case-law

GEMA v. OpenAI: Germany Rules on AI Memorization

By Nora Whitfield · August 8, 2026

Every case in this pillar so far involves US courts and US copyright law. GEMA v. OpenAI doesn't, and that's exactly why it's worth understanding. It's the clearest sign yet that the legal fight over AI training isn't just an American story, and the result in Germany was more decisive against OpenAI than anything a US court has produced so far.

Who GEMA is and what happened

GEMA is Germany's music rights collecting society, the organization that manages licensing and royalty collection for composers, lyricists, and music publishers, roughly analogous to ASCAP or BMI in the US. GEMA sued two companies in the OpenAI group over the lyrics of nine well-known German songs, including "Atemlos" by Kristina Bach (also performed by Helene Fischer), Herbert Grönemeyer's "Männer," and Reinhard Mey's "Über den Wolken," seeking injunctive relief, disclosure of information, and damages.

On November 11, 2025, Munich Regional Court I ruled largely in GEMA's favor. The court found that ChatGPT had memorized the lyrics to these songs closely enough that the underlying model could reproduce them, largely faithfully, when prompted, and that this memorization itself constituted copyright reproduction under German and EU law. GEMA's secondary claim, alleging violations of the songwriters' personality rights, was dismissed, but the core copyright claims for injunctive relief, disclosure, and damages were upheld.

Why the reasoning matters more than the verdict

The legal reasoning here is the part worth understanding, because it cuts more directly against a common AI-industry defense than most US rulings have. AI companies frequently argue that training falls under text and data mining exceptions, provisions in both EU and (in different form) US law that permit certain automated analysis of copyrighted material without a license, on the theory that the analysis itself doesn't reproduce protectable expression to anyone.

The Munich court rejected that defense specifically as applied to memorization. Its reasoning: the text and data mining exception covers the initial analytical phase of training, the process of a model learning statistical patterns from text. It does not cover a model retaining and later reproducing specific protected expression closely enough that a user prompt can extract it largely verbatim. That's a meaningful distinction. Training a model to understand language patterns is one thing. A model that can be prompted to output the actual lyrics of a specific copyrighted song, largely intact, is functioning, in the court's view, as a reproduction mechanism, not merely an analytical one, and reproduction is squarely within the rights a copyright holder controls.

This is a sharper and more direct finding against an AI company's core defense than most of what's come out of US courts so far, where the fair-use analysis in cases like Bartz and Kadrey has generally been more favorable to the AI companies on the training question itself, while remaining tougher on how training data was acquired.

What this means for AI authors

If you're a songwriter, composer, or lyricist, particularly one with a substantial or well-known catalog, GEMA v. OpenAI is a concrete example of a court finding real, provable memorization and reproduction of specific protected lyrics, not a training-in-the-abstract argument. That's a meaningfully stronger evidentiary position than most AI copyright plaintiffs have managed to establish, and it's worth knowing this kind of claim has succeeded, in at least one major jurisdiction, if you're weighing whether your own catalog might have a similar claim.

More broadly, this case is a reminder that the legal landscape around AI training isn't uniform across jurisdictions, and a practice that survives one country's fair-use or text-and-data-mining framework may not survive another's. If your work is distributed internationally, or if you're licensing or negotiating around AI use of your material, it's worth knowing that European courts, at least in this instance, have been more willing to find liability for output-level memorization than US courts have been so far. That's a real, if narrow, data point about where the safer and less safe jurisdictions currently sit for a plaintiff with a strong memorization claim.

Copyrightable is not a law firm and doesn't provide legal advice, and this summary describes a foreign court ruling under German and EU law, which differs in important ways from US copyright law. Confirm current status, including any appeal, at official sources before relying on this summary.