Why Germany is becoming the AI lawsuit capital of Europe

The November 2024 decision of the Munich Regional Court in GEMA v. OpenAI received less English-language press than it deserved. It was the first European court of any consequence to hold that a generative AI operator was liable for the reproduction of copyrighted material in its outputs — and, more importantly, the first to limit the EU's text-and-data-mining exception in a way that other European courts will inevitably follow.

That ruling matters far beyond German rights-collection law. It tells every AI lab operating in Europe that the legal architecture they have spent the last three years designing around — Articles 3 and 4 of the 2019 DSM Directive — does not extend to memorized output. And it suggests that European AI litigation, long overshadowed by the U.S. publisher cases, is going to be more, not less, demanding for defendants.

The TDM exception and what it actually says

Article 4 of the DSM Directive permits text-and-data mining of lawfully accessible works for any purpose, subject to a rights holder's express opt-out. AI labs have read this — reasonably — as a near-blanket permission to train models on internet-accessible material so long as they respect robots.txt-style signals. The TDM defense has been the European equivalent of fair use: the all-purpose answer to copyright complaints against training.

GEMA, the German collecting society for music composers and lyricists, tested this defense with surgical precision. It did not sue OpenAI for training. It sued for output. GEMA prompted ChatGPT to produce the lyrics to nine well-known German songs in its repertoire — works by Herbert Grönemeyer, Reinhard Mey, and others — and obtained near-verbatim reproductions. The Landgericht München I, in a declaratory judgment, held the reproductions infringed.

The court's reasoning on Article 4 is the sentence that mattered. The TDM exception, the court held, exists to enable the temporary processing of works for analytical purposes. It does not extend to a state of affairs in which a copyrighted work has been encoded into a model in a manner that allows near-verbatim retrieval. Once a work can be regurgitated, the use is no longer mining; it is reproduction.

Why this is harder than fair use

The U.S. fair-use doctrine is multi-factor and fact-specific. A defendant who loses on training-data fair use can imagine a more careful corpus, a more transformative downstream product, a smaller appropriation, or a less commercial use producing a different outcome. The TDM exception is none of those things. It is a categorical rule — yes or no — bounded by an opt-out mechanism and a definitional question about what counts as "data mining."

If, as Munich now holds, memorized output is by definition outside the exception, the line between training and infringement is a technical one: it depends on the model's tendency to memorize. That makes the question almost entirely empirical. Plaintiffs will run regurgitation prompts; defendants will respond with deduplication, output filtering, and lower-temperature decoding. The doctrinal fight has effectively been replaced by an evidentiary one.

Why Munich, and why now

The Munich Regional Court has long been one of Europe's preferred IP forums. Its civil chambers have specialist judges, fast-tracked procedure, and a body of case law on infringement and injunctive relief that is unusually plaintiff-friendly. German collecting societies — GEMA for music, VG Wort for text — are also institutionally well suited to AI litigation. They have standing without any individual rights holder needing to step forward, and they have decades of experience converting court victories into licensing revenue.

The legal architecture is also more favorable to plaintiffs in Germany than in many other European jurisdictions. Disclosure obligations under §§ 101 and 101a UrhG allow rights holders to compel information about infringing chains; preliminary injunctions are routinely available for IP claims; and German courts have repeatedly extended injunctive relief beyond Germany's borders for cross-border digital infringement. None of that is true to the same degree in France, Italy, or Spain.

Who's next

Three observations for what follows. First, the German publishing collectives — VG Wort and the Börsenverein — have been studying AI litigation for two years. Expect a parallel text-and-author action in Germany along GEMA's template. Second, French SACEM and Italian SIAE have similar structural advantages and similar political will. The TDM-output limitation, if a French or Italian court adopts the Munich reasoning, will be very hard to dislodge across the EU.

Third, the appellate posture matters. OpenAI has appealed; the Oberlandesgericht München and ultimately the Bundesgerichtshof will get the case. A reversal is possible — the TDM jurisprudence is thin enough that a clean-slate appellate opinion could go either way. But the Munich Regional Court's reasoning has the doctrinal pull that first-mover decisions on novel legal questions usually do, and the burden will be on the appellant.

The shorter version: European AI training law just got narrower than the major labs assumed it would. Operators that are betting their European deployments on Article 4 alone are betting on something Munich already rejected.