AI training and copyright: the important distinctions

Most arguments about AI and copyright go wrong in the first sentence. Someone cites a ruling that training was fair use and concludes AI is legal, or cites a $1.5 billion settlement and concludes it is not. Both treat separate legal questions as one. A court can approve training copies and condemn how they were obtained in the same order. A settlement can pay for past copying while leaving output claims untouched. A lawfully trained model can still produce an output that infringes, or one no one can own.

This analysis separates those questions, using the rulings, statutes and regulator documents themselves. It is legal context, not legal advice. The law differs by country, key rulings are under appeal, and a real decision about a dataset or product needs a qualified lawyer who knows the facts.

Four questions that get blurred together

Each question has its own legal test, evidence and, often, defendant. Answering one does not answer the next.

AI Chronicle

Question 1: was copying for training allowed?

Training involves making copies: collecting, storing, converting and feeding works through training runs. The US Copyright Office concluded that several stages of development “implicate the owners’ exclusive rights”, so the question is whether a defence applies.[1] In the US that is fair use, a four-factor test covering the purpose of the use, the nature of the work, how much was taken and the effect on the market. The report is still a pre-publication version, though the Office expects no “substantive changes”.[2] It gives no blanket answer, and draws the line at substitution and source:

“especially where this is accomplished through illegal access, goes beyond established fair use boundaries.”

— US Copyright Office, Copyright and Artificial Intelligence, Part 3: Generative AI Training (pre-publication, May 2025), on commercial use of vast numbers of works to produce competing expressive content.[1]

Three US rulings show how much the facts matter.

Bartz v. Anthropic (June 2025). Judge Alsup held that using books to train Claude was fair use, calling the purpose transformative, “spectacularly so”.[3] Scanning print books Anthropic had bought was also fair use. But the same order held that “Anthropic had no entitlement to use pirated copies for its central library”, and sent the downloaded pirate-library copies to trial.[3] The authors had not alleged that any output infringed.[3]

That trial never happened. The parties settled for $1.5 billion, and Judge Araceli Martínez-Olguín granted final approval on 20 July 2026.[4] Class members release claims about past inputs, but “they do not give up claims about past AI outputs, nor claims of any kind about future conduct”.[4] First payments are expected on or before 15 November 2026.[5] The settlement is about how books were acquired, not a verdict that training is unlawful.

Kadrey v. Meta (June 2025). Judge Chhabria granted Meta summary judgment on training, but on a thin record. He wrote that “in many circumstances it will be illegal to copy copyright-protected works to train generative AI models without permission”, and said the ruling was narrow:[6]

“these plaintiffs made the wrong arguments and failed to develop a record in support of the right one.”

— Judge Vince Chhabria, Kadrey v. Meta Platforms (N.D. Cal., 25 June 2025).[6]

The argument he thought neglected was market dilution: a model flooding the market with works that compete with the originals.[6] A separate claim, that Meta distributed books while torrenting them, was expanded in March 2026 and was still in discovery in September.[7]

Thomson Reuters v. Ross (affirmed September 2026). Ross trained a legal search tool on Westlaw headnotes to compete with Westlaw. Judge Bibas held the use not transformative because Ross “took the headnotes to make it easier to develop a competing legal research tool”.[8] On 29 September 2026 the Third Circuit affirmed, in the first federal appellate ruling on fair use in AI training, calling it “no more than an ordinary copyright case” and noting Ross's system “was not a generative AI”.[9]

Together, the US cases say neither “legal” nor “illegal”. The answer depends on what was copied, from where, for what product, and with what effect on the originals' market.

Question 2: does the model, or its output, contain the work?

The second question is about what sits inside the model and what comes out. A model retaining training examples is memorisation; reproducing them in outputs is often called regurgitation. Our explainers on how AI image generators work and how AI video generators work cover the mechanics.

The Copyright Office's view is that when a model can output near-copies of a training example without being fed that expression, it “must exist in some form in the model's weights”, so copying the weights may itself implicate the reproduction right.[1]

Two European courts reached opposite-looking results on this, and both are under appeal.

GEMA v. OpenAI (Munich, November 2025). The Munich Regional Court I largely upheld claims by GEMA, the German music collecting society, over nine song lyrics. According to the court's press release, it found the lyrics were reproducibly contained in two of OpenAI's models and that both the memorisation and the outputs infringed.[10] It held that the text and data mining exception covers copies made to assemble a training corpus, but not reproduction of works inside the model.[10] OpenAI appealed to the Higher Regional Court of Munich in December 2025 (case 6 U 3662/25 e), according to published reports; we found no appellate ruling as of October 2026.[11]

Getty Images v. Stability AI (London, November 2025). Getty dropped its training claim, having no evidence training happened in the UK, and its output claim after Stability blocked the prompts in evidence.[12] What remained was whether the model was an imported “infringing copy”. Mrs Justice Joanna Smith held it was not, because Stable Diffusion “does not store or reproduce any Copyright Works (and has never done so)”.[12] Getty won only a narrow trade mark claim over watermarks, findings “historic and extremely limited in scope”.[12] Getty has permission to appeal; the UK government noted in March 2026 that the judgment is “currently subject to appeal”.[13][14]

The results are less contradictory than they look. Munich had evidence of memorised, near-verbatim text; London was told the model stored no images. Whether a model memorised a work is a question of fact. An EU-wide answer may come from Like Company v Google (C-250/25), where a Hungarian court asked the Court of Justice whether training is reproduction, whether Article 4 covers it, and whether chatbot answers displaying press content are the provider's act.[15] The hearing was on 10 March 2026; we found no opinion or judgment by mid-October 2026.[15]

Question 3: who owns what the AI produces?

Ownership is separate from infringement. An output can be unowned and harmless, owned and infringing, or anything between.

The US Copyright Office concluded in January 2025 that AI-assisted works can be protected where a human determines the expressive elements, including through creative selection, arrangement or modification. On prompting: “Prompts alone, however, at this stage are unlikely to satisfy those requirements.”[16] On 2 March 2026 the Supreme Court denied Stephen Thaler's petition, leaving in place the ruling that refused registration of an image he credited to his AI system.[17] Our guide to AI for marketing covers what that means for campaign assets.

The UK is unusual. Section 9(3) of its copyright act treats the author of a “computer-generated” work as the person who made the arrangements for its creation. The government's March 2026 report says that provision contains “a legal contradiction” with the originality test, and proposes that, absent evidence of its value, “it should be removed”.[14] Japan's Agency for Cultural Affairs takes a case-by-case view: detailed instructions that specify creative expression, iterative correction and human edits can count; long prompts that merely suggest an idea do not.[18] It also notes that a copyrightable output can still infringe an existing work.[18]

Question 4: whose law applies?

Copyright is territorial. Where the copying, training and use happened can decide the case, as Getty's abandoned training claim showed.[12] The four systems below give different starting points.

Sources: the statutes, court records and official reports cited. Simplified.

The EU. Article 4 of the 2019 Directive on copyright in the Digital Single Market allows reproductions of “lawfully accessible” works for text and data mining, on condition that use has not been:[19]

“expressly reserved by their rightholders in an appropriate manner, such as machine-readable means”

— Directive (EU) 2019/790, Article 4(3).[19]

The AI Act builds on this. Article 53(1)(c) requires every provider of a general-purpose AI model to have a copyright policy that identifies and complies with those reservations; Article 53(1)(d) requires a public training-content summary on the Commission's template.[20] The template is mandatory; the duty applies from 2 August 2025, with earlier models given until 2 August 2027.[23] As our EU AI Act explainer sets out, openly released models are covered too.

The General-Purpose AI Code of Practice turns this into commitments for signatories: crawlers that follow robots.txt, no circumventing paywalls, excluding sites recognised as persistently infringing, safeguards against outputs that reproduce training content, and a complaints contact for rights holders.[24] Adherence “does not constitute conclusive evidence of compliance”.[24] Our guide to open versus closed AI models explains why openness does not remove these duties.

The UK. Section 29A allows computational-analysis copies only “for the sole purpose of research for a non-commercial purpose”.[21] A December 2024 consultation proposed a broad exception with opt-out. The March 2026 report required by the Data (Use and Access) Act 2025 dropped it:

“a broad copyright exception with opt-out is no longer the government’s preferred way forward.”

— UK government, Report on Copyright and Artificial Intelligence (March 2026).[14]

It says it will not reform the law “until we are confident” the changes meet its objectives.[14]

Japan. Article 30-4 permits exploiting a work where the purpose is not to “personally enjoy or cause another person to enjoy” its thoughts or sentiments, as in data analysis, unless the use “would unreasonably prejudice the interests of the copyright owner”.[22] Non-binding guidance from the Agency for Cultural Affairs says the exception fails where training also aims to output the works' expression, such as fine-tuning to imitate them; treats robots.txt blocks on databases offered for licensing as relevant to the proviso; and warns against known pirate sources.[18]

Licensing and opt-outs: what they settle and what they don't

Licences and opt-outs answer the input question, and their force depends on the jurisdiction. An EU opt-out matters because Article 4 is conditional on it.[19] The US has no statutory opt-out, so robots.txt there is a technical signal, not a condition of an exception. Licensing grows where fair use risk is real: the Copyright Office found it “fast emerging in certain sectors” but inconsistent.[1] The Third Circuit counted harm to a potential market for licensing training data against Ross.[9]

They do not answer questions 2 and 3. A licensed dataset can still yield a model that regurgitates a work. A respected opt-out says nothing about who owns the output. Records of inputs and output origins matter for both; our explainer on content credentials covers one way to record provenance.

A worked example: one studio, four questions

This is a hypothetical example. It describes no real company, licence or dispute. The readings are illustrative, not legal conclusions, and a real assessment would depend on contract terms and facts this sketch leaves out.

A 15-person design studio in Manchester fine-tunes an openly released image model on two sets of material: past client work it owns, and a stock library it licenses for use in client projects. It then uses the fine-tuned model to create a mascot for a client in Germany.

With modest fine-tuning compute, the studio is probably not a new model provider under the AI Act, so the Article 53 duties stay with the original provider.[20] And if the stock licence is silent on training, the first problem is contractual, before copyright is reached.

Questions readers commonly ask

Is it legal to train AI on copyrighted work?

It depends on the country and the facts. US courts have found some training fair use and some not, depending on source, purpose and market effect.[3][6][9] The EU permits mining lawfully accessible works unless rights holders opt out; the UK only for non-commercial research; Japan for most non-enjoyment uses.[19][21][22]

Did the Anthropic settlement show that training is illegal?

No. The same case had held training itself fair use; the settled claims concerned pirated library copies, and the release excludes outputs and future conduct.[3][4]

If I block AI crawlers with robots.txt, am I protected?

In the EU, a machine-readable reservation conditions the Article 4 exception, and Code signatories commit to follow robots.txt.[19][24] Elsewhere it is a signal, not a legal bar, and it does not reach copies already collected.

Can I copyright something I made with AI?

In the US, only the human-authored parts: your own material, creative selection and arrangement, or creative modification. Prompts alone are unlikely to be enough.[16] The UK's computer-generated works rule is uncertain and may be removed.[14]

Does using an AI tool expose my business to infringement claims?

Possibly, through outputs. Publishing an output substantially similar to a protected work can infringe whoever trained the model. Check commercial outputs, especially for recognisable characters, logos and text.[18]

The next step

Records of where content came from are becoming central to all four questions. Next, read Content credentials explained: what provenance can show.

About these sources

All pages were accessed on 16 October 2026. Court decisions and statutes are cited for what they state, not as legal advice. Where a ruling is under appeal, that is noted. The studio in the worked example is hypothetical.

Related reading

• Who is accountable when an AI agent acts?
• How to spot AI-generated content: signals and limits