Suno Loses GEMA Copyright Case: What It Means

Europe now has its first court ruling on whether generative music AI can be trained on copyrighted songs without a license — and the answer is no. On Friday, July 31, 2026, the Munich Regional Court ruled that Suno, the U.S. AI music startup, infringed copyright by training its models on protected recordings and storing them in a form that can be reproduced on demand. The case was brought by GEMA, Germany’s state-mandated collecting society for composers, lyricists, and music publishers, and the court found for GEMA on nearly every point.

The decision lands as courts and regulators across multiple jurisdictions grapple with the same unresolved question at the center of the AI boom: whether ingesting copyrighted work to train a model is fair use, a licensable act, or an infringement. Munich has now given one clear answer for the European market.

What the court decided

Presiding judge Elke Schwager, of the court’s 42nd Civil Chamber, found two distinct violations. First, storing the protected songs inside Suno’s model infringes the reproduction right — the court concluded the works were retained in the model parameters in a recoverable form, not merely analyzed and discarded. Second, serving generated outputs to users infringes the right of making works available to the public.

The ruling centered on six musical works, a list that spans decades of European pop: Boney M.’s “Rasputin” and “Daddy Cool,” Lou Bega’s “Mambo No. 5,” and tracks associated with Alphaville and Helene Fischer. GEMA’s legal team demonstrated the problem directly, entering the original lyrics, a target style, and the song title into Suno’s prompt box and producing outputs the court found to be recognizably the protected works. The court stated it was “convinced that the musical pieces in question are reproducibly contained in the defendant’s models” — specifically versions v3.5 and v4 — and noted those models were stored on servers in Germany, anchoring jurisdiction.

The court granted GEMA injunctive relief, disclosure, and damages. In practice, Suno must stop reproducing the six works without a license and must disclose the revenue it generated from their unlicensed use. The damages figure itself will be set in a separate proceeding, tied to what Suno earned during the period of infringement.

The defenses the court rejected

Suno did not concede quietly. It contested the case on several grounds, and the court dismantled each in turn.

Suno argued that the musical elements GEMA relied on were too generic to be protected, that its outputs were not recognizably similar to the originals, and that its model stored only mathematical patterns rather than copies of songs. The court disagreed on all three, finding the outputs confusingly similar and the underlying works reproducibly present in the weights.

The most consequential rejection concerned the text-and-data-mining (TDM) exception — the provision in EU and German law that lets researchers and companies analyze large datasets, and the legal foundation most AI labs lean on to justify training. The judges were blunt: the exception “does not apply, because the works were not only analysed during training but retained in the models in a reproducible form.” That distinction — between analyzing data and retaining it in recoverable form — is the hinge of the entire decision, and it is what makes the ruling dangerous for the broader industry.

Suno had also invoked U.S. fair use doctrine as a shield for its training activity. The court held that German copyright law governed conduct affecting German rightsholders and works stored on German servers, and that a foreign fair-use defense did not override it.

Why memorization is the whole argument

The technical crux of the case is a phenomenon researchers call memorization: when a model, rather than learning general patterns, effectively encodes specific training examples in its weights and can regurgitate them. It is the same failure mode that surfaces in text and image models, and it is precisely what separates a defensible “learning from data” narrative from an indefensible “storing copies” one.

Suno’s position rested on the claim that a neural network holds only abstract statistical relationships — a claim that, if accepted, would place training comfortably inside the TDM exception. GEMA’s evidence undercut it. By showing that carefully constructed prompts could pull recognizable versions of specific songs back out of the model, GEMA reframed the weights not as a lossy abstraction but as a compressed store of the training corpus. Once the court accepted that framing, the reproduction right was engaged and the TDM exception fell away.

This is the same conceptual territory covered in our explainer on how diffusion and generative models work: these systems are trained to reproduce the distribution of their training data, and when that data is copyrighted and the reproduction is faithful, the line between generation and copying gets very thin.

GEMA’s alternative: license it

GEMA did not arrive at the courtroom empty-handed. Ahead of the ruling, it launched PLAI by GEMA, a fully licensed training dataset of roughly 178,000 audio files drawn from about 57,000 works across more than 60 genres. The package bundles authors’ rights, master rights, the sound files themselves, and comprehensive metadata from a single source — an explicit answer to the “we had no legal way to license this at scale” argument that AI firms often raise.

The pairing is deliberate. GEMA is not simply trying to stop AI music generation; it is trying to route it through a licensing regime it controls, the same way it already licenses radio play, streaming, and live performance for its members. The lawsuit establishes the obligation; PLAI offers the compliant path.

Suno’s response

Suno said it disagrees with the ruling and is evaluating all available options, including an appeal. An appeal to a higher German court — and potentially referrals on EU-law questions — could take years to resolve, and the injunction’s practical bite in the meantime will depend on enforcement. But the direction of travel in Europe is now unambiguous, and Suno is not the only defendant watching: rival Udio faces similar exposure, and the same GEMA repertoire underpins a large share of European music.

What it means

This is a precedent-setting loss, and its weight comes from how the court reasoned, not just that it ruled against Suno.

The memorization framing is the real threat. By drawing the line at whether works are “retained in a reproducible form,” the court gave rightsholders across every media type a template: don’t argue about the abstract legality of training, prove that specific protected works can be pulled back out. Any generative model — music, image, text, or code — that can be prompted to reproduce its training data is now exposed to the same argument in Europe. That is a much narrower and more litigable claim than “training is infringement,” and it is far harder for labs to dismiss.

The TDM exception looks a lot weaker. Much of the European AI industry’s training strategy assumed the text-and-data-mining carve-out provided cover. Munich has signaled that the exception protects analysis, not retention — and drawing that boundary in practice will require labs to prove a negative about what their weights contain. Combined with the enforcement powers that arrive under the EU AI Act’s GPAI rules, European AI providers face a tightening compliance environment on two fronts at once: content licensing and model governance.

Licensing is becoming the default, not the exception. The contrast with the deal-making happening elsewhere is stark. Where music AI firms fought and lost, other content owners have moved to monetize — see the Getty Images licensing arrangement with OpenAI. Expect more collecting societies and publishers to follow GEMA’s playbook: sue to establish the obligation, then sell the licensed dataset that satisfies it. For AI companies, “scrape now, settle later” is looking like an increasingly expensive strategy.

Who wins, who loses. Rightsholders and collecting societies win a durable bargaining position. Licensed-data vendors win a market. AI music startups that built on unlicensed corpora — Suno chief among them — lose their cheapest input and gain a recurring cost. The open question is whether smaller labs can afford licensed data at all, or whether the ruling quietly consolidates generative music around the few players big enough to pay. That tension between openness and compliance is playing out across the whole field, from open-weight model releases to the courtroom fights over AI and trade secrets.

What to watch next. The damages proceeding will put a number on the cost of training without a license — the first hard figure the industry can price against. Watch, too, for whether Suno’s appeal produces a referral to the EU Court of Justice on the scope of the TDM exception; a ruling there would bind the entire bloc. And watch the U.S., where Suno and Udio face separate litigation from the major labels. Munich has shown one way that case could go if a court accepts the memorization argument.