Loading
AI copyright is the fast-moving body of US and EU law governing three separate questions that get conflated under one label: whether training a model on copyrighted material is fair use, whether AI output that competes commercially with the works it was trained on infringes those works, and whether output produced by AI without meaningful human authorship can be copyrighted at all. Three landmark rulings answered each question differently between 2025 and 2026, training on lawfully acquired books was found fair use in Bartz v. Anthropic, a directly competing AI research tool was found not fair use in Thomson Reuters v. Ross, and the US Supreme Court left standing a rule that purely AI-generated works cannot be copyrighted in Thaler v. Perlmutter. None of this is fully settled: the Anthropic case ended in a record settlement rather than an appellate ruling, and the Thomson Reuters case is still awaiting a decision from the Third Circuit as of mid-2026. For governance teams, the practical stakes, vendor indemnities, data provenance, and output-review processes, matter long before any of these doctrines fully stabilize.
Run the free AI Health CheckAI Copyright, the body of law and contract concerning rights in AI training data, AI-generated outputs, and the use of copyrighted material by AI systems.
AI copyright is contested across three distinct questions. First, whether AI training on copyrighted material is fair use: Judge Alsup's Bartz v Anthropic ruling (N.D. Cal., June 23, 2025) found training on lawfully acquired books was transformative fair use, though retaining pirated copies was not; a separate $1.5 billion settlement of the piracy claims received final court approval in July 2026 but did not test or overturn the fair-use finding on appeal. Second, whether AI outputs that compete commercially with their training sources are fair use: Judge Bibas's Thomson Reuters v Ross ruling (D. Del., February 11, 2025) said no, but that ruling is not yet final, Ross's appeal was argued before the Third Circuit in June 2026 with a decision still pending. Third, whether AI-generated work is copyrightable at all, the US Supreme Court denied certiorari in Thaler v Perlmutter (March 2026), settling that purely AI-generated work is not.
Source: Bartz v Anthropic (N.D. Cal. 2025); Thomson Reuters v Ross (D. Del. 2025, appeal pending, 3d Cir.); Thaler v Perlmutter; EU AI Act Article 53
'AI copyright' is not one legal test but shorthand for at least three distinct disputes that courts are resolving separately, often with opposite outcomes. The first question is upstream: is copying copyrighted works to train a model infringement, or fair use because the model does not reproduce the works themselves? The second is downstream: even if training is fair use, is it still infringement when the resulting product competes commercially with the market for the original works, or with the market for licensing them as AI training data? The third concerns the output side entirely: can a work created by AI, without meaningful human creative input, be copyrighted by anyone at all?
Because these are separate questions, a ruling that favors an AI developer on one of them says nothing about how the same court, or a different one, will rule on another. Anthropic won on the training question and still faced billion-dollar liability on the piracy question, in the same case. Governance teams that treat 'AI copyright' as a single resolved question risk missing that each thread is still moving, on different timelines, in different courts and jurisdictions.
Bartz v. Anthropic, training on legally acquired copies is fair use
In June 2025, Judge William Alsup (N.D. Cal.) ruled that training Claude on lawfully acquired books was 'quintessentially transformative' fair use, but that copying and permanently retaining pirated books from shadow libraries to build an internal training library was not. The piracy claims were resolved through a $1.5 billion class settlement covering roughly 480,000 works, which received final court approval on July 20, 2026, the largest copyright class-action recovery in US history. The fair-use holding on legitimately acquired training data was never tested on appeal because the case settled.
Thomson Reuters v. Ross Intelligence, a directly competing product was not fair use
In February 2025, Judge Stephanos Bibas (D. Del.) rejected Ross Intelligence's fair-use defense and found it liable for copying 2,243 Thomson Reuters Westlaw headnotes to train a competing legal-research tool, reasoning that a commercial product built to substitute for Westlaw was not transformative and harmed Thomson Reuters' market. Ross's interlocutory appeal reached the Third Circuit, which heard oral argument on June 11, 2026; no decision had issued as of this writing, making this the first pending federal appellate test of AI-training fair use.
Thaler v. Perlmutter, AI cannot be a copyright author
The US Supreme Court denied certiorari on March 2, 2026, leaving in place the D.C. Circuit's ruling that the Copyright Act requires human authorship, so a work generated entirely by AI, with no human creative contribution, cannot be copyrighted. Stephen Thaler's attempt to register an image as authored solely by his 'Creativity Machine' was refused by the Copyright Office and every reviewing court. Works combining AI tools with meaningful human creative authorship remain copyrightable.
Getty Images v. Stability AI (UK)
The UK High Court ruled in November 2025 that Stability AI did not commit secondary copyright infringement, but the court never reached the core question of whether training on copyrighted images is infringement, because Getty conceded the training itself did not take place in the UK. UK courts have not yet issued a first-instance ruling on that core training question.
New York Times v. OpenAI and Microsoft
Filed in December 2023 and still active in the Southern District of New York, this case alleges millions of Times articles were used without permission to train ChatGPT. In June 2026 the Times narrowed its complaint after a Supreme Court ruling raised the bar for holding platforms liable for users' infringement, and refocused claims on Microsoft's alleged role in supplying the computing infrastructure used for training. No fair-use ruling has been issued.
EU AI Act Article 53, transparency, not a liability test
Article 53(1)(d) of the EU AI Act requires providers of general-purpose AI models to publish a 'sufficiently detailed summary' of training content, using a template the European Commission adopted for models placed on the market after August 2, 2025 (with a transition to August 2, 2027 for earlier models). It does not decide whether training infringes copyright, but is designed to give rights holders the information needed to identify and enforce their rights.
None of this is settled enough to write into policy as a fixed rule. The one holding every side agrees on, human authorship is required for copyright, is stable because the Supreme Court declined to disturb it. Everything else, whether training on copyrighted material is fair use, and when a competing output crosses the line, is still being litigated in parallel cases with different facts and different circuits, and a ruling favorable to one AI developer does not bind another.
Practical governance responses that do not depend on how the litigation ends: obtain contractual warranties and indemnities from AI vendors about training-data provenance; review AI-generated outputs for close similarity to identifiable copyrighted works before commercial use; keep records of what data went into any model your organization fine-tunes or deploys; and track the Article 53 training-content summaries now being published by general-purpose AI providers as a public source of provenance information.
Is training an AI model on copyrighted books or articles fair use in the US?
It depends on how the works were obtained and what the resulting product does. In Bartz v. Anthropic, a federal judge held in June 2025 that training on lawfully acquired (purchased or licensed) books was 'quintessentially transformative' fair use, but that retaining pirated copies in a permanent internal library was not. In Thomson Reuters v. Ross, a different court held in February 2025 that training was not fair use where the resulting product competed directly with the market for the original works. There is no single settled answer: the only US appellate court to hear argument on the question, the Third Circuit, had not ruled as of July 2026.
Did the Anthropic $1.5 billion settlement decide whether AI training on copyrighted books is legal?
No. The settlement, which received final court approval on July 20, 2026, resolved claims tied to Anthropic's use of pirated books downloaded from shadow libraries such as LibGen. It did not overturn, and was never tested on appeal against, the judge's separate 2025 finding that training on lawfully acquired books was fair use. That fair-use finding stands only as an unappealed district-court ruling.
Can I copyright an image, story, or song that AI generated for me?
Only the parts you meaningfully created yourself. By denying certiorari in Thaler v. Perlmutter in March 2026, the US Supreme Court left in place a rule that a work with no human creative authorship cannot be copyrighted. The US Copyright Office has said works combining AI-generated elements with substantial human creative input, selection, arrangement, or editing can be copyrighted, but protection covers only the human-authored contribution, not the AI-generated material itself.
Is the Thomson Reuters v. Ross Intelligence ruling final?
Not yet. Judge Bibas's February 2025 decision rejecting Ross's fair-use defense is on interlocutory appeal to the Third Circuit, which heard oral argument on June 11, 2026. As of mid-2026 no appellate decision has issued, so the ruling could still be affirmed, reversed, or narrowed, and it remains the first pending federal appellate test of AI-training fair use.
Does the EU require AI companies to disclose what copyrighted material they used for training?
Yes, in a limited way. Article 53(1)(d) of the EU AI Act requires providers of general-purpose AI models to publish a public summary of training content, using a template the European Commission adopted for models placed on the market after August 2, 2025 (with a transition period to August 2, 2027 for earlier models). This is a transparency obligation intended to help rights holders identify potential uses of their work, not a ruling on whether the underlying training was lawful.
Why did Anthropic win on training but lose on piracy in the same case?
The court treated them as two separate acts. Training a model on a lawfully acquired book was found transformative because the model does not reproduce the book to end users. But downloading and permanently storing pirated copies to build an internal library was found to be a separate, non-transformative act of infringement, regardless of what the copies were later used for, which is why Anthropic still faced billions in piracy liability despite winning the training question.
Last reviewed July 2026
This page is general information about What Is AI Copyright?, not legal, regulatory, or professional advice, and does not capture every nuance or exception. Requirements change and can be fact-specific. Always verify against primary sources and your own qualified legal counsel before relying on it.