The full text of the Munich I Regional Court judgment in GEMA v. Suno runs to 143 pages. Circulating today in English translation, it shows a German civil chamber conducting its own musicological analysis, applying United States copyright law directly, and concluding that the fair use doctrine does not shield the training of a music generator whose outputs reproduce the works it was trained on.

The 42nd Civil Chamber of the Munich I Regional Court decided case 42 O 763/25 on 31 July 2026, following an oral hearing held on 9 March 2026. The operative part orders the defendant to stop reproducing six musical works for training purposes in the United States, to stop storing them as parameters or other data structures inside the model, to stop offering that model in Germany, and to stop distributing arrangements of the works. Each future violation carries a court-determined fine of up to 250,000 euros, or alternatively custodial detention of up to six months.

According to the judgment, the plaintiff is a collecting society administering rights assigned by composers, lyricists and music publishers. The defendant, based in the United States, develops, operates and licenses an AI music generator reachable in Germany through a website hosted in the United States. Versions 3.5 and 4 of that generator are the subject of the case.

What the operative part actually orders

Beyond the injunction, the chamber granted an information claim and a declaratory damages ruling, both limited in scope. The defendant must disclose the number and extent of the infringing acts and the revenue derived from them, covering everything since 1 July 2023, but only for three of the six works plus the chorus of a fourth. The same limitation applies to the damages declaration. In all other respects the complaint was dismissed.

Two further orders carry weight beyond the parties. The plaintiff was authorised to publish the operative part of the judgment, once final, in the Süddeutsche Zeitung on a weekend, on page 3 of the business section, in quarter-page landscape format, at the defendant's expense. The chamber justified that order on the ground that the legal assessment of whether a model commits copyright infringement affects every author whose works were used to train a model, not only the litigants.

The defendant was also ordered to pay pre-trial attorney fees of 5,049.70 euros plus interest at five percentage points above the base rate from 30 July 2025, calculated on a dispute value of 600,000 euros. Costs fall entirely on the defendant. Provisional enforcement of the four injunction limbs requires security of 150,000 euros in each case, 70,000 euros for the information and damages items, and 10,000 euros for the publication order.

Six songs, 176 prompts

The works at issue are "Atemlos durch die Nacht" by Kristina Bach, "Rasputin" by Frank Farian, Fred Jay and George Reyam, "Big in Japan" and "Forever Young" by Marian Gold, Bernhard Lloyd and Frank Mertens, "Daddy Cool", and the chorus of "Mambo No. 5 A Little Bit of". For the last of these, only the arranger's rights in the chorus were asserted.

The prompt data is the most quoted number in the file. Employees and legal representatives of the plaintiff entered the original lyrics, a music style and a title into the generator, without any specification of melody, harmony, rhythm or arrangement. Producing the disputed output for "Atemlos durch die Nacht" took 176 identical prompts. "Big in Japan" took 124. The others took 14, 12, 8 and 4.

That figure cuts both ways in the judgment, and understanding why explains the shape of the ruling.

Memorisation treated as reproduction

The chamber inferred memorisation under section 286 of the German Code of Civil Procedure by comparing training data with outputs generated from simple, open-ended prompts. Exact identity between output and original is not required. The judgment records that the defendant's training corpus contained millions of complete audio recordings, that complete recordings were necessary because fragments cannot teach how musical form unfolds over minutes, and that the works at issue were among them.

Technically, the reasoning rests on determinism. Both the transformer and diffusion components produce identical results for identical internal states and inputs, with randomness introduced downstream by the selection algorithm applied after the model output. Audio was tokenised into segments of roughly 10 to 50 milliseconds, encoding features such as volume, timbre and spectral shape, with detail deliberately discarded through quantisation. The chamber nonetheless held that fixation in probability values suffices for a reproduction, drawing analogies to MP3 and JPEG compression.

Because the model is stored on edge servers inside Germany to serve German users, the chamber found a reproduction under sections 15(1) and 16 of the German Copyright Act occurring on German territory. The judgment also notes that the defendant has been aware of the memorisation phenomenon since at least a 2021 study by Carlini and co-authors.

The chamber conducted the musicological comparison itself, refusing the defendant's applications for expert evidence, and counted its judges both among listeners reasonably familiar with musical matters and among average listeners. Passages of the judgment work through alternating bass figures, five-eighth-note upbeats and the substitution of a minor sixth for a major sixth, concluding that such deviations do not break the auditory connection to the original.

Why the text and data mining exception failed

Section 44b of the German Copyright Act, which implements Article 4 of the DSM Directive, did not save the training. The chamber gave two independent reasons.

First, the permanent reproduction inside the model does not serve further data analysis. Text and data mining covers the analysis of patterns; the retention of the work itself in the parameters goes beyond that. The judgment states the position in one sentence that has already circulated widely: "If, given the current state of the art, the memorization of training data cannot be prevented, training models using copyright-protected training data is not covered by the text and data mining exception."

Second, there was no lawful access. The defendant obtained the recordings by stream-ripping them from YouTube, circumventing the rolling cipher, a mechanism that generates a frequently changing hidden URL for the actual media file. The chamber classified that mechanism as an effective technological measure under section 95a of the German Copyright Act and, separately, under 17 U.S.C. 1201(a)(1)(A). It also found three overlapping usage reservations on the platform side: a robots.txt crawler prohibition constituting a machine-readable reservation, a scraping prohibition in the terms of use, and the rolling cipher itself.

The chamber placed the burden on the defendant to show from which sources each work was obtained and whether restrictions were declared there. On the EU AI Act, the judgment cites the Commission's practical guide directly, noting that adherence to the code of practice does not amount to compliance with Union copyright law.

The right that failed, and the one that caught it

Not every claim succeeded. The chamber rejected infringement of the making-available right under section 19a of the German Copyright Act, and the 176-prompt figure is why. If a retrieval requires up to 176 identical prompts, access does not occur at a time of the user's choosing, which is a defining element of that right.

The claim was instead upheld under the unnamed right of communication to the public in section 15(2), read in conformity with Article 3(1) of the InfoSoc Directive. The chamber held that the defendant performs the act of communication directly by providing the model, that quantitative public access is met because the service is open to anyone with an internet connection and already has more than one million paying subscribers, and that qualitative public access is met because training transforms the works into memorised works made available to a new audience through a distinguishable technical process.

The chamber also refused the hosting safe harbour in Article 6 of the Digital Services Act. The outputs are the defendant's own content, not third-party information passively transmitted, because the prompts were simple and open-ended and did not predetermine what the model produced.

A German court running the US four-factor test

The most unusual passage is the fair use analysis. Because the training took place in the United States, the chamber ascertained US law itself under section 293 of the German Code of Civil Procedure and ran the four-factor test under 17 U.S.C. 107.

On the first factor, the chamber found no transformative use, because the model does not create something new relative to the works used to train it but generates copies of them. On the second, the works are highly creative and the defendant exploited that creativity specifically. On the third, complete recordings were used. On the fourth, the outputs substitute for the originals in the market for light music, and the chamber placed the burden of proof on the defendant under US law rather than applying the German concept of a secondary burden on the plaintiff.

The chamber distinguished rather than defied the two US decisions on point. In Anthropic's case, the court had treated training as transformative precisely because no reproduction of the books appeared in the outputs, a point the judgment quotes at length. PPC Land reported in September 2025 that Anthropic agreed to a settlement of at least 1.5 billion dollars over pirated copies, after the same judge held that downloading pirated books was not fair use. In the Meta case, the court granted summary judgment on fair use for Llama training in June 2025, again in the absence of substantially similar outputs.

Here, the chamber wrote, recognisable infringing outputs have been identified, which distinguishes both cases. On market harm it added an arithmetic point: given a generator that recorded 2.21 million visitors in a single month in 2024, a significant market effect is evident even if only every fourth or every 176th prompt yields a substantially similar piece.

The judgment then tempers its own reach: "Not all models memorize their training data. The rejection of fair use in this specific case does not stand in the way of progress or competition, nor does it conflict with the public interest."

For the injunction itself the chamber applied the eBay standard, the four-part US equity test, reasoning that German civil law is familiar with equity through section 242 of the Civil Code. The same provision was used by analogy to close a conflict-of-laws gap on the information claim.

No referral to Luxembourg

The defendant had asked, in a brief dated 26 November 2025, for a preliminary reference to the Court of Justice of the European Union under Article 267(2) TFEU. The chamber declined, holding that existing guidance on reproduction and communication to the public under the InfoSoc Directive is sufficient, and that no interpretation of the text and data mining exception is needed because the facts fall outside it.

The judgment is not final and can be appealed.

Why this matters for the marketing community

The direct commercial exposure sits with anyone who ships generative output into paid media. Dr. Ursula Feindor-Schmidt, a copyright, media and AI specialist at the Munich firm LAUSEN who published an analysis of the full text, put the consequence plainly: "anyone offering generative models in Europe will not get around licensing and cannot ignore the European legal position."

Three threads in PPC Land's coverage converge on this ruling.

The first is the divergence between jurisdictions on whether model weights contain works at all. In November 2025 the High Court of England and Wales found that Stable Diffusion does not store copies of training images in its weights, a factual finding that pointed the opposite way from the Munich chamber's conclusion. Two European courts, two readings of the same technical question, and no appellate answer yet in either.

The second is compliance theatre. The judgment's treatment of the EU AI Act code of practice removes a defence that model providers have signalled they would rely on. PPC Land covered the publication of the final General-Purpose AI Code of Practice in July 2025 and the Commission guidelines that accompanied it, which European creative sector bodies rejected as inadequate days later. The Munich chamber has now confirmed the creators' reading: signing the code settles nothing about copyright liability.

The third is data acquisition. The rolling cipher finding turns scraping method into a licensing question, because circumvention destroys the lawful access precondition for the text and data mining exception. That sits alongside the European Data Protection Board's July 2026 guidelines on web scraping for generative AI, which narrowed the data protection routes to the same training material, and the European Commission's antitrust probe into how Google uses publisher and YouTube content for AI, opened in December 2025. Copyright, data protection and competition law are now pressing on the same pipeline from three directions.

Munich is becoming the venue where these questions get answered. The same court held Google directly liable for defamatory AI Overviews output in May 2026, with written reasoning published in July that stripped away the host privilege on three grounds. The common thread across both rulings is authorship: where a system generates new text or new audio, the operator answers for it.

For advertisers and agencies commissioning AI-generated audio, the practical effect is a provenance problem rather than an abstract legal one. An asset generated by a model trained on unlicensed repertoire may be recognisable as an adaptation of a protected work, and section 23(1) of the German Copyright Act reaches adaptations, not only identical copies. The publication order in the Süddeutsche Zeitung guarantees that the finding reaches beyond trade press, once the judgment becomes final.

Timeline

Summary

Who: The Munich I Regional Court's 42nd Civil Chamber, ruling on claims brought by a German collecting society against the United States operator of the Suno AI music generator. Dr. Ursula Feindor-Schmidt of the Munich firm LAUSEN published an analysis of the full text.

What: A final judgment in case 42 O 763/25 granting an injunction against training reproductions in the United States, storage of works inside model parameters, provision of the model in Germany and distribution of arrangements, backed by fines of up to 250,000 euros per violation, plus information and damages claims covering three works and one chorus since 1 July 2023, publication of the operative part in a national newspaper, and 5,049.70 euros in pre-trial fees. The making-available claim under section 19a was rejected; the unnamed right of communication to the public succeeded instead. The chamber applied US law to the training and rejected fair use.

When: The judgment was issued on 31 July 2026 after an oral hearing on 9 March 2026. The 143-page English translation circulated on 3 August 2026. The decision is not final and can be appealed.

Where: Munich, Germany, with extraterritorial reach over training conducted in the United States and effects across the European Union market.

Why: The chamber held that memorisation of training data inside model parameters constitutes a reproduction, that the text and data mining exception in section 44b of the German Copyright Act does not cover it, that circumventing YouTube's rolling cipher destroyed the lawful access precondition for that exception, and that recognisable infringing outputs distinguish the case from the two US decisions finding fair use in AI training.