Suno Lost in Germany. Canada Still Needs Its Own Rules
Matthew Leo · Published August 2, 2026 · AI
A German court has rejected the idea that an AI music company can treat recognizable reproductions of copyrighted songs as an accidental side effect of machine learning.
On July 31, the Munich Regional Court largely granted claims brought by German music-licensing organization GEMA against Suno. The case concerned six musical works that appeared in Suno’s training data and could be reproduced through its systems.
The judgment is important, but its boundaries matter just as much. It is not final, it does not establish that every Suno song infringes copyright, and it does not decide how a Canadian court would apply Canada’s Copyright Act.
What the court actually decided
The Munich court said the six works were reproducibly contained in versions 3.5 and 4 of Suno’s models. In the court’s analysis, the system had memorized protected musical content rather than merely learning abstract facts about rhythm or style.
GEMA generated the disputed outputs using prompts containing the original lyrics, musical style and work title. The prompts did not specify melodies, harmony, rhythm or arrangement, yet the court found recognizable original elements in the results.
The court held Suno responsible because the company selected the training material, designed and operated the models, and produced outputs substantially determined by those models. It also found that the relevant text-and-data-mining exception did not protect the reproductions identified in this case.
The dispute did not concern infringement by the lyrics themselves. That detail prevents the decision from being simplified into a general finding that entering any copyrighted words into an AI music service automatically creates liability.
The ruling is significant but not final
Reuters reported that Suno must disclose revenue connected to the infringement and may face damages that have not yet been quantified. Suno said it disagrees with the decision and is evaluating its options, including an appeal.
The court’s press release also states plainly that the judgment is not legally final. Any appeal could narrow, overturn or confirm the reasoning.
It would therefore be premature to call this the universal answer to AI training. The case involved specific songs, specific model versions, identifiable outputs and a particular combination of German and American copyright questions.
Its strongest lesson is narrower: when a model can reproduce recognizable protected material after simple prompts, calling the output “statistically generated” does not automatically end the copyright analysis.
Canada has been studying the same fault line
The Canadian government consulted creators, technology companies, academics and other stakeholders about generative AI and copyright between October 2023 and January 2024.
ISED’s consultation report identifies three unresolved policy areas: using protected works to train AI, deciding who owns AI-generated material and determining liability when outputs infringe.
Canadian creators generally argued that training should require consent and compensation. Technology-sector participants were more likely to support legal clarification or a text-and-data-mining exception that would permit broader training.
Those positions are difficult to reconcile because they involve different ideas about what training does. Creators see their work being copied into a commercial process. Developers often argue that models learn patterns rather than storing expressive works.
The Munich ruling addressed that factual dispute directly. For the six songs before it, the court concluded that memorization and reproduction had occurred.
Canadian policy needs to distinguish learning from extraction
Canada does not need to declare that every computational examination of a work requires the same licence. Research, indexing, accessibility tools and commercial generative systems can involve different purposes and risks.
It does need a clearer rule for systems that reproduce substantial protected expression. A company should not receive a broader practical right to distribute recognizable elements merely because a model stands between the source file and the output.
A workable Canadian framework should consider:
- Whether protected works were copied during training
- Whether rights holders reserved or granted permission
- Whether the model can reproduce recognizable expression
- How often extraction occurs under ordinary prompts
- Whether developers disclose meaningful information about training sources
- Who is responsible when the provider’s design makes infringement predictable
These questions are more useful than arguing over whether a model “creates” or “copies” in the abstract. A system can generate mostly original outputs while still memorizing particular works badly enough to create a legal and commercial problem.
Transparency must make licensing possible
Canada’s consultation found substantial support for training-data transparency. Rights holders cannot evaluate a licence, correction or legal claim when they have no practical way to learn whether their work entered a commercial model.
That does not require publishing every private or sensitive training record on an open website. It could require model providers to preserve source records, publish meaningful dataset summaries and offer a process through which a verified rights holder can investigate suspected use.
Licensing also needs to be workable for smaller developers. A system that requires a startup to negotiate individually with millions of creators would favour the largest technology companies. Collective licensing may provide one route, but it must not become an excuse to pay creators an amount too small to measure while calling the problem solved.
What Canadian musicians should preserve now
Musicians cannot write national policy by themselves, but they can keep evidence that becomes useful when a dispute arises. Preserve dated project files, session recordings, stems, release information, distributor records and contracts explaining who owns each part of a song.
When a suspicious output appears, record the complete prompt, model version, date, account and resulting audio. A social-media clip without the surrounding generation details may show similarity, but it provides less evidence about how the system produced it.
Artists using AI tools should perform the same discipline in reverse. Keep the service terms and licence that applied when the music was created, disclose meaningful AI involvement to collaborators and avoid promising exclusive rights that the contract or creative process may not support.
Canada should not wait for a foreign judgment to become policy
The Munich decision does not bind a Canadian court, and an appeal remains possible. Still, it turns a theoretical debate into a concrete finding: particular protected songs entered the training process, remained reproducible in the models and emerged in recognizable outputs.
Canada should now clarify when training requires permission, what transparency developers owe rights holders and who carries responsibility when a model reproduces protected expression.
Leaving those questions unanswered does not preserve innovation. It transfers the cost of uncertainty to musicians, developers and smaller companies least able to spend years resolving it in court.
Tags: Suno, AI music, copyright, Canadian musicians, generative AI