GEMA v Suno: the AI-training ruling that does not (yet) bind LLMs

A ruling, and the ruling behind it

On 31 July 2026 the 42nd Civil Chamber of the Munich I Regional Court (Landgericht München I) handed down a judgment against Suno, the US provider of an AI music generator (case 42 O 763/25, presiding judge Elke Schwager). The court found that Suno infringed the copyright in six compositions — "Atemlos durch die Nacht" (Kristina Bach, recorded by Helene Fischer), "Daddy Cool" and "Rasputin" (Boney M., written by Frank Farian and others), "Forever Young" and "Big in Japan" (Alphaville), and the chorus of "Mambo No. 5" (Lou Bega) — and ordered it to stop, to disclose its revenues, and to pay damages, the amount still to be set (JUVE Patent, 31 July 2026; tagesschau, 31 July 2026). Songs, not recordings: the performers' and record producers' claims were not asserted, and only the musical works were in dispute (Bird & Bird, 7 August 2026).

The judgment was not the first of its kind. In November 2025 the same chamber, before the same judge, had already ruled for GEMA against OpenAI, holding that ChatGPT's reproduction of famous song lyrics (most visibly Herbert Grönemeyer's "Mensch") was an infringement (LG München I, 11 November 2025, 42 O 14139/24, GRUR-RS 2025, 30204). That was the first time a European court had found for creators whose works were used by a generative AI system (Music Business Worldwide, 31 July 2026). The Suno case is the sequel: music instead of lyrics, melody instead of text, and a ruling that this time directly condemns the use of works during training.

Because both decisions remain first-instance, neither is yet legally final. Judgments of German regional courts bind only the parties in the case at hand; German law has no system of binding precedent in the common-law sense; in practice, it is the Federal Court of Justice (Bundesgerichtshof, BGH) and the Court of Justice of the European Union whose rulings fix the law, and neither has yet ruled on AI training. OpenAI's appeal is pending at the Munich Court of Appeal (OLG München, case 6 U 3662/25), and Suno announced it was evaluating "all available options, including an appeal" (Suno statement, 31 July 2026). As of early October 2026 no appellate decision in either case had been reported.

Four ways to infringe

The court's core finding is deceptively simple: a work that an AI model can reliably reproduce is a copy of that work. On the facts it decided that the six songs were "reproducibly contained" in Suno's models v3.5 and v4 — "These models were stored on servers in Germany," the court's press release read — and that this "memorisation" was confirmed "by comparing the musical pieces contained in the training data with their playback in the outputs. Given the complexity and length of the musical pieces, randomness as the cause of their playback can be ruled out" (LG München I press release, quoted in Music Ally, 31 July 2026).

From there the court found four separate infringements (Bird & Bird, 7 August 2026):

  1. Training in the United States — Suno's downloads, format conversions and backups of the recordings were reproductions under US law, an infringement unless "fair use" applied; the court held it did not (17 U.S.C. §§ 106(1), 101).
  2. The model in Germany — fixing the works in the model's parameters is a reproduction under § 16 UrhG (implementing Article 2 of the InfoSoc Directive), because the works are materially embodied in the weights on German servers.
  3. Offering the model — making the generator available to the public infringes the "unnamed" right of communication to the public (§ 15(2) UrhG, Article 3(1) InfoSoc Directive): the offering is complete once access is enabled, and actual retrieval of a protected work is unnecessary.
  4. The outputs — recognisably similar outputs (measured note-by-note from an average listener's perspective) are reproductions communicated to the public, for which Suno — not the end user — is directly liable.

The last point matters beyond music. Suno argued that whatever resemblance appeared was the doing of its users, who typed the prompts. The court rejected this: a user who only supplies lyrics, a style tag and a title — nothing about melody, harmony, rhythm or arrangement — cannot be said to "steer" a model towards a particular song. The same input, parameters and probabilities determine everything; repeating a prompt does not steer (Bird & Bird, 7 August 2026). The user is the trigger, not the author of the copy.

Memorisation is not text and data mining

The decisive legal question for the wider AI world is what this says about the EU's text-and-data-mining (TDM) exception, Article 4 of the DSM Directive, implemented in Germany as § 44b UrhG. This is the provision nearly every AI company in Europe has relied on: it permits reproductions needed to analyse lawfully accessible works, subject to the rights holder's option to reserve rights.

The Munich court gave with one hand and took with the other. It first accepted that § 44b does apply to generative AI training at all — the copies made while assembling and preparing the training corpus, such as format conversions and backups for analysis, fall within it (Bird & Bird, 7 August 2026; relying on Recital 18(1) of the DSM Directive and the German legislator's reference to machine learning). Then it held that the exception stops precisely where the model's parameters begin: memorisation is not analysis but reproduction, so the premise of the exception — that analysis leaves the author's own exploitation untouched — no longer holds. Because the exception offers no remuneration, extending it to memorisation by analogy would leave authors unprotected, contrary to Recital 17 of the DSM Directive (Bird & Bird, 7 August 2026).

The earlier OpenAI ruling went further still, holding that training the model itself is not covered by § 44b at all, which in the court's view covers only the preparation of the training material (LG München I, 42 O 14139/24, headnote 4). And the Suno court added a sentence with the widest possible reach: if current technology cannot prevent memorisation, then training on protected works is entirely excluded from the exception — "a business model that helps itself to the intellectual property of others free of charge is, in the court's words, unknown to both EU and German law" (Bird & Bird, 7 August 2026).

There was a second, independent reason the exception failed: lawful access. Suno had obtained the recordings by "stream-ripping" them from YouTube, circumventing that platform's download restriction, the "Rolling Cipher". The court treated the Rolling Cipher as an effective technical measure under § 95a UrhG, and access obtained by defeating it is not lawful access within Article 4 and Recital 18(2) of the DSM Directive (Bird & Bird, 7 August 2026; JUVE Patent, 31 July 2026). Suno also invoked the EU AI Act; the court dismissed this as conflating transparency with authorisation — complying with Article 53(1)(c) and (d) of Regulation (EU) 2024/1689 (disclosing a training-data summary) does not replace a licence, and the AI Act's Code of Practice itself "does not constitute compliance with Union law on copyright" (Bird & Bird, 7 August 2026).

Fair use, decided in Munich

Because the training took place in the United States, the court had to decide whether US law turned those parts of the claim into an infringement. A German court establishes foreign law of its own motion, and it did so here: it applied US copyright law to the US acts, refused Suno's request for an expert on American law, and answered the question that US judges have been struggling with for two years — is training on copyrighted works fair use?

Its answer: no, on all four factors. The court distinguished the two 2025 Californian decisions that had leaned the other way, Bartz v Anthropic and Kadrey v Meta, on a single factual point. There, the training material never reappeared in the outputs; as Bartz put it, "if the outputs seen by users had been infringing, Authors would have a different case". Here the works did resurface, from simple open-ended prompts, so the use was not transformative, was commercial, appropriated the full expressive core of the works, and substituted for the originals — and the circumvention of YouTube's download protection weighed as bad faith under 17 U.S.C. § 1201(a) (Bird & Bird, 7 August 2026; JUVE Patent, 31 July 2026; Music Ally, 31 July 2026).

A venue rule that reaches across the Atlantic

The court's jurisdiction over the American training has no direct precedent, and rests on an accident of German procedural law. Section 131 of the German Collecting Societies Act (VGG) gives collecting societies a privileged forum for all their copyright claims — even, the court held, claims about acts abroad, drawn into the German court by a rule designed for domestic concentration. Three limits were noted: the gateway is only open to collecting societies, at least one infringing act must occur in Germany, and the recognition and enforcement of the German judgment in the United States remains an open question the ruling does not address (Bird & Bird, 7 August 2026). For a trade body with the deep pockets of GEMA this is a powerful instrument; for individual authors it says nothing.

A European landmark for music

Strip away the technicalities and the practical significance for the music business is easy to state. This is the first European ruling — even if only first-instance — to hold that an AI company must licence the music it trains on, and both sides of the industry treat it as a turning point. GEMA's chief executive Tobias Holzmüller called it a decisive strengthening of "Europe's position as a cultural hub"; the IFPI said it "reinforces the principle that the use of music to develop commercial AI products requires authorisation" (Music Ally, 31 July 2026). The German lawyer Christian Solmecke distilled what other AI-music providers will take from it: "No AI music provider in Europe can ignore this judgment" (tagesschau, 31 July 2026).

The commercial picture had already moved before the ruling. Warner Music settled its US lawsuit against Suno in November 2025 and signed a licensing partnership, with current models to be replaced by licensed ones in 2026; Universal and Sony are still litigating, and Udio settled with both UMG and Warner (tagesschau, 31 July 2026; MBW, 31 July 2026). The Danish collecting society Koda has its own claim against Suno, and more than 1,800 artists back US class actions against Suno and Udio (MBW, 31 July 2026, citing Reuters). GEMA has simultaneously built the instrument it wants AI companies to buy: PLAI, a licensed dataset for training music models (GEMA). For rights holders worldwide the Munich judgment is, in the German press's word, negotiating leverage — "Verhandlungsmasse" (tagesschau, 31 July 2026).

Nothing about the market will change overnight: Suno remains available in Germany, the case is about licensing and payment, not a ban, and the damages figure has not been set (AI Musicpreneur, 31 July 2026). What has changed is the default assumption on which every licensing conversation now starts.

Does it transfer to LLMs? The hinges

The question the reader of this site will care about is narrower: does this ruling become a precedent for training large language models on text? The honest answer is that it transfers only where its conditions are met, and those conditions are specific. The direct-liability finding rests, as the court itself framed it, on two cumulative facts — the works are memorised in the model, and they surface in response to simple, open-ended prompts. A model that does not memorise, or a service whose infringing outputs appear only under "steering" prompts designed to force a result, falls outside the finding and would instead be assessed under the general intermediary-liability framework (Bird & Bird, 7 August 2026). The court therefore did not hold that "AI training infringes copyright". It held that memorisation-and-reproduction infringes, and that when a model demonstrably does that, neither the TDM exception, nor the AI Act, nor the user's prompt, nor US fair use will save the provider.

For text models the first hinge — memorisation proven by outputs — is the narrow one. LLMs can and do reproduce their training data, but usually for a tiny share of inputs. The canonical demonstrations are deliberate extraction attacks: the 2021 study that pulled individual records from GPT-2's training data by prompting it with recognisable prefixes, and the later finding that a model 10 times larger could be made to output legitimate training data many times more often (Carlini et al., "Extracting Training Data from Large Language Models", USENIX Security 2021; Carlini et al., "Quantifying Memorization Across Neural Language Models", 2022, arXiv:2202.07646). In daily use, verbatim or near-verbatim regurgitation concentrates in exactly the high-value, high-duplication categories a collecting society cares about: song lyrics, book passages, code, news articles. A German court following the Munich reasoning would face limited difficulty holding that an LLM which outputs "substantially similar" protected passages has reproduced and communicated them, that the provider of the relevant ChatGPT-like service is directly liable because it exercises control over the outputs, and that memorised copies inside the model are reproductions under § 16 UrhG. The "offer of the model as communication to the public" theory would in principle extend to any generative model offered in Germany.

Nor is the strictness of the court a quirk of music. Two features of its reasoning press hard on providers of any generative technology. First, the court held that an exact or discrete copy is not required for a reproduction: what matters is what is taken over, not how much — a model may encode a work lossily, the way an MP3 discards detail, and still reproduce it. Second, it placed the memorisation risk squarely in the provider's sphere: a provider cannot point to its own inability to prevent memorisation as a defence, because if the technology cannot prevent it, training on protected works falls outside the exception altogether. That last point, if it survives appeal, is the one with the widest blast radius for LLMs.

Where the reasoning breaks for text

There are, however, three places where the transfer genuinely breaks. The first is the lawful-access finding. Suno lost on TDM partly because it ripped recordings from YouTube across a technical protection measure. The web text on which LLMs are trained is largely crawled openly; a rights holder's words may be paywalled or behind a login, but much of it is "freely available" in the sense of Article 4. The stream-ripping point is a gift for rights holders who publish behind access controls, but it does not automatically disqualify ordinary web crawling.

The second break is the collecting-society venue. The extraterritorial limb of the injunction — the bit that reaches across the Atlantic — runs through § 131 VGG, which exists only for collecting societies. An author suing over LLM training would have to establish jurisdiction by the ordinary routes: a domestic act of infringement in Germany (an output served here, a model hosted here) plus the usual rules. That is still plausible, and it is how the court got jurisdiction over the German acts in the first place, but it is not the same powerful shortcut.

The third break is the evidentiary one, and it is the most important. In both German cases GEMA won because the works resurfaced in outputs that the court itself could compare, note-by-note, with the originals; the court even declined expert evidence, judging as "the average listener". For an ordinary LLM the equivalent showing would have to be made work-by-work and passage-by-passage, and the outcomes of extraction attacks show that memorisation is concentrated and non-uniform. A plaintiff who cannot point to reproduced passages has no memorisation holding and no output infringement; what is left is the claim that the training corpus itself involved unlawful copying — which brings us to the question the Munich court did not, and could not, settle.

The question that is actually pending

The general question — is copying works into a training corpus at all a copyright use, and does the TDM exception cover it — is not finally answered by GEMA v Suno; it is pending before the two courts whose answers would actually bind. In Germany the BGH heard the LAION case (Kneschke v LAION, I ZR 281/25) on 3 September 2026: the photographer Robert Kneschke's claim that the LAION-5B dataset, roughly 5.85 billion scraped images, copied his photos for AI training without permission. The BGH indicated it may refer the scope of the TDM exception to the CJEU, and its decision is awaited (ad-hoc-news, 3 September 2026). In parallel, the CJEU has a reference on the same provision (Like Company v Google, C-250/25), with the Advocate General's opinion expected in September 2026 (Bird & Bird, 7 August 2026). If the BGH or the CJEU reads Article 4 broadly, then analysis-oriented training on lawfully accessible works is protected in the EU (subject to the rights holder's reservation of rights), and the Munich decisions end up confined to the memorisation-and-reproduction corner. If they read it narrowly, the Munich dictum becomes the template: training on works whose memorisation cannot be prevented is outside the exception, and EU-based LLM training needs licences.

The industry's case, in full

The steelman of the losing side deserves to be stated clearly. Suno's technical claim — that a model does not store its training data, only learned parameters, and that "reproducible repetition" is a property of a subset of high-memorisation works, not the normal operation of the generator — is not fringe; it is the working assumption of most of AI research. The 2021–2022 extraction studies that prove memorisation exists also show it is rare and uneven. The court did not test this by hearing Suno's expert evidence; it refused Suno's applications, judged the outputs itself, and applied what it took from the studies GEMA put before it. A critic would say the factual finding of "memorisation" was built by a party and confirmed by a court that had declined to be contradicted.

The broader objection is policy. The tech coalition Chamber of Progress — whose members include Suno, OpenAI, Google and Midjourney — argues that GEMA's real project is the collective licensing of all training material: "It's a cruel irony that, in pursuing that objective in the name of European artists, it may instead deny those creators access to one of the most transformative creative technologies in history" (Music Ally, 31 July 2026). The EU's own TDM exception was designed so that analysis of content does not require permission; a rule that any storage in a model is a reproduction turns that design inside out, since no large model can guarantee zero storage. On this reading the ruling would not protect individual creators but would entrench collecting societies as gatekeepers, and would do so at exactly the moment US litigation is pointing the other way. Whether Massachusetts judges in the UMG/Sony v Suno case follow a German court's application of their own fair-use law remains to be seen; they are not bound by it.

Precise about "precedent"

It is worth being exact about the word. In the common-law sense the German ruling is not precedent for anything: it binds only the parties, on these facts. In the German and EU sense, only the BGH and the CJEU set rules that lower courts must follow, and neither has spoken on AI training yet. What the Munich court provides is persuasive guidance — a careful, citation-rich application of existing copyright law by a specialised chamber that has now written the same analysis twice, ten months apart, for two different modalities. That consistency is what gives the reasoning weight. My reading, in confidence terms: the core holding — that a model reproducibly containing and emitting a protected work reproduces it, and that the TDM exception does not cover that — is likely to survive the appeals in substance (roughly a 2-in-3 chance it is substantially confirmed, on the strength of the two consistent decisions and mainstream copyright doctrine), while the two expansive claims that matter most for LLMs are genuinely unsettled: that training the model as such falls outside the TDM exception, and that a provider cannot rely on the impossibility of preventing memorisation. Both could go either way depending on what the BGH and the CJEU do with Article 4; I would not put better than even money on either surviving unchanged.

What it changes, either way

Whatever the appellate courts do, the case rearranges the negotiation. Two effects are already visible. On the rights-holder side, the belief that "training is legally safe in Europe" is gone; the burden of explanation has moved onto AI companies, and every collecting society with GEMA's resources now has a template for jurisdiction, evidence and argument. On the licensing side, the market is consolidating around exactly the outcome GEMA predicted: settled deals with Warner, a proprietary licensed dataset, and an expectation that catalogue owners are paid. None of that required a final judgment.

For LLM training data specifically, the practical answer to the question in our title is: not yet — but the runway is shorter than it looks. The ruling does not ban training on ordinary text; the matter that would do that is pending before the BGH and the CJEU. What it does is give rights holders a judicial statement — twice — that a model which reproduces their works is infringing, that its operator is the responsible party, that offering it in Europe is itself a use of the work, and that European courts will assert jurisdiction over US companies that serve the European market. Any text of value that an LLM can be shown to emit, substantially similar, from a simple request, now sits squarely within that logic. The question for LLM providers is no longer whether the analysis is plausible; it is whether they can demonstrate, as Suno could not, that their models do not encode the works they trained on.

Sources

Legal decisions and records

Law-firm analysis

Reporting

Parties and industry statements

Primary research on memorisation

Statute and case references cited in the text: § 15(2), 16, 19a, 44b, 95a UrhG; § 131 VGG; Article 2 and 3(1) InfoSoc Directive 2001/29/EC; Article 4 and Recitals 17–18 DSM Directive (EU) 2019/790; Article 53(1)(c)–(d) AI Act (EU) 2024/1689 and its Recitals 107–108; Article 8(1) Rome II Regulation (EC) 864/2007; Articles 106(1), 101, 107, 1201 U.S.C.; CJEU C-426/21 Ocilion, C-188/24 & C-190/24 WebGroup Czech Republic, C-580/23 & C-795/23 Mio, C-682/18 & C-683/18 YouTube and Cyando, C-433/20 Austro-Mechana.

For music this is a European landmark; for LLM training data it is a marker, not a precedent, because the decisive questions are still before the Federal Court of Justice and the CJEU.