What recent court cases teach businesses about training data

Two questions that courts consider separately

Before turning to the cases, it is important to understand that courts in the EU and the US are actually dealing with two different questions that are often mixed together in public discussions.

First: where the training data came from and how it was used in the training process. The key concepts here are lawful access, licensing, and the text and data mining (TDM) exception.

Second: what the model reproduces after training. This is where memorisation comes in: a phenomenon in which a model has “memorised” protected content to such an extent that it can reproduce it almost verbatim in response to a simple prompt.

These two questions can be decided differently. The training may be lawful, while the output constitutes an infringement. Or, conversely, the training may be unlawful, not because of memorisation, but because of the way the data was obtained. The cases below illustrate both scenarios.

GEMA v OpenAI: when TDM does not protect against memorisation

GEMA v OpenAI

Landgericht München I (LG München I) — case 42 O 14139/24 — decision of 11 November 2025 — first instance, under appeal

The German collecting society GEMA filed a lawsuit against OpenAI over the use of lyrics from copyrighted songs to train GPT-4 and GPT-4o. Nine works were involved, including “Atemlos” by Christina Bach and “Über den Wolken” by Reinhard Mey. GEMA argued that the lyrics were not merely used for analysis: they had been memorised in the model’s parameters and could be reproduced almost verbatim in response to a simple prompt.

The court found that memorisation does constitute reproduction within the meaning of copyright law. However, the text and data mining (TDM) exception relied upon by OpenAI applies only to the first stage: collecting and analysing data. It does not extend to the subsequent permanent incorporation of protected text into the model’s weights and its subsequent reproduction in the output.

OpenAI argued that the models do not store specific texts but instead learn only statistical relationships. The court rejected this argument, finding it decisive that ChatGPT reproduced lines from the songs almost verbatim in response to a simple prompt. The fact of reproduction was sufficient.

The decision is under appeal. The question of how higher courts, and potentially the CJEU, will interpret the relationship between the TDM exception and memorisation remains open. But there is already a first point of reference.

→ Lawful training does not equal permission to memorise. If a model can reproduce protected text in response to a simple prompt, this is a separate legal issue, regardless of whether the data was obtained lawfully.

Bartz v Anthropic: when the infringement was not where it seemed

Bartz et al. v. Anthropic PBC

U.S. District Court, Northern District of California — No. C24-05417 — June 2025 decision (fair use), settlement in September 2025, final approval on 20 July 2026

Three authors — Andrea Bartz, Charles Graeber, and Kirk Wallace Johnson — filed a lawsuit against Anthropic over the use of their books to train Claude. Anthropic obtained the books in two ways: it purchased physical copies, scanned them, and stored them digitally; and it downloaded them from the pirate libraries LibGen and PiLiMi.

Judge William Alsup clearly distinguished between these two methods. Using lawfully purchased books to train AI constituted transformative use, and therefore fair use. Downloading books from pirated sources, however, constituted infringement, and fair use did not apply. The court put it directly: “a person who copies a textbook from a pirate website has already broken the law — period.”

Anthropic prevailed on the issue of training the model on the books themselves. But it lost on the way those books had been obtained. In September 2025, the parties settled the case: Anthropic agreed to pay $1.5 billion, approximately $3,000 for each of the nearly 500,000 pirated books. The settlement received final approval on 20 July 2026.

The figure is telling: Anthropic was not paying for training the model on someone else’s books. It was paying for how those books had been obtained. The distinction is significant — and relevant to other companies building AI products.

→ The source of the data is critical, regardless of the purpose for which it is used. Fair use may cover AI training, but it does not protect unlawful acquisition of materials. The provenance of every dataset is a legal issue, not merely a technical one.

The three-gate framework: where the risk arises

To systematise the lessons from both cases, it is useful to think about three separate points where a legal issue with training data may arise.

The practical significance of this framework is that a problem can arise independently at any of the three levels. A company may have lawful access to the data (Gate 1), use it in a way covered by TDM (Gate 2), and still face a lawsuit because the model reproduces protected text in its output (Gate 3). This is exactly what the GEMA case demonstrated.

Opt-out and rights in the EU: what both sides need to do

Alongside the court cases, the EU has a regulatory framework that is often underestimated. Article 4(3) of Directive 2019/790 (CDSM) allows rightsholders to reserve their rights for text and data mining purposes through a machine-readable opt-out, for example, via a robots.txt file or metadata.

The AI Act adds another layer: Article 53 requires providers of GPAI models to identify and respect machine-readable rights reservations, even if the model was trained outside the EU but is placed on the EU market. This means that an opt-out properly implemented by a rightsholder must be respected technically.

Two practical takeaways. For rightsholders: an opt-out is not a guarantee against all risks, but it significantly strengthens their position in a dispute. It should be implemented in a machine-readable format and kept up to date. For AI developers: checking for opt-out signals before training a model is an obligation, not a recommendation. The case Like Company v Google (CJEU C-250/25) has raised the question of what exactly qualifies as a valid machine-readable opt-out. The case is still pending.

What this means for businesses

Both cases illustrate one thing: the issue of training data cannot be left for later or delegated solely to engineers. Below are specific measures that make sense regardless of whether a company develops AI systems or simply uses third-party models.

For AI developers and providers

• Document the provenance of every dataset. Where did the data come from, under what terms was it obtained, what licences apply, and are there any opt-outs from rightsholders? This is not just legal protection. It can also be important evidence in a future dispute or due diligence process.
• Respect machine-readable opt-outs. Technical checks before training a model are an obligation under Article 53 of the AI Act for GPAI models placed on the EU market.
• Implement technical safeguards against memorisation. Deduplicate data before training, use output filters, and test for the reproduction of protected content. GEMA v OpenAI showed that courts look at the fact of reproduction, not the technical architecture behind it.
• Have a procedure for responding to rightsholder complaints. Notice-and-takedown procedures for AI outputs are not yet standardised, but having no procedure at all can weaken a company’s position in a dispute.

For companies using third-party AI systems

• Review the provider’s terms regarding training data. What guarantees does the provider offer? Is there an indemnification clause covering third-party claims related to training data? This is already becoming a standard part of AI vendor due diligence.
• Do not upload confidential materials to AI systems without reviewing the terms. If the provider uses input data for further model training, this may affect the rights of third parties whose materials you upload.
• Monitor developments in the cases. Bartz v Anthropic is an important precedent, but it will not be the last. Similar lawsuits against OpenAI, Google, and Meta are at different stages. The legal framework is being shaped right now.

Instead of a conclusion

Bartz v Anthropic showed that training on lawfully obtained data can qualify as fair use, but the way the data was obtained can make all the difference. GEMA v OpenAI showed that the TDM exception does not cover memorisation, and that courts look at the output, not just the training process.

Both cases point to one practical rule: license the data rather than simply taking it, and control the output. Companies that document data provenance and implement output safeguards now will be in a significantly stronger position, both in court and during due diligence.