Munich Court: OpenAI Infringed Copyright by Training ChatGPT on German Song Lyrics
12 November 2025 - A regional court in Munich has ruled that OpenAI infringed German copyright by training and outputting lyrics from protected songs, including works by Herbert Grönemeyer. The decision, issued on Tuesday, 11 November, centers on nine German songs cited by the court.
The case was brought by GEMA, which represents composers, songwriters, and music publishers. Judge Elke Schwager ordered OpenAI to pay compensation for using the protected lyrics in model training and for reproducing them in outputs. The amount is not yet disclosed.
OpenAI argued its models do not store or copy specific works but learn statistical patterns, and that outputs depend on user prompts. The court rejected that position. It found the model's "memorization" during training and the reproduction of lyrics in ChatGPT responses both violate the exclusive right of exploitation.
"We hope this decision opens the way for dialogue with OpenAI on how to provide fair compensation for copyright holders," said GEMA legal adviser Kai Welp. GEMA's CEO Tobias Holzmueller added, "Internet is not a supermarket, and human creative work is not a free material that can be taken at will."
OpenAI disagreed with the ruling and said it is reviewing next steps. The company stated the case involves a small number of lyrics and does not affect daily use by millions of users and developers in Germany.
Key legal takeaways
- Training as a rights-relevant act: The court treated training on protected lyrics as a use that can infringe, not just the final output.
- Memorization matters: Model "memorization" of protected text can create liability, especially where verbatim or near-verbatim output is possible.
- Developer liability: The ruling rejects shifting responsibility entirely to end users where the system can reproduce protected lyrics.
- Licensing and opt-outs: In the EU, text-and-data-mining exceptions have limits, and rights holders can reserve rights. Compliance requires proof of licensing or respect for reservations.
Context for EU practitioners
The judgment may influence how EU courts view AI training on copyrighted works, especially lyrics and other short-form texts prone to verbatim recall. Expect closer scrutiny of dataset provenance, opt-outs, and memorization controls.
For reference, see the EU's copyright framework under the DSM Directive, including text and data mining provisions and rights reservations by rightholders: Directive (EU) 2019/790. For background on the rightholder organization involved, visit GEMA.
Practical steps for legal teams advising AI developers
- Run a rights audit: Identify training sources, confirm licenses, and document any reservations by rightholders. Maintain a dataset registry and a license evidence file.
- Respect reservations: Implement machine-readable opt-out detection and exclusion workflows for lyrics and other high-risk content.
- Reduce lyric leakage: Deploy memorization-reduction techniques, add lyric-blocking prompts and content filters, and monitor for verbatim output.
- Update terms and indemnities: Revisit customer terms, user policies, and publisher agreements. Calibrate indemnities, caps, and exclusions for training-related claims.
- Set up incident response: Create a process to receive, triage, and remediate claims (notice, model adjustments, retraining, and settlement protocols).
- Governance and logs: Keep detailed training and evaluation logs to evidence compliance and to support defenses or licensing negotiations.
What this means for rights holders
- Leverage reservations and licensing: Assert reservations over text-and-data mining where available, and engage in licensing discussions specific to AI training.
- Evidence collection: Capture instances of lyric reproduction and model behavior indicating memorization to support injunctions or damages.
- Collective action: Coordinate through CMOs for efficient enforcement and standardized licensing frameworks.
What to watch next
- Appeal and damages: Whether OpenAI appeals, and how the court quantifies compensation (statutory schemes, actual damages, or licensing proxies).
- Scope of injunctive relief: Any training or output restrictions that follow in enforcement proceedings.
- Global spillover: Similar claims are growing in other jurisdictions. Earlier this year, major Bollywood labels filed suit in New Delhi alleging unauthorized use of voice recordings for AI training.
Bottom line for counsel: treat training data like any other copyright-sensitive ingestion. Secure rights, respect reservations, reduce verbatim output risk, and prepare for audits and discovery. The cost of getting this wrong now includes both damages and product constraints later.
Your membership also unlocks:
