meta’s Llama models) granted Meta summary judgment (a ruling made without a full trial) on fair use on this record only, because these thirteen plaintiffs failed to prove the training harmed the market for their works. The judge wrote that these plaintiffs “made the wrong arguments,” and that in cases like this one “the plaintiffs will often win.” It is not a class action. For the wider picture of who is liable for what AI does, see AI in law.⁶

Why do the cases split? Market harm and market dilution

The fourth factor, the effect on the market, is doing most of the work, and the contested new idea is market dilution. In Kadrey the judge called harm to the market “more important than the purpose for which the copies are made,” and said better-evidenced plaintiffs will often win.^2,6

Market dilution is the argument that a model trained on an author’s works can flood the market with similar machine-made works and so cut demand for the originals, even when no single output copies any one work. The Copyright Office’s Part 3 says that “the speed and scale at which AI systems generate content pose a serious risk of diluting markets for works of the same kind as in their training data.” That is the Office’s own reading of the fourth factor, and in Kadrey the plaintiffs did not prove it. It is not settled law as of 11 November 2026.^2,6

Can AI-generated output be copyrighted?

No, not the purely AI-generated parts. Whether you can copyright AI output comes back to authorship: US copyright needs a human author. The Copyright Office will not register material that is entirely generated by AI, and it concluded that prompts alone do not give a person enough control over the result to count as authorship. So copyright does not attach to the machine-made parts on their own.^1,8

Human authorship is the requirement that a work’s original expression come from a person, not a machine. What does get protection is a person’s own expression in a work, together with their creative selection, coordination, arrangement or modification of AI output; copyrightability is decided case by case. That also answers who owns AI-generated art: the human author of the protectable human parts, not the model. To register an AI-assisted work, an applicant discloses the AI-generated material and claims the human contribution. The Office concluded existing law is adequate and that no new or separate right for AI output is needed.¹

Thaler v. Perlmutter upheld the refusal to register a wholly AI-generated image, holding that the Copyright Act requires a human author. The Supreme Court declined to hear the appeal on 2 March 2026.^8,9

The training cases and the output question are separate. Whether a model’s output itself infringes someone else’s work is still largely untested. In Bartz the judge noted that if the outputs seen by users had been infringing, the authors would have a different case, but the authors challenged only the inputs.⁵

What does it mean if you use, license or dispute AI material?

Three points come out of the sources, and the first is ownership: you may not own the purely AI-generated parts of what you make. Where the training data came from matters as much as what the model does with it. And provenance, a documented record of where a work or its training data came from, together with licensing and indemnity, is what to check before anyone relies on AI material commercially.^1,2,5

If you license or sell AI content, expect a US registration to cover only the human-authored portion, and disclose the AI contribution, so the copyright risk of using AI-generated content commercially is mostly about what you do not own. If you train or deploy a model, the clearest dividing line in the orders is lawful sourcing: lawfully acquired works fared better than pirated ones. On whether you need a licence to train AI on copyrighted data, no court has held that training is lawful in general, so licences and AI copyright indemnity terms are the practical answer where the use is commercial.^2,5,6

The risk-management standards an organisation uses to govern this are in AI governance standards. Who answers when an AI output makes a false and damaging claim, a related accountability question, is in AI defamation: who is liable when AI output lies about you. The section overview is AI in dispute resolution.

What is still undecided?

A great deal. The Copyright Office said dozens of lawsuits are pending and decided none of them.²

• Part 3 is a pre-publication draft, not law; the Office says a final version will follow without any substantive changes expected in the analysis or conclusions.
• Bartz set the pirated-copies claim for trial; the court decided only fair use on the training and on the format-shift.
• Kadrey is limited to the thirteen plaintiffs and does not make training lawful in general.
• Thomson Reuters v. ROSS went up on interlocutory appeal; the Third Circuit (opinion filed 29 September 2026) decided only the two certified questions, copyrightability and fair use, and the rest of the case has not been tried.
• The copyright suits against OpenAI are consolidated as In re: OpenAI, Inc., Copyright Infringement Litigation, No. 1:25-md-3143, in the Southern District of New York. The Third Circuit’s September 2026 ROSS opinion described that litigation as pending and noted the Department of Justice’s statement of interest filed in it. We could not open the court’s own docket, so we state nothing about its merits as of 11 November 2026.
• We could not confirm any Anthropic copyright-settlement figure against the court’s own approval order, so we leave it out.
• Whether AI outputs, as opposed to training inputs, infringe is largely untested.^2,4,5,6,7

What does this mean for anyone building or buying AI?

In our view, the detail matters less than the direction these sources share. Existing copyright law already reaches generative AI. The Copyright Office frames training as infringement that fair use must excuse, and the courts decide it on the ordinary four factors. The reports and the pending cases are filling in the detail, not building a new regime. A person or an organisation stays accountable for what a model is trained on and for what it produces. The clearest dividing line in the orders is lawful sourcing, not how clever the use is: lawfully bought works fared better than pirated ones. Because copyright needs a human author, anyone relying on AI-generated material should assume they may not own the purely machine-made parts, and should check provenance and licences before they lean on it. The pace of the technology is ahead of the settled case law, so the safest reading is that the ordinary duties of authorship and fair use already govern, and the cases and reports are catching up.^1,2,4,5,6