Munich Court Sanctions OpenAI for the Unauthorised Use of Song Lyrics

1. A landmark European ruling on the use of copyrighted works by AI

In a judgment delivered on 11 November 2025, the Regional Court of Munich (Landgericht München I) found two companies of the American group operating the generative AI models GPT-4 and GPT-4o liable for copyright infringement involving song lyrics.

This ruling is of major significance for the legal framework applicable to generative AI in Europe, addressing both the training of large language models and the generation of text in response to user queries.

The claimant, GEMA, Germany’s collective management organisation for musical works, represented the authors and publishers of nine well-known German songs. It alleged that the defendant entities had:

• used the lyrics without authorisation during model training, and
• enabled the reproduction, in whole or in part, of these lyrics within model outputs.

‍

2. Facts: ingestion of lyrics and their reproduction in AI outputs

‍

2.1. Works concerned

The case concerned songs such as Atemlos, Männer, Bochum, Junge, Über den Wolken, along with several children’s songs by Rolf Zuckowski, all protected by copyright.

‍

2.2. Presence of lyrics in the training data

The GPT-4 and GPT-4o models were trained on extensive datasets compiled through automated web crawling.

These datasets contained the full lyrics of the songs, even though the works had not been made freely available online nor licensed for this purpose.

‍

2.3. Reproduction of lyrics in generated outputs

GEMA demonstrated, supported by screenshots, that the models:

• generated significant or nearly complete excerpts of the lyrics,
• without performing any real-time internet search (the online search feature had been actively disabled),
• and did so consistently, despite variations in how the prompts were phrased.

This showed that the models could reproduce the claimants’ copyrighted works directly from their internal representations.

‍

3. Legal issues addressed by the Court

The Court examined several foundational questions for copyright law in the context of generative AI:

1. Does training a large language model on copyrighted works constitute a copyright-relevant reproduction?
2. Do AI-generated excerpts of lyrics amount to reproduction or making available to the public?
3. Can the Text and Data Mining (TDM) exceptions justify such usage?
4. Can the probabilistic or “hallucinatory” nature of outputs absolve the operator of liability?
5. Can a collective management organisation obtain broad injunctive and monetary remedies in such a context?

‍

4. The Court’s reasoning

‍

4.1. Training on copyrighted works constitutes reproduction

The Court adopted a clear and pragmatic approach:

Even if the model does not store the lyrics in readable form, the ability to regenerate recognisable excerpts demonstrates the existence of an internal fixation of the work, sufficient to qualify as a reproduction under copyright law.

Whether the data is stored as vectors, weights or other transformed formats is irrelevant:

If the model can reconstruct the work, it necessarily holds a legally relevant copy of it.

The Court rejected the argument that a model only manipulates statistical abstractions unrelated to the original expression.

‍

4.2. Outputs themselves constitute acts of infringement

The Court held that the generation of lyrics by GPT-4 qualifies as:

• a reproduction, since the text appears on the user’s screen and in chat history;
• an act of making available to the public, because the content can be accessed, viewed again or shared;
• and, where altered versions are generated, an unauthorised adaptation.

The defendants were found to exercise sufficient control over the design, training and filtering of the models to be directly responsible.

‍

4.3. TDM exceptions do not apply

The defendants invoked two TDM exceptions under the DSM Directive:

• the commercial TDM exception (subject to opt-out), and
• the research TDM exception.

Both were rejected:

1. GEMA had issued a valid opt-out, which barred any use of the works for machine learning purposes.
2. The companies operating GPT-4 pursue commercial objectives, incompatible with the strict conditions of the research exception.

This is a key part of the judgment:

TDM exceptions cannot be relied upon to justify training commercial AI models on protected works where rights-holders have opted out.

‍

4.4. Probabilistic behaviour and “hallucinations” do not negate liability

The Court dismissed arguments claiming that reproductions were accidental or the result of unpredictable stochastic processes:

The generation of protected text is a direct consequence of the model’s training.

The operator remains fully liable.

‍

4.5. Limited dismissal of moral rights claims

The Court held that accidental distortions did not amount to an intentional infringement of the authors’ moral rights.

Some moral-rights-based claims were therefore dismissed.

‍

5. Remedies ordered

The Court imposed a series of significant measures:

• a ban on using the lyrics within the GPT-4 and GPT-4o models;
• a ban on generating excerpts or adapted versions of the songs;
• an obligation to disclose all relevant uses and revenues;
• a finding of liability for all resulting damages;
• authorisation for judicial publication;
• and an order that the defendants bear 80% of litigation costs.

‍

6. Significance and impact

This Munich ruling marks a fundamental moment in the legal governance of generative AI in Europe:

• model training is subject to copyright law, even when the works are technically transformed;
• opt-outs under the TDM framework must be strictly observed;
• model operators are directly responsible for infringing outputs;
• generative AI cannot be assimilated to neutral search engines or hosting platforms;
• the EU confirms its position as one of the strictest jurisdictions worldwide on the use of copyrighted works in AI training.

This decision is expected to have global repercussions, influencing ongoing and future litigation in the United States, the United Kingdom and beyond, and shaping future negotiations between rights-holders and AI developers.

‍

‍

This analysis concerns a decision rendered under German law. For authoritative interpretation or further advice, readers should consult a lawyer specialising in German copyright law.