Corsino San Miguel

On 11 November, the 42nd Civil Chamber of the Munich Regional Court delivered the first European judgment to hold an AI developer directly liable for both training and outputs involving copyrighted works. The case – GEMA v OpenAI – concerns ChatGPT’s reproduction of German song lyrics and whether those lyrics were unlawfully embedded in the models GPT-4 and GPT-4o.

But while headlines describe a “landmark victory” for rights-holders, it is crucial to remember: this is a regional, first-instance ruling, already being appealed, and its reasoning is far from settled. Its ultimate fate may lie not in Munich Higher Regional Court but in Luxembourg, before the Court of Justice of the European Union (CJEU).

Still, even as a provisional step, it signals a dramatic shift in Europe’s direction of travel – and stands in marked contrast to the UK approach as I noted in my Scottish Legal News analysis of the Getty ruling, where the High Court declined to equate learning with copying.

The core issue: memorisation as reproduction

GEMA – Germany’s collecting society representing over 100,000 music creators – argued that OpenAI had used lyrics from nine well-known German songs during training, and that ChatGPT could reproduce those lyrics almost verbatim in response to simple prompts. OpenAI denied copying, insisting that large models do not “store” text but learn statistical relationships.

The Munich court sided with GEMA, finding that:

• the ability of ChatGPT to reproduce substantial lyrics constituted memorisation, not pattern analysis;
• this memorisation amounted to a reproduction under Article 2 of the Directive 2001/29/EC (InfoSoc Directive) and §16 UrhG (German Copyright Act); and
• the verbatim outputs were additional acts of reproduction and communication to the public.

Importantly, the court accepted scientific literature showing that “reproducible determinations” of training data can exist within model parameters – a point of high technical controversy and one likely to be tested on appeal.

Why the TDM exception did not save OpenAI

OpenAI relied on the EU text and data-mining (“TDM”) exceptions in Articles 3 and 4 of Directive (EU) 2019/790 (the “CDSM Directive”, implemented in Sections 60d and 44b UrhG), arguing that such uses were, in any event, covered by those exceptions. The court rejected that defence, holding that:

• TDM covers extraction of patterns, syntax, and statistics;
• memorisation of entire works exceeds this scope; and
• permanent embodiment of lyrics interferes with authors’ economic rights.

This interpretation drastically narrows the TDM exception and, if upheld, would place Europe on a markedly more restrictive path than both the UK and the US.

A European approach diverging from UK and US courts

The Munich decision stands in sharp contrast with Getty v Stability AI.

The English High Court found no evidence that Stable Diffusion’s model weights contained reproductions of Getty photographs. The judge refused to expand the definition of an “infringing copy” without statutory authority. Training that occurred outside the UK meant the core infringement claims could not even proceed.

U.S. courts continue to treat AI training through the lens of fair use. In Bartz v Anthropic, Judge Alsup emphasised that training can amount to fair use where it is transformative and non-substitutive, characterising the process as one that extracts statistical patterns rather than reproducing expression. Crucially, he drew a line between Anthropic’s creation of a vast “central library” of pirated books – a non-transformative use that ultimately led to a $1.5bn settlement for the plaintiff authors – and the subsequent training step, which he considered “highly transformative” and therefore protected by fair use.

A similar approach emerged in Kadrey v. Meta Platforms, Inc, where the court found that training large-language models on lawfully-acquired copyrighted works plausibly falls within §107 fair use, noting the use was “exceedingly transformative” at the training stage. Crucially, U.S. courts separate training from output: while training may qualify as fair use, any downstream verbatim reproduction could still give rise to liability.

The Munich ruling strengthens a distinctly EU-style approach: copyright as an economic property right requiring prior authorisation, with little tolerance for unlicensed training.
But again, this is only one chamber of one regional court. Its reasoning has not yet been endorsed by the higher courts or the CJEU.

The responsibility question – user or provider?

OpenAI argued that outputs arise from user prompts and therefore liability, if any, rests with the user. The court disagreed: selecting training data, designing the architecture, and determining model parameters make OpenAI responsible for both memorisation and output. This finding, if upheld, could have industry-wide consequences, imposing affirmative duties on AI providers to prevent infringing outputs – even when caused by user prompts.

Business implications – with the appeal looming

Although not yet binding, the ruling hints at a future where:

• blanket licensing could become the norm for AI training in Europe;
• developers must document training data to comply with Article 53 AI Act;
• memorisation triggers liability regardless of where training occurred;
• and companies may need to retrain or filter models to avoid infringement.

But this future is not guaranteed. The appeal could reshape or even overturn the decision. A reference to the CJEU remains entirely possible, especially given the EU-wide implications for LLM developers, rights-holders, and the AI Act’s enforcement.

Why this matters for the UK

For Scottish legal practitioners, the Munich ruling is strategically important:

• UK courts (as seen in Getty) treat memorisation very differently.
• While the UK recognises a narrow TDM exception under section 29A CDPA, it applies only to non-commercial research
• The UK retains flexibility but faces trade and data-transfer constraints with the US and EU.
• Divergent regimes complicate compliance for Scottish law firms advising multinational clients.

In short, AI companies may soon face three incompatible copyright regimes: the UK, the EU, and the US – each defining “copying” differently.

A landmark – but not a settlement

Europe may have fired the opening shot, but the battle is far from over. The Munich decision is not final, not EU-wide, and not endorsed by higher courts. The appeal could reshape every core finding – from memorisation to TDM to developer liability. The CJEU may ultimately be asked to decide whether mathematical parameters can constitute “copies” in the digital age.

Still, even at this early stage, the ruling is a powerful signal: Europe is willing to treat AI training as reproduction unless licensed. By contrast, the UK’s approach in Getty reflects judicial restraint – insisting that Parliament, not the courts, must determine the boundaries of authorship in the algorithmic era. The legal landscape is diverging. And for now, one truth holds: the future of AI copyright will be written not by any single court, but by appellate judges and legislators on both sides of the Channel.

Dr Corsino San Miguel is a member of the AI Research Group and the Public Sector AI Task Force at the Scottish Government Legal Directorate. The views expressed here are personal.