GEMA v. OpenAI

What GEMA alleged

GEMA, the German collecting society that licenses music on behalf of composers, lyricists and publishers, sued OpenAI in November 2024 over nine German songs, arguing that ChatGPT reproduces their lyrics through memorization inside the model and communicates them to the public whenever a chatbot output reproduces the text. GEMA describes it as the first collecting-society suit brought against OpenAI anywhere.

The nine works — including “Atemlos”, “Männer”, “Über den Wolken”, “Bochum” and “36 Grad” — are among the best-known songs in German popular music, chosen, GEMA’s filings suggest, precisely because their familiarity made memorization straightforward to demonstrate.

The suit sought an injunction, damages, and disclosure of the scope of use and revenue tied to it — the standard remedy package for a collecting society litigating on behalf of an aggregated repertoire rather than a single rights holder.

The legal question: does the text-and-data-mining exception cover what a model retains?

The EU’s text-and-data-mining exception, built into the Digital Single Market Directive, is the central defense AI developers rely on for training within Europe: it permits reproduction for the purpose of TDM, subject to a rights holder’s opt-out. GEMA argued the exception only covers the transient copies made during the analytical process of mining text, not the permanent retention of a work’s expression inside a model’s trained weights.

OpenAI argued the copies embedded in a trained model are not conventional, retrievable copies of the kind copyright law targets, and that the TDM exception should cover the training pipeline as a whole, including whatever a model incidentally retains as a byproduct of the process.

The ruling and its limits

The Munich Regional Court I sided with GEMA on November 11, 2025. It held that memorization is a reproduction under Section 16 of the German Copyright Act, that the TDM exception stops at the analysis phase and doesn’t reach copies retained afterward, and that chatbot outputs reproducing the lyrics count as communication to the public attributable to OpenAI, not to the user who typed the prompt. The court ordered OpenAI to cease the conduct, pay damages, and disclose the scope of use and revenue.

The ruling is a first-instance decision under German copyright law, tied to nine specific songs — it doesn’t resolve the broader, multi-jurisdictional question of where AI training itself takes place or what law governs it, and it is not yet final.

“Memorization of a work’s lyrics within a model’s parameters is itself a reproduction under Section 16 of the German Copyright Act, and the text-and-data-mining exception does not extend to copies retained after the analytical phase of training.”— Judgment, Case No. 42 O 14139/24 (LG München I, Nov. 11, 2025)

Where it stands

OpenAI announced it would appeal shortly after the judgment issued, and the case is expected to proceed to the Munich Higher Regional Court (Oberlandesgericht München). As of July 22, 2026, no appellate ruling has been reported, and commentary continues to treat the first-instance decision as under active review rather than settled law.

No further procedural detail — such as a hearing date at the appellate court — has been publicly confirmed as of this writing.

Why it matters for training-data buyers

This is the first European judgment to hold an AI developer directly liable for what its model memorized from training data. If that reasoning survives appeal, operators with EU exposure cannot rely on the TDM opt-out regime to cover a model that retains protected expression verbatim — the risk turns on what the model can be made to emit, not only on how the training set was assembled.

For data buyers, that is a different diligence test than US fair use: it means testing a model for memorization and regurgitation of specific or well-known works, not just confirming that a dataset was properly licensed at the point of collection. It also matters that a collecting society can bring this kind of claim at repertoire scale on behalf of its members, which changes the enforcement economics compared with a single rights holder suing alone — and Suno now faces the same Munich chamber over a companion case, with a verdict due July 31, 2026.

Practically, that pushes provenance obligations further downstream than most licensing agreements currently reach. A dataset can be properly licensed at acquisition and still leave a buyer exposed if the resulting model reproduces protected expression on demand — which is why output-level testing against known repertoire is becoming as important to EU diligence as the underlying license paperwork.