US copyright law has two separate AI questions. First, a work needs a human author to be protected, so purely AI-made material is not. Second, whether training AI on copyrighted works is fair use has no single answer yet. Trial judges and one appeals court have ruled, with different results depending on the facts. This page summarizes the Copyright Office reports and the main rulings as of 24 October 2026.
01Two questions that often get mixed up
The first question is about outputs: can a picture, text or song made with AI be protected by copyright? The second is about inputs: may a company copy books, images or articles to train an AI system without permission? Different rules and cases apply to each. This page is general information, not legal advice. For a specific work or dispute, ask a copyright lawyer.
02Can a work made with AI be copyrighted?
The Copyright Office's January 2025 report says copyright protects the original expression of a human author, even if the work also contains AI-generated material. It does not protect purely AI-generated material, or material where a person lacks enough control over the expressive elements. It says that with today's tools, prompts alone do not give enough control. A person can hold copyright in their own contributions, such as creative selection and arrangement, or edits to AI output. The Office says each case turns on its facts.
In Thaler v. Perlmutter, the D.C. Circuit affirmed on 18 March 2025 the refusal to register an image that listed an AI system as its only author. The Supreme Court declined to hear the case on 2 March 2026 and gave no opinion, so the appeals court decision stands.
03Is training AI on copyrighted works allowed?
Fair use is a legal defense that lets some copying happen without permission. A court weighs several factors, including harm to the market for the original. The Office's May 2025 report on training says the key question is whether copying works for AI training is fair use. It says the answer depends on the works used, where they came from, why they were used and what limits exist on the outputs. It says commercial use of vast numbers of works to make competing content, especially through illegal access, goes beyond established fair use. It is labeled a pre-publication version, with a final version to follow.
Courts have started to rule. The results differ because the facts differ.
- Bartz v. Anthropic, 23 June 2025. A California federal judge found it was fair use to train language models on books the company lawfully obtained, and to scan purchased print books into digital copies. Keeping pirated copies in a central library was not. The case later settled as a class action, and the court granted final approval on 20 July 2026, according to the official settlement website.
- Kadrey v. Meta, 25 June 2025. Another California judge ruled for Meta because thirteen authors offered no meaningful evidence of harm to the market for their books. The judge wrote that in most cases such copying would likely be illegal, since fair use typically does not cover copying that significantly harms the market for the originals.
- Thomson Reuters v. ROSS, 29 September 2026. The Third Circuit, an appeals court, held that training a legal search tool on Westlaw headnotes was not fair use. ROSS meant to compete directly, and its tool returned existing court passages instead of writing new text. The court said concerns raised in a Justice Department filing in the OpenAI copyright litigation do not apply to its facts.
04What the White House and Congress have said
The White House's March 2026 legislative recommendations say the administration believes training on copyrighted material does not violate copyright law. They acknowledge arguments to the contrary and support letting courts decide. They ask Congress not to act in ways that affect that question. They suggest Congress consider collective licensing frameworks, and a federal framework on AI-made digital replicas of a person's voice or likeness.
The Copyright Office's training report says legislation would be premature while licensing markets develop. We found no new federal copyright statute on AI as of 24 October 2026.
05If you make or publish AI-assisted work
These steps are habits, not a guarantee of protection.
- Keep a record of what you made yourself and what the AI produced, including drafts, edits and your choices.
- If you register a work, expect to disclose more than a trivial amount of AI-generated material and describe your own contribution, as the Office's guidance describes.
- Follow the court cases if you are a creator worried about training. Rights depend on the facts of each case.
A court ruling applies to its own facts. One judge's fair use decision does not decide your situation, and none of this page is legal advice.
?What is not known
- Whether other appeals courts or the Supreme Court will decide fair use for generative AI training. The Third Circuit case involved a tool that did not write new text.
- How much human input makes an AI-assisted work protectable. The Copyright Office says this is decided case by case.
- Whether the Office's final training report will differ from the pre-publication version, and whether Congress will act.
QQuestions people ask
Can I copyright something I made with an AI tool?
Only the parts a person authored. The Copyright Office says prompts alone are not enough with current tools, but your own writing, selection, arrangement and edits can be protected.
Is it illegal to train AI on copyrighted books?
No court has given one answer for all cases. One judge found training on lawfully obtained books was fair use, and an appeals court found no fair use for a competing search tool. Pirated copies were treated differently.
Where can I read the Copyright Office reports?
They are on copyright.gov in its artificial intelligence section. The links to the two parts we used are in the sources below.
Last checked October 24, 2026. Rules and settings change, so open the source to see the current version.
- Copyright and Artificial Intelligence, Part 2: Copyrightability U.S. Copyright Office The January 2025 report; we used its conclusions on human authorship and prompts, and its registration guidance.
- Copyright and Artificial Intelligence, Part 3: Generative AI Training (pre-publication version) U.S. Copyright Office The May 2025 report; we used its conclusion on fair use, licensing and legislation.
- Thaler v. Perlmutter, No. 23-5233 U.S. Court of Appeals for the D.C. Circuit The 18 March 2025 opinion on whether a machine can be an author.
- Thaler v. Perlmutter, Supreme Court docket 25-449 Supreme Court of the United States The docket; we used it for the 2 March 2026 denial of review.
- Bartz v. Anthropic, Order on Fair Use U.S. District Court, Northern District of California The 23 June 2025 order; we used its holdings on training, scanning and pirated copies.
- Kadrey v. Meta, order on summary judgment U.S. District Court, Northern District of California The 25 June 2025 order; we used its reasoning on market harm.
- Thomson Reuters v. ROSS Intelligence, No. 25-2153 U.S. Court of Appeals for the Third Circuit The 29 September 2026 opinion; we used the fair use holding and its note on generative AI.
- Bartz v. Anthropic settlement website Court-appointed settlement administrator The official case site; we used it for the final approval date of 20 July 2026.
- National Policy Framework for Artificial Intelligence, Legislative Recommendations The White House The March 2026 recommendations; we used their section on intellectual property.
General information, not legal, medical, financial or investment advice. We are not affiliated with any company, agency or standards body named here. See our editorial policy and corrections, or tell us about a mistake at [email protected].