When does learning become memorizing?

AI & Society Sep 9, 2026

The GEMA v. Suno ruling doesn't ban AI-generated music. It asks a sharper question — and gives a concrete answer for six songs.

Six songs were at the center of a German copyright ruling issued on 31 July 2026. Rasputin. Forever Young. Atemlos durch die Nacht. Daddy Cool. Big in Japan. The chorus of Mambo No. 5. You probably recognize them. So, the Landgericht München I found, did Suno's model — a little too well.

The ruling from Germany's Regional Court Munich I (case 42 O 763/25) is the most detailed court decision yet on what it means for a generative AI model to "know" a song. Most early coverage collapsed it into two camps: either a landmark ban on AI music generation, or a narrow German procedural decision with no global reach. Neither reading is quite right. The case turns on a distinction that will matter far beyond Munich — and far beyond music.

Three layers, three legal questions

What makes the GEMA ruling unusual is that the court refused to treat AI training as a single event. Instead, it analyzed three distinct actions and applied separate legal reasoning to each.

The first is what happened in the United States. Suno obtained the six compositions by stream-ripping them from YouTube, bypassing what the judgment explicitly calls the "Rolling Cipher" — a YouTube mechanism that routes audio through a hidden URL generated by a frequently-rotating algorithm, separate from the visible page URL, specifically designed to prevent bulk downloading. The German court applied US law to this act — and found that fair use did not cover it under these circumstances. That conclusion is striking, but its authority is limited: it is a German court interpreting American law, not a US federal ruling. The dozens of American cases involving AI and music copyright move forward on their own track.

The second layer is what happened inside Suno's model in Germany. This is where the case gets technically interesting. The court did not accept the standard defense that a neural network only learns statistical patterns — that it absorbs structure, not content. Comparing the six original compositions with outputs generated by the model, the court found that the works had been stored reproducibly in the model's parameters or associated data structures. That reproducibility, not the training process in the abstract, was the basis for treating the model itself as containing unauthorized copies.

The third layer is the output. When GEMA's lawyers prompted the model using text, a title, and a genre description, the resulting audio reproduced elements that the court attributed to the originals — not to the prompts. Suno had argued that GEMA had manipulated the test conditions to engineer an infringing output. The court rejected this: melody, harmony, and rhythm were not fully determined by the text prompt alone. The model, not the prompter, made the musical choices that matched the source material. GEMA submitted 338 test prompts across the six works in total, selecting the most demonstrative output per song — evidence that the court was evaluating a consistent pattern rather than an isolated anomaly.

Spectrogram of AI-generated 1980s synthesiser pop music showing frequency bands and harmonic patterns as a colorful horizontal heatmap across time — A spectrogram of AI-generated 1980s synthesiser pop music created with Riffusion — the same genre as several of the six songs at the center of the GEMA v. Suno case. The horizontal axis is time; the vertical axis is frequency. Each colored band represents a harmonic layer of the generated audio. (image: Wikimedia Commons, Benlisquare, CC BY-SA 4.0)

What "memorization" means legally

A neural network trained on a hundred million songs doesn't store those songs in any obvious sense. It encodes relationships between musical features — patterns of pitch, rhythm, chord progression — at a level of abstraction that makes it genuinely difficult to point to where any individual song "lives" in the model's weights. We explored a related tension — what AI actually does to a specific, beloved song — in our earlier piece on what AI does to the song that united Europe for seventy years.

The Munich court didn't resolve the general question of whether model weights constitute copies. It resolved a specific factual one: whether these model weights, derived from these six works using this extraction method, produced outputs that reproduced recognizable content from the originals. The answer was yes — and that was enough.

This distinction matters for what comes next. The ruling doesn't establish that every model trained on protected music is infringing by definition. It establishes a test: can a court, by comparing an original with model output, demonstrate traceable reproduction? If so, the model doesn't get a pass just because the mechanism is statistical rather than literal.

What the ruling doesn't say

It doesn't say that EU law prohibits AI companies from training on copyrighted music. The text-and-data-mining exception in Article 4 of the DSM Directive — and its German implementation in §44b of the Urheberrechtsgesetz — explicitly allows mining for lawfully accessible works, unless rights holders have reserved their works through a machine-readable opt-out.

What the ruling says is that neither provision helped Suno here. The access was not lawful: stream-ripping YouTube with circumvention of a technical measure sits outside the definition of "lawful access." And even if access had been clean, the reproducible storage of recognizable content crossed the line between permitted analysis and prohibited reproduction. The ruling adds an unexpected secondary argument: the court found that the Rolling Cipher itself carries the legal weight of a machine-readable TDM reservation under §44b UrhG — meaning YouTube had declared the opt-out through three parallel channels simultaneously: its robots.txt file, its terms of service, and the technical protection mechanism.

The judgment is also not final. Suno was reportedly considering an appeal. The Court of Justice of the European Union has not yet ruled on how the DSM Directive applies to generative models that demonstrably memorize training data. And the German reasoning on US fair use carries no weight in American federal courts.

What remains, after those caveats, is a first-instance ruling with unusually precise factual grounding — six named works, a documented extraction method, and tested outputs — that gives rights holders a replicable investigative playbook.

The market has already started moving

Courts and markets often move in the same direction without coordinating. Before the Munich ruling, the music industry had already started replacing litigation with licensing.

In October 2025, Universal Music Group and Udio announced a settlement and a licensing agreement for a new platform built on authorized AI music generation. In November 2025, Warner Music Group and Suno announced a similar settlement and a partnership to develop models trained on licensed repertoire. Neither announcement acknowledged legal liability. Both acknowledged that licensing was now the business model. This shift connects to a broader repositioning: AI music platforms have been reframing themselves from tools to co-creators, and licensed repertoire is what makes that claim credible.

This shift doesn't vindicate the position that unlicensed training is fine. It reflects something more pragmatic: the market moves faster than litigation, and major platforms calculated that licensed infrastructure was a better long-term bet than fighting each case in court. For smaller AI music companies without the cash to settle or the catalog relationships to license, the calculus looks different.

Rights holders' organizations point out that deals between labels and AI platforms don't automatically resolve claims from individual composers, publishers, and performers. GEMA and its counterparts across Europe have been explicit: any licensing regime needs to be transparent, and the value flowing back to creators must be demonstrably fair — not a payment from platform to label that disappears before reaching the people who made the music.

The Dutch parallel

Germany's §44b UrhG and the Netherlands' Article 15o of the Copyright Act are both implementations of DSM Article 4. The conditions are functionally similar: lawful access is required, and rights holders can exclude their works by making a machine-readable reservation.

BumaStemra, the Dutch performing rights organization, stated in 2024 that it had exercised this opt-out right for its repertoire, making prior authorization a condition for mining BumaStemra-administered works for AI development. Its response to the Munich ruling was measured: the German decision is significant, but it is not a Dutch ruling and does not create binding precedent in the Netherlands.

The practical implication for Dutch AI companies is not that GEMA v. Suno applies to them directly. It is that a Dutch court asked to evaluate an analogous claim — mining after a valid opt-out, or outputs that demonstrably reproduce BumaStemra-licensed material — would be deciding under legislation with the same architecture and could take the Munich reasoning seriously.

No separate Dutch case against Suno or Udio appears in the public record.

The question that remains

The GEMA v. Suno ruling doesn't settle whether AI can train on music. Courts, legislators, and rights holders will keep arguing about that for years. What the ruling settles, for six songs and one model in one jurisdiction, is a narrower question: when does a model's exposure to music cross from learning patterns into storing and replaying content?

The answer the Munich court gave is operational rather than philosophical. It looked at outputs. It compared them to originals. It found traceable content. That test can be run again, in other jurisdictions, on other models and other repertoires.

The music industry's real argument was never about whether AI can analyze music. It was about who pays when a model stops analyzing and starts performing.