Germany delivers landmark copyright ruling against OpenAI: What it means for AI and IP

A landmark decision in Munich

The Regional Court of Munich (LG München I) has issued a landmark judgment in GEMA v OpenAI (Case No. 42 O 14139/24), holding that the use of copyrighted song lyrics for training generative AI models without a licence violates German copyright law. This is the first decision in Europe to address these questions directly, providing clarity after the UK ruling left key issues unresolved. It sets a precedent that could reshape AI compliance strategies across Europe.

The court found that:

• The judgment explicitly cites §§ 15, 16, 19a, 44b UrhG and Articles 2, 3 InfoSoc Directive and Article 4 DSM Directive.
• It clarifies that § 44b UrhG (Text and Data Mining) does not apply because the models permanently memorize lyrics, exceeding the scope of transient analysis.

• Reproduction (§16 UrhG) occurs when lyrics are stored within the model.
• Public communication (§19a UrhG) applies when ChatGPT outputs those lyrics on user prompts.
• The text and data mining exception does not apply where substantial parts of protected works appear in outputs, especially after GEMA exercised its opt-out rights.

Claims based on personality rights were dismissed, but the ruling underscores that AI providers cannot sidestep licensing obligations when their models can reproduce protected works.

What was at stake?

The case centered on lyrics from well-known German songs, including “Atemlos” and “Wie schön, dass du geboren bist”, which could be generated by ChatGPT through simple prompts. The court rejected OpenAI’s argument that users, not the platform, were responsible for any infringement and dismissed reliance on innovation-driven copyright exceptions.

Significantly, the injunction goes beyond prohibiting reproduction:

• OpenAI must cease storing unlicensed German lyrics on infrastructure in Germany.
• The judgment must be published in a local newspaper – a symbolic but striking remedy.
• Damages will be assessed later and enforcement will require proving specific prompts, a practical challenge for rights holders.

Legal reasoning: key takeaways

The court confirmed memorization within model versions '4' and '4o', rejecting arguments about mere statistical correlation. A systematic comparison of training data and outputs demonstrated full reproduction.

The court clarified that:

• Responsibility lies with OpenAI as the operator, not end-users, due to decisions on architecture and dataset selection.

• Even statistical representations of lyrics do not shield liability when outputs reproduce substantial parts of protected works.
• Slightly modified outputs still constitute reproduction.
• Text and data mining exceptions cannot justify outputs that replicate original works.

This reasoning signals that technical arguments about “correlation storage” will not suffice to avoid copyright liability.

Comparison: Getty v Stability AI

Just days earlier, on 4 November 2025, the English High Court delivered its first AI-related copyright judgment in Getty v Stability AI (see discussion on our NRF Inside Tech Law-Blog). Getty alleged that Stability AI used its copyrighted images to train the Stable Diffusion model without authorization. 

The decision was narrowly framed:

• Primary copyright claims were dismissed because the training occurred outside the UK, leaving no territorial basis for infringement.
• Secondary claims that the model itself constituted an infringing copy were also rejected, as Stable Diffusion does not store protected works in its model weights (statistical parameters).
• The Court did, however, find limited trademark infringement due to Getty watermarks appearing in outputs, confirming that liability rests with the provider, not end-users.

Crucially, the judgment leaves key questions unresolved, particularly whether training on copyrighted material within the UK jurisdiction would amount to infringement. Technical arguments about model weights as statistical representations were decisive.

This contrasts sharply with the German ruling in GEMA v OpenAI, which imposed strict liability for memorization and reproduction of lyrics, rejecting the TDM exception and signaling a more rigorous European approach.

While the UK case focused on image datasets and avoided broader copyright questions, both decisions share a common theme: namely that European courts are increasingly scrutinizing unlicensed ingestion of protected works for AI training, though the UK approach remains narrower than Germany’s.

Together, these cases mark a turning point in European IP enforcement against generative AI.

Industry and regulatory outlook

• Licensing is not optional: Both rulings make clear that copyright exceptions have limits.
• Strategic litigation: GEMA frames this case as a defense of its licensing model, while OpenAI has so far declined to negotiate – a stance that may become untenable.
• European advantage: Unlike the U.S., where fair use and high statutory damages dominate, Europe’s approach emphasizes licensing and proportional remedies. This could become a competitive differentiator for compliant AI providers.
• Next steps: Appeals are expected and a referral to the Court of Justice of the EU could pave the way for harmonized standards.

Why this matters

These decisions are more than legal milestones, they signal a new era where AI innovation must coexist with robust IP protection. For technology companies, the message is clear: transparency and licensing are not optional but essential for sustainable AI development.

The judgment is not yet final; appeals and potential referral to the CJEU remain possible.

Further reading:

https://www.insidetechlaw.com/blog/2025/11/getty-v-stability-ai-first-ai-judgment-made-but-key-questions-remain-unanswered