GEMA v. Suno

What GEMA alleges

GEMA’s second AI suit, filed in January 2025, targets Suno’s text-to-music generator rather than a text chatbot. GEMA alleges Suno trained on protected recordings without a license and that its outputs are recognizably similar to specific, well-known songs — including “Forever Young” and “Big in Japan” by Alphaville, “Atemlos” by Kristina Bach, “Mambo No. 5” by Lou Bega, and “Rasputin” and “Daddy Cool” by Boney M.

The relief sought mirrors the OpenAI case: an injunction, information about the scope of use, and damages. But the underlying medium is different — this case tests audio generation and musical similarity, not the reproduction of text-based lyrics, and it arrived barely two months after GEMA’s first suit against OpenAI, suggesting a deliberate strategy of pursuing multiple AI developers in the same forum in quick succession.

The legal question: does memorization-as-reproduction extend from text to audio?

The Munich court has already held, in GEMA v. OpenAI, that memorization within a model is a reproduction and that the EU text-and-data-mining exception doesn’t cover what a model retains after training. The open question here is whether that reasoning, developed for the memorization of discrete strings of lyric text, extends to a music-generation model whose outputs may echo a recording’s melody, arrangement and production choices rather than reproducing identical text.

That distinction raises a harder evidentiary question than the OpenAI case did: musical similarity is a matter of degree — melodic, harmonic and production resemblance — rather than the more binary question of whether identical lyric text appears in an output. Courts assessing that kind of similarity typically rely on expert musicological comparison, which makes how the Munich court weighs the comparative-listening exercise from the March hearing especially significant.

The evidentiary moment: playing songs in the courtroom

At the March 9, 2026 hearing, the court had original recordings and Suno-generated tracks played back-to-back — a direct approach to assessing similarity rather than relying solely on written expert reports. It’s a concrete signal of how seriously the court is treating the comparative-listening exercise at the heart of the case.

The same Munich chamber ruled against OpenAI on memorization in November 2025, which gives some indication of how the court approaches AI-model liability generally — without predicting how it will resolve the harder audio-similarity question this case presents.

Where it stands: verdict delayed to July 31, 2026

The verdict was originally scheduled for June 12, 2026. On May 26, 2026, the court moved the date to July 31, 2026, citing internal administrative reasons rather than any substantive development. No ruling had issued as of July 22, 2026.

Because no decision has been reached, there is no holding yet to report — the case sits at the most consequential possible moment: a verdict expected within days of this article’s publication, and one likely to draw immediate comparison to the OpenAI ruling from the same courtroom.

Why it matters for training data

If GEMA wins and the court extends its memorization-as-reproduction reasoning from lyrics to generated audio, music-model operators with EU users would carry licensing obligations tied to what their models can output, regardless of where the underlying training occurred — echoing the structural logic of the OpenAI ruling.

Collecting societies like GEMA aggregate claims across large repertoires cheaply, which makes Germany — and the EU generally — a comparatively low-cost enforcement venue next to per-work US litigation. For anyone sourcing or supplying audio training data, the practical takeaway is to document repertoire clearance and test model outputs against well-known reference recordings, not only to verify licensing at the point of acquisition.

The outcome will also inform how Suno’s parallel US litigation, Labels v. Suno, gets read by observers on both sides of the Atlantic: a GEMA win here, even under German law, would strengthen the argument that music-generation output — not just training method — is where enforceable liability concentrates.