AI Copyright Law: Ownership, Training and Fair Use (2027)

As of November 4, 2026, US copyright law requires a human author, so an AI system cannot be the author of a work (Thaler v. Perlmutter, D.C. Cir. 2025), and the Copyright Office says purely AI-generated material is not protected. Whether training a generative AI model on copyrighted works is fair use remains undecided at the appellate level in the opinions we read, and the two generative AI fair use rulings on this page come from federal judges in the same district who disagree on key points. Liability for what an AI tool outputs is also still being worked out.

Status as of November 4, 2026: The Supreme Court denied review in Thaler on March 2, 2026, without an opinion. In September 2026, two federal appeals courts ruled in AI cases: the Third Circuit in ROSS, on a non-generative tool, and the Ninth Circuit in Doe v. GitHub, on a DMCA claim. Neither decided whether training a generative model is fair use. The $1.5 billion Bartz v. Anthropic settlement received final approval on July 20, 2026, and class counsel reported on September 2 that the two appeals filed (9th Cir. Nos. 26-5391, 26-5455, per the district docket) concern only the attorney's fee award. Summary judgment was argued in Concord v. Anthropic on October 21, 2026, with no ruling in the docket entries we read through November 3.

Jurisdiction scope: This page covers United States federal copyright law as it applies to artificial intelligence: ownership of AI-generated works, the training-data cases, and liability for output. It does not address patent law, and it covers trade-secret law only in passing. For software specifically, see the code guide linked below; for other countries, see how AI and copyright differ worldwide.

Can AI-Generated Work Be Copyrighted? Thaler v. Perlmutter and the Copyright Office

Not if no human authored it. In Thaler v. Perlmutter, the D.C. Circuit affirmed the denial of a registration for an image where Dr. Stephen Thaler listed his AI system, the "Creativity Machine," as sole author. The court held that the Copyright Act of 1976 "requires all eligible work to be authored in the first instance by a human being."

That ruling is narrower than many headlines suggest. The court did not decide whether the Constitution requires a human author, and it did not reach Thaler's argument that he was the author, because he waived it before the agency. The Supreme Court then denied review on March 2, 2026 (No. 25-449). That denial came without an opinion and is not a ruling on the merits, so headlines saying the Supreme Court ruled on AI copyright are wrong.

The D.C. Circuit also said what its rule does not do. The requirement "does not prohibit copyrighting work that was made by or with the assistance of artificial intelligence." The author must be a human being, not the machine.

What the Copyright Office says about prompts and AI-assisted work

The harder question is how much human involvement is enough, and Thaler did not answer it. The Copyright Office addressed it in Part 2 of its AI report, published January 29, 2025. These are the Office's conclusions:

• Using AI as a tool to assist human creativity "does not affect the availability of copyright protection for the output."
• Copyright "does not extend to purely AI-generated material, or material where there is insufficient human control over the expressive elements."
• "[G]iven current generally available technology, prompts alone do not provide sufficient human control to make users of an AI system the authors of the output."
• Human authors keep copyright in their own works that are perceptible in AI outputs, and in "the creative selection, coordination, or arrangement of material in the outputs, or creative modifications of the outputs."
• Whether human contributions are enough "must be analyzed on a case-by-case basis."

The Office also concluded: "The case has not been made for additional copyright or sui generis protection for AI-generated content." Its example of the split result is Zarya of the Dawn: the registration covered the human-written text and the human selection and arrangement of the elements, but not the Midjourney images themselves.

Registering a work that contains AI material

The Office's March 2023 guidance, 88 Fed. Reg. 16,190, gives applicants "a duty to disclose the inclusion of AI-generated content in a work submitted for registration and to provide a brief explanation of the human author's contributions." AI-generated content that is more than de minimis should be explicitly excluded from the application.

The guidance warns that omitting this information can have consequences: the Office may cancel the registration, and under § 411(b) a court may disregard it in an infringement action if the applicant knowingly gave inaccurate information and accurate information would have led to refusal. A second court challenge to the Office's refusal to register an AI-generated output, Allen v. Perlmutter, was filed in the District of Colorado in 2024. The case metadata we read on November 4, 2026, showed it open, with a last filing on June 9, 2026.

Is AI Training Fair Use? Bartz, Kadrey and Thomson Reuters v. Ross

No appellate opinion we read answers this for generative AI. Fair use under 17 U.S.C. § 107 weighs four factors, and the courts below have stressed different ones.

Thomson Reuters v. ROSS Intelligence (3d Cir. 2026)

On September 29, 2026, the Third Circuit wrote (No. 25-2153): "We therefore hold that Thomson Reuters's materials are copyrightable and that ROSS's use was not fair." It affirmed partial summary judgment on an interlocutory appeal, on ROSS's use of 2,243 Westlaw headnotes to train a legal search tool meant to compete with Westlaw.

The panel found ROSS's use "minimally transformative, at best," because it shared the same ultimate purpose as Westlaw. The second factor weighed slightly toward fair use, and the first, third and fourth against it.

The court drew the line itself. ROSS's AI "was not a generative AI," and "[u]nlike the AI models in Bartz and In re: OpenAI, ROSS's AI platform cannot generate original expression." So the decision does not resolve the generative AI question, though its treatment of the licensing market matters to it, as discussed below.

Bartz v. Anthropic: fair use order and $1.5 billion settlement

Judge Alsup ruled on June 23, 2025, that using the books at issue to train Claude "was exceedingly transformative and was a fair use under Section 107 of the Copyright Act." Scanning print books Anthropic had bought was also fair use, for a different reason: it replaced purchased copies with digital ones, without adding copies or redistributing them.

The pirated books came out the other way. "Anthropic had no entitlement to use pirated copies for its central library," and building "a permanent, general-purpose library was not itself a fair use excusing Anthropic's piracy." The order set a trial on the pirated copies and damages. The authors did not claim that any output reaching users infringed, and the judge wrote that if outputs were infringing, they "would have a different case."

The case then settled. Judge Martínez-Olguín granted final approval on July 20, 2026 (ECF 680), of a non-reversionary $1.5 billion fund, covering the 482,460 works on the Works List at an estimated $3,000 or so per work, less costs and fees. She awarded attorney's fees of $101,561,111. Class counsel's September 2, 2026 status report gives the final number of claimed works as 449,731.

The release is limited. Quoting the preliminary approval order, the final order says class members give up claims "broadly related to past AI inputs," but "do not give up claims about past AI outputs, nor claims of any kind about future conduct (on or after August 25, 2025)."

Notices of appeal were filed August 18 and 19, 2026 (ECF 682, 683), and the appeals are 9th Cir. Nos. 26-5391 and 26-5455, according to the district docket we read on November 4, 2026. Class counsel's September 2 report (ECF 688) says no class member or objector filed a timely appeal, and that both notices came from counsel excluded from the fee award and relate solely to that award.

The report adds that an objector mailed a request postmarked August 25, which class counsel says is not a notice of appeal, adding that any notice she does file would be untimely. We did not read the notices or any ruling on that request, so this description is class counsel's.

Class counsel also reported that the parties agree the settlement's Effective Date has passed, because appeals solely about fees do not hold it up. On September 15, 2026, the judge signed a stipulation by class counsel and Anthropic's counsel approving updates to the settlement website (ECF 692). Those updates tell class members the parties agree the Effective Date occurred "assuming that no late-filed appeal is permitted to go forward." The order is not a ruling on the appeals.

On payments, ECF 688 says class members "should expect distributions to start by early November, and no later than November 15, 2026," for works where all rightsholders agree and all payment information is in. Other works can take longer. Counsel estimated the first distribution at about $2,203.56 per claimed work, subject to adjustment and before Anthropic's final $450 million payment plus interest, and each work's amount is shared among its claimants.

Those are counsel's expectations, not a promise of any amount or date. The settlement website gave the same expected date on November 4, 2026, and we read nothing showing whether payments have gone out.

Kadrey v. Meta: what the ruling does and does not say

Meta won summary judgment on training, but the opinion is not the fair-use win it is often reported to be. In Judge Chhabria's words, "this ruling does not stand for the proposition that Meta's use of copyrighted materials to train its language models is lawful. It stands only for the proposition that these plaintiffs made the wrong arguments and failed to develop a record in support of the right one."

He also limited its reach: "This is not a class action, so the ruling only affects the rights of these thirteen authors." And he stated his general view, which goes beyond what he decided: "in many circumstances it will be illegal to copy copyright-protected works to train generative AI models without permission."

The theory he considered most promising is market dilution, meaning a model trained on protected works floods the market with similar works. The plaintiffs "presented no meaningful evidence on market dilution at all," so that factor favored Meta on this record. In a March 25, 2026 order he repeated the point: "the reason Meta won that argument was not that copying protected works to train AI models actually is fair use." It won because counsel had not marshaled meaningful evidence on market harm.

In the orders we read, the case continues on torrenting. The June 2025 order left open the claim that Meta unlawfully distributed works while torrenting them, and the March 2026 order allowed a contributory infringement claim on the same facts. The docket remained open, with its last entry on November 3, 2026, in the copy we read.

Where the rulings disagree

Market dilution. Kadrey treats it as potentially decisive and calls Bartz's schoolchildren comparison an "inapt analogy." Bartz said competing works of that kind are not "the kind of competitive or creative displacement that concerns the Copyright Act."

Pirated sources. Bartz analyzed the pirated library separately from training. Kadrey cited Bartz as contrary authority and held that Meta's downloading "must still be considered in light of its ultimate, highly transformative purpose."

Licensing markets. Bartz and Kadrey both refused to count lost fees for licensing books as AI training data. The Third Circuit went the other way on ROSS's record, finding that the market for licensing headnotes to train AI "is rapidly developing" and that ROSS "usurped" Thomson Reuters's chance to enter it. These are different courts ruling on different facts.

What the Copyright Office and the Justice Department have said

Part 3 of the Copyright Office's report, a pre-publication version released May 9, 2025, says training uses are "likely to be transformative," but that fairness depends on "what works were used, from what source, for what purpose, and with what controls on the outputs." In the Office's view, "making commercial use of vast troves of copyrighted works to produce expressive content that competes with them in existing markets, especially where this is accomplished through illegal access, goes beyond established fair use boundaries."

The Office also said the knowing use of pirated data "should weigh against fair use without being determinative," and that government intervention on licensing "would be premature at this time." As of November 4, 2026, its AI page still listed Part 3 as pre-publication.

The Justice Department filed a statement of interest in In re OpenAI on September 1, 2026. According to the Third Circuit's description, it relied on Bartz to contend that training a large language model is transformative. We did not read the statement itself.

Who Is Liable for AI Output? Users, Vendors and Doe v. GitHub

Output can infringe, and the Copyright Office says outputs that replicate or closely resemble protected works "likely infringe the reproduction right." It deferred "the allocation of potential liability" to a later part of its report, which its AI page did not list as of November 4, 2026.

If you publish AI output

Infringement requires "(1) ownership of a valid copyright, and (2) copying of constituent elements of the work that are original" (Feist, 1991). Under 17 U.S.C. § 106 the owner controls reproduction, adaptation and distribution, and under § 501 anyone who violates those rights is an infringer.

Intent is not an element. The Supreme Court said in Buck v. Jewell-LaSalle Realty Co. (1931), "Intention to infringe is not essential under the act." So a person who publishes output containing someone else's protected expression can be liable even without knowing where it came from.

Statutory damages under § 504(c)(1) run from $750 to $30,000 per work, as the court considers just. Under § 504(c)(2) a court "in its discretion may reduce" the award to as little as $200, but only if the infringer "sustains the burden of proving" that he or she "was not aware and had no reason to believe" the acts were infringing. The court may raise the award to as much as $150,000 if the owner proves willfulness.

Two limits matter. Under §§ 401(d) and 402(d), a proper copyright notice on copies the defendant could access generally defeats an innocent-infringer plea. And § 412 bars statutory damages and attorney's fees for most infringement that began before registration, subject to a three-month window after first publication.

If you build or sell the AI tool

No opinion on this page decides whether a generative AI vendor is liable for what users generate with its tool. The closest general authority is Cox Communications v. Sony Music, decided March 25, 2026. It held that a provider is contributorily liable "only if it intended that the provided service be used for infringement," shown only if it induced the infringement or the service is tailored to it.

A service is tailored to infringement if it is "not capable of 'substantial' or 'commercially significant' noninfringing uses." Mere knowledge that a service will be used to infringe is not enough. Cox was an internet provider, vicarious liability was not before the Court, and we did not read a decision applying Cox to an AI tool.

Separately, the music publishers' case against Anthropic, Concord v. Anthropic, was argued on summary judgment on October 21, 2026, and the court had ordered supplemental briefs on Doe v. GitHub, according to the docket entries we read through November 3.

Doe v. GitHub (9th Cir. 2026) decided a DMCA claim only

On September 16, 2026, the Ninth Circuit (No. 24-7700) affirmed dismissal of a claim under 17 U.S.C. § 1202(b), the DMCA provision on removing copyright management information such as author names and license terms. It held that, as alleged, Copilot and Codex "do not 'remove or alter' copyright management information (CMI) from a copy of an existing protected work but instead create new works that never contained that information."

The limits are in the opinion. The output "may or may not infringe plaintiffs' copyrights," and on substantial similarity the court wrote, "We express no view on whether that similarity would allow plaintiffs to assert a claim for copyright infringement." The separate "input" theory was forfeited.

The court added that if Copilot "functioned like a search engine and produced outputs that were identical to plaintiffs' code but did not contain CMI, then plaintiffs might have a stronger claim." The plaintiffs' breach of contract claims remain in the district court, according to the opinion.

AI-Generated Code and Open Source Licenses

Coding assistants raise these questions directly, because they are trained on public code that often carries license conditions. Open source is copyrighted code offered under a license, not public domain. In Jacobsen v. Katzer (Fed. Cir. 2008), the court treated the terms of an open source license as "enforceable copyright conditions."

We cover this in a dedicated guide: is AI-generated code copyright infringement?. It explains the license families, who is liable when a tool like Copilot suggests infringing code, and why purely AI-written code may not be copyrightable.

Why There Is No AI Copyright Law by State

Copyright is one national system. Under 17 U.S.C. § 301(a), rights "equivalent to any of the exclusive rights" of copyright in a fixed work are governed exclusively by federal law, and no one is entitled to an equivalent right under state law. That means who owns a work, and whether it infringes, does not turn on state AI statutes.

State law still reaches nearby ground, because § 301(b)(3) preserves rights that are not equivalent. A single incident can raise trade-secret misappropriation under state law or breach of a software license in addition to a federal copyright claim.

Some state AI laws touch training data without being copyright laws. California's AB 2013 requires developers of generative AI systems to post documentation about their training data. It has been challenged in X.AI v. Bonta, where a preliminary injunction was denied on March 4, 2026, and Ninth Circuit argument was set for November 18, 2026 (No. 26-1591).

Voice and likeness raise separate questions. See deepfake laws; the Copyright Office's Part 1 report (July 31, 2024) addressed digital replicas. For state AI regulation generally, see AI laws and regulation in the United States.

Pending AI Copyright Cases as of November 2026

Each status below is the date of the docket copy or data we read, not a promise that nothing has changed since.

We read the court opinions and orders this page quotes. We did not read the merits briefing in the pending cases, and we state no view of how any of them will come out.

What Is Settled and What Is Still Moving

What Changes in 2027

Scheduled law. As of November 4, 2026, we found no federal statute on AI and copyright scheduled to take effect in 2027. A search of the copies of Title 17 chapters 1, 3, 4, 5 and 12 posted by the Copyright Office for "artificial intelligence" returned nothing, and the bills below remain bills. That was not a survey of every pending bill.

Bills, as Congress.gov showed them on November 4, 2026. None of these is law, and we do not predict passage.

• S. 2455, the TRAIN Act, would create an administrative subpoena process to help copyright owners learn whether their works trained an AI model. It was referred to the Judiciary Committee on July 24, 2025.
• S. 2367, the AI Accountability and Personal Data Protection Act, would establish a federal tort for use of individuals' data without express, prior consent. It was referred to the Judiciary Committee on July 21, 2025.
• S. 4591, the NO FAKES Act of 2026, was reported with a substitute and placed on the Senate Calendar (No. 446) on June 24, 2026. It concerns digital replicas. We did not read the substitute text.

Dates on court calendars.

• November 18, 2026: Ninth Circuit argument in X.AI v. Bonta on California's AB 2013, as the court's calendar showed on November 4.
• January 6, 2027: status report due in Getty Images v. Stability AI, per the docket we read on November 4.
• January 13, 2027: the Getty stay runs until this date. We did not read why the parties sought it.

Decisions with no date. Any of these could change the answers above: a summary judgment ruling in Concord or In re OpenAI, rulings on the Kadrey torrenting claims, a Ninth Circuit ruling in the two Bartz appeals, which class counsel describes as limited to fees, and a decision in Allen v. Perlmutter. The Copyright Office says the final Part 3 will publish "in the future," and it has not given a date for the later part on output liability.

This is general legal information, not legal advice. It describes United States federal copyright law and does not address your specific facts. Several cited cases are still pending. An attorney licensed in your jurisdiction can advise you on AI-generated or open source material.

Change Log

• November 4, 2026: Rewritten from the primary opinions. Kadrey v. Meta is no longer described as a holding that training is fair use. Added the Bartz settlement approval and class counsel's account of the fee appeals, Copyright Office Part 3, Cox v. Sony, the limits of Thaler's denial of review, a pending case table, and the 2027 outlook.
• November 4, 2026: Corrected for two September 2026 appellate rulings, Thomson Reuters v. ROSS and Doe v. GitHub, replacing the statement that no case had produced an appellate ruling.
• September 23, 2026: Date of the prior version's last review.

Related articles

• Is AI-generated code copyright infringement?
• How AI and copyright differ worldwide
• AI model licensing and open-source attribution
• AI laws and regulation in the United States

Frequently Asked Questions

Can AI-generated content be copyrighted in the United States?

Not if no human authored it. The D.C. Circuit held in Thaler v. Perlmutter (2025) that the Copyright Act requires a human author, and the Copyright Office says purely AI-generated material is not protected. Work made with AI can be protected to the extent a human authored it, which the Office judges case by case.

Did the Supreme Court rule that AI cannot hold a copyright?

No. On March 2, 2026, the Supreme Court denied review in Thaler v. Perlmutter (No. 25-449) without an opinion. That is not a merits ruling. The D.C. Circuit's 2025 decision stands.

Are AI prompts enough to copyright the output?

The Copyright Office says no, given current generally available technology: prompts alone do not provide sufficient human control. Human-written inputs that are perceptible in the output, and creative selection, arrangement or modification of the output, can be protected, judged case by case.

Do I have to disclose AI when I register a copyright?

The Copyright Office's March 2023 guidance says applicants have a duty to disclose AI-generated content in a work submitted for registration and to briefly explain the human author's contributions. AI-generated content that is more than de minimis should be excluded from the claim.

Is it legal to train AI on copyrighted data?

As of November 4, 2026, no appellate opinion we read decides this for generative AI. Bartz v. Anthropic found training fair use but not a permanent pirated library. Kadrey v. Meta ruled for Meta only on its record and said the ruling does not make training lawful. Thomson Reuters v. ROSS rejected fair use for a non-generative tool built to compete with the source.

Is the Anthropic settlement final?

The court granted final approval of the $1.5 billion Bartz v. Anthropic settlement on July 20, 2026. Class counsel reported on September 2, 2026, that no class member or objector filed a timely appeal and that the two appeals filed (9th Cir. Nos. 26-5391, 26-5455, per the district docket) concern only the attorney's fee award. The parties told the court the Effective Date has passed, assuming no late-filed appeal goes forward. Counsel said to expect first payments by November 15, 2026, for works with no allocation dispute and complete payment information; that is an expectation, not a promise. The settlement releases past claims about AI inputs for listed works, not claims about AI outputs or conduct on or after August 25, 2025.

Can I be sued for AI output that copied someone's work?

Yes. Copyright infringement does not require intent, so you can be liable for publishing output that copies protected expression even if you did not know. A court may reduce statutory damages to as little as $200, but only if you prove you were not aware and had no reason to believe you were infringing.

Is an AI company liable when its tool produces infringing output?

That is unresolved. Cox v. Sony (2026) limits contributory liability to providers that induced infringement or sold a service tailored to it, but it involved an internet provider. Doe v. GitHub (9th Cir. 2026) decided only a DMCA claim and took no view on copyright infringement.

Is AI copyright law different in California than in other states?

Copyright itself is the same in every state, because 17 U.S.C. § 301 bars state rights equivalent to copyright. States can still regulate related matters. California's AB 2013 requires generative AI developers to post training data documentation, and trade secret and contract claims arise under state law.

Does the human-authorship rule apply to AI images and text, not just code?

Yes. Thaler involved an image, and the Copyright Act's human-author requirement is not limited to one type of work. The Copyright Office's conclusions on purely AI-generated material apply across categories of works.