Germany - GEMA v OpenAI, 11 November 2025

Claimant: Gesellschaft für musikalische Aufführungs- und mechanische Vervielfältigungsrechte ('GEMA')

Defendant: two companies (Irish and American) within the OpenAI group ('OpenAI')

GEMA is a collective management organisation that administers the rights of use for musical works entrusted to it by composers, lyricists, and music publishers.

OpenAI is a developer and provider of large language models, and of chatbots built upon them.

GEMA has filed an action against OpenAI before the Munich District Court, alleging copyright infringement on the grounds that the language models provided by OpenAI memorise copyright-protected song lyrics and reproduce them in their output as a response to simple prompts.

GEMA’s claims

GEMA argued that the memorisation of the copyright-protected song lyrics within the language models and the subsequent reproduction in AI-generated output (so-called 'regurgitation') constituted an unlawful reproduction, a communication to the public, and an adaptation under Sec.  16, 19a, 23 German Copyright Act. In addition, GEMA maintained that this infringed upon the personal rights of the lyricists.

OpenAI’s defence

OpenAI disputed the claims, asserting that its models did not store or copy specific training data but instead learned statistical patterns from large datasets. Since outputs were generated dynamically in response to user prompts, OpenAI argued that responsibility for any alleged infringements lies with the user rather than with OpenAI. It further maintained that any potential use of copyrighted material fell under the text and data mining exemption provided in Sec. 44b German Copyright Act. Additionally, OpenAI asserted that by making their works publicly available online, rights holders implicitly consented to such use.

Judgement of the Court

The court ruled predominantly in favour of GEMA. The claims for injunctive relief, information, and damages were largely upheld, while allegations concerning personal rights violations were dismissed.

Memorisation as an unlawful reproduction

The court held that memorisation in the language models met the requirements of a reproduction. It emphasised that it was irrelevant that the works were present only as probability values distributed across multiple parameters; it was sufficient that the model could render the content recognisably as song lyrics in response to prompts. Although the works were only indirectly perceptible (i.e., accessible via certain prompts), this did not change their classification as a reproduction in the large language models.

Regurgitation as an unauthorised adaption, reproduction and public communication

The court classified the reproduction of protected content in the AI-generated output as an adaptation within the meaning of Sec. 23, a reproduction pursuant to Sec. 16, and a communication to the public under Sec.19a German Copyright Act. Even text versions containing elements added by the AI model ('hallucinations') were deemed protected, provided that the original lyrics remained recognisable. The court further found that a communication to the public had taken place, as the chatbot was accessible online to the public at any time.

No justification under legal exceptions

The court held that the infringements were not justified by any legal basis. The text and data mining exception under Sec. 44b German Copyright Act did not cover the use at issue, as the AI training process involved the incorporation of the protected lyrics themselves rather than merely abstract information derived from them. Consequently, the exception did not justify either the memorisation of the works within the AI model or their regurgitation in the output. The court also rejected any implied consent, noting that memorisation and reproduction of protected works by AI models cannot be considered a normal or expected use of copyright-protected content.

Responsibility lies with OpenAI

Contrary to OpenAI’s argument that users are solely responsible for generated outputs, the court ruled that, as developer and provider of the models, OpenAI bore responsibility for the infringements.