AI Regulation Tracker / Courts and copyright

A German Court Says Memorized Lyrics Inside an LLM Are Copyright Infringement

On November 11, 2025, the Regional Court of Munich I ruled that OpenAI infringed German copyright when its models memorized and reproduced protected song lyrics without a license. The court rejected the text-and-data-mining defense. For US attorneys advising on AI, this is the first major European judgment squarely on what a model stores, not just how it was trained.

By Anthony Guerriero · The Leveraged Years Regulatory Research · Published July 21, 2026 · Last updated July 21, 2026

The short version

On November 11, 2025, the Regional Court of Munich I (Landgericht Munchen I), case 42 O 14139/24, held that OpenAI infringed German copyright in a suit brought by GEMA, the German music rights society, over nine well-known German songs. The court found that when a large language model memorizes protected lyrics and reproduces them in output from simple prompts, that is a reproduction under copyright law, and that it is not saved by the text-and-data-mining exception, which covers only the analytical training phase. It largely granted GEMA's claims for injunctive relief, disclosure, and damages. Two cautions on how to read it. First, it is a first-instance judgment and OpenAI has said it will appeal, so it is not final. Second, it is German law, not US law. But it is the clearest signal yet that a court will look at what a model has stored inside it, and treat regurgitation of training content as infringement. US attorneys should factor that theory into diligence, licensing, and output-filtering advice now.

Regulatory Briefing

Instrument: Civil judgment finding copyright infringement (injunctive relief, disclosure, and damages granted); first instance
Authority: Regional Court of Munich I (Landgericht Munchen I), 42nd Civil Chamber
Case: GEMA v. OpenAI, case No. 42 O 14139/24
Jurisdiction: Germany; persuasive across the EU on the same TDM and reproduction questions
Bindingness: Binding as a first-instance judgment between the parties, but not final. OpenAI has announced an appeal to the Higher Regional Court of Munich, and the judgment may be reviewed on appeal.
What Changed: A court held that memorization and reproduction of protected works inside and out of an LLM is infringement, and is not covered by the text-and-data-mining exception.
Who This Affects: AI developers, licensors and rightsholders, and the US and EU attorneys advising on training data, licensing, and output filtering.
Judgment Date: November 11, 2025
Primary Source: Regional Court of Munich I, judgment of Nov. 11, 2025 (42 O 14139/24), as reported by CMS, cms.law
Last Verified: July 21, 2026

GEMA, the collecting society that manages music rights for a huge share of German and international songwriters, sued OpenAI in Munich and picked its ground carefully. Not a sprawling claim about the entire training corpus, but nine specific, famous German songs, including "Atemlos" by Kristina Bach, "Manner" by Herbert Gronemeyer, and "Uber den Wolken" by Reinhard Mey. The claim was that OpenAI used those lyrics to train the models behind ChatGPT without a license, and that the models can spit the lyrics back out. On November 11, 2025, the Regional Court of Munich I agreed on the parts that matter.

The court's central move is the one to understand. It found that simple user prompts led ChatGPT to reproduce substantial parts of the original lyrics almost verbatim, and it treated that as a reproduction of the protected work. Crucially, it held that the storage of the content inside the model counts too. In the court's words, "it is irrelevant that the content is only contained in the language models in the form of probability values." That sentence is doing enormous work. It says you cannot escape copyright by pointing out that the model does not hold a literal copy of the text, only weights and probabilities. If the protected expression can be pulled back out, the law treats it as reproduced.

What about the text-and-data-mining exception?

This was OpenAI's main shield, and the court took it away. EU and German copyright law include a text-and-data-mining exception that lets you analyze large bodies of copyrighted material to extract patterns and information. The theory is that training is analysis, and analysis is permitted. The Munich court drew a hard line around that. It held the TDM exception covers only the initial analytical phase of training, and does not extend to memorization inside the model or to regurgitation in the output. As the court framed it, the limitation "does not permit memorisation in the AI language model or regurgitation." So training-as-analysis may be fine, but a model that has effectively memorized a work and can reproduce it is a different act that the exception does not reach.

The court largely upheld GEMA's claims for injunctive relief, disclosure, and damages, while dismissing a secondary claim that the songwriters' personality rights had been violated. So the copyright core landed and the personality-rights theory did not, which is itself a useful signal about which arguments have traction in this space.

How much weight should US attorneys give a Munich judgment?

Read it precisely, which cuts both ways. On the cautious side, this is a first-instance decision from a regional court, OpenAI has announced it will appeal to the Higher Regional Court of Munich, and it applies German law implementing an EU directive, not US law. It does not bind a US court and it does not settle the US fair-use question, which turns on a different four-factor test and is being fought out in cases like the Anthropic and New York Times litigation. Anyone telling clients that "OpenAI lost, so training is illegal" is overreading a non-final foreign judgment.

On the other side, do not wave it away. It is the most developed judicial reasoning yet on a theory that US plaintiffs are also pressing: that the problem is not only ingestion of copyrighted works during training, but what the model retains and can output afterward. That memorization-and-regurgitation framing is exactly what rightsholders in the US are using to argue around fair use, because a model that reproduces protected expression looks less like transformative analysis and more like copying. When a serious court builds out that theory in detail, it becomes a reference point that plaintiffs' lawyers everywhere will cite.

What should attorneys and their clients do now?

For counsel advising AI developers, treat output as a live copyright surface, not just training inputs. Ask clients what testing they do to detect memorization and verbatim regurgitation of known works, and whether they have output filters and guardrails that stop a model from reproducing protected text on a simple prompt. For clients operating in or selling into the EU, the exposure is concrete now, because a German court has granted an injunction on exactly this theory. For those advising rightsholders, the judgment is a template: identify specific works, show that simple prompts reproduce them, and argue that memorization sits outside the TDM exception. And across the board, revisit licensing. A negotiated license for training and reproduction is starting to look less like a nice-to-have and more like the cheapest available insurance against exactly this claim.

Questions professionals are asking

What did the Munich court actually decide?

It held that OpenAI infringed German copyright because its models memorized nine protected song lyrics and reproduced substantial parts of them in output from simple prompts. The court granted GEMA injunctive relief, disclosure, and damages, and ruled that storing the content as probability values inside the model does not avoid infringement.

Does the text-and-data-mining exception protect AI training?

Only partly, per this court. The Munich Regional Court I held that the TDM exception covers the initial analytical phase of training but "does not permit memorisation in the AI language model or regurgitation." So a model that memorizes and reproduces protected works performs an act the exception does not reach.

Is this ruling final, and does it bind US courts?

No on both. It is a first-instance judgment and OpenAI has announced an appeal to the Higher Regional Court of Munich, so it is not final. It applies German law implementing an EU directive, not US law, and does not bind US courts or resolve the US fair-use question. Treat it as a strong signal, not settled doctrine.

What should attorneys advising on AI do about it?

Treat model output as a copyright surface, not just training inputs. Ask clients what memorization and regurgitation testing they run and whether output filters stop reproduction of known works on simple prompts. For EU-facing clients the exposure is concrete now, and across the board a negotiated training and reproduction license looks increasingly like cheap insurance.

RELATED BRIEFINGS

Browse the full AI Regulation News tracker

Informational analysis for working professionals, not legal advice. Confirm how any court ruling or copyright question applies to your situation with qualified counsel in the relevant jurisdiction.