GEMA v OpenAI: Potential hurdles for AI service providers

What has happened?

The GEMA, a German collecting society which takes care of the fiduciary administration of copyrights of music creators, claimed that OpenAI was using the lyrics of famous songs of its clients for AI training purposes in an unlawful manner.  Test results had shown that OpenAI tools like ChatGPT were reproducing such lyrics literally within their outputs. GEMA claimed that this behaviour constitutes a breach of its clients’ copyrights.

What did the court decide?

The court (LG München I, Urt. 42 O 14139/24 v. 11.11.2025) upheld the claim in most respects:

• It was assumed that the large language model must have included a reproduction (“memorisation”) of protected works in the Sense of § 16 UrhG (Urheberrechtsgesetz – German Copyright Act) as lyrics were literally included within its output – which then also constitutes a violation of copyright holders’ right to make such content publicly available according to § 19a UrhG
• Further, in particular the exception for “Text and Data Mining” according to § 44b UrhG does not apply according to the court’s decision as this clause may only cover mere data extraction for AI training purposes, not the literal reproduction of protected works
• Finally, the court did however not see any evidence for an infringement of personal rights which had been claimed by GEMA due to the fact that altered lyrics may become falsely attributed to affected artists

What does this imply for AI service providers?

The ruling strengthens the rights of copyright holders when being faced with large language models containing reproductions of their content. The court makes clear that in its opinion the described utilisation of protected content is not covered by statutory copyright exceptions – and thus requires other legal justification such as the particular copyright holders’ consent.

It is obvious that this decision will be challenged in front of all competent national and likely European courts. If it stands, it will impose a significant hurdle for LLM providers and their business models where availability of large amounts of training data from sources protected by copyright is crucial. It would become much more difficult to acquire such data when the consent of affected content creators is deemed to be required.  This would also be likely to significantly affect commercial framework conditions of connected AI services.

It will be interesting to see whether this assessment stands throughout the likely appeals process  – or if in particular the range of copyright exceptions explicitly allowing data analysis for training purposes may be interpreted more widely to make more room for development of innovative AI-driven services in the future.