Sony and Universal Expand Their Fight With Suno Over AI Music Training

Updated September 26, 2026 4:23 pm

Sony Music and Universal Music Group have launched a fresh legal attack on AI music company Suno, alleging that its newest v6 model is still built on copyrighted recordings and that the startup is trying to obscure that history by training the system on outputs from earlier infringing models. The new complaint matters because it could help define whether AI companies can legally “cleanse” tainted training data by reusing the results of models built on unlicensed music.

The lawsuit adds a new layer to one of the music industry’s most closely watched copyright fights. Rather than targeting only Suno’s alleged use of songs scraped from YouTube and other sources, the labels now say the company’s latest model is a product of what they describe as “model laundering” — a process in which one infringing system produces data that is then fed into another, supposedly newer model.

What Sony and UMG are accusing Suno of

At the center of the case is Suno’s recently released v6 model. According to the labels, the company did not truly start from scratch when it trained that system. Instead, they argue, it relied on outputs created by older Suno models that were themselves trained on unauthorized copies of recorded music.

That distinction is crucial. If the older models were built on infringing data, the labels argue, then the newer model still carries the legal and commercial benefits of those copies, even if the raw music files were not fed directly into v6.

“Training a ‘new’ model on the outputs of an infringing model does not eliminate the infringement,” the complaint argues in substance, adding that the process simply moves the value of the copyrighted works from the original recordings into the later models and their outputs.

The labels say that Suno’s approach does not represent a clean restart. Instead, they contend, the latest model is still derived from the same underlying material they believe was copied without permission.

Why this lawsuit is different

This new filing goes beyond the earlier dispute over direct training on copyrighted music. The record labels are trying to close what they see as a potential loophole: if a company is accused of using infringing data once, can it later point to a different model and claim the new version is legally separate?

Sony and UMG say the answer is no. Their theory is that Suno may have attempted to hide or dilute the provenance of its training data by using outputs from predecessor models, a method they describe as distillation. In practical terms, that means the company may have taught a new model to imitate the behavior of an older system that already absorbed value from copyrighted recordings.

If a court accepts that reasoning, it could have implications well beyond Suno. Other AI companies that train successive generations of models on synthetic or machine-generated outputs may also face questions about whether that process breaks the chain of infringement or simply extends it.

How the labels describe “model laundering”

The phrase “model laundering” is the labels’ shorthand for what they see as a legal wash cycle: an allegedly infringing model produces outputs, those outputs become training material for a successor model, and the company then presents the successor as a fresh product.

In the complaint, the music companies argue that this does not erase the original copying. Instead, they say, the copyrighted expression remains embedded in the model pipeline and continues to shape the newer system’s behavior. That means, from their perspective, the new model is still downstream of unauthorized use.

The case language is designed to do more than just win a lawsuit. It also appears aimed at setting a broader precedent for how courts should evaluate model-to-model training in the age of generative AI.

What is distillation and why does it matter?

Distillation is a training technique in which a new model learns by imitating the outputs of a larger or earlier “teacher” model. In Suno’s case, the labels say this method was used to help train v6 to reproduce the performance of previous systems.

The legal issue is not distillation itself, but the source of the teacher model’s knowledge. If that teacher was trained on copyrighted recordings without permission, the labels argue that the student model inherits the benefits of that disputed training.

That argument raises an important question for AI developers: can a company avoid copyright liability simply by using model outputs instead of direct source material? Sony and UMG say no, at least not when the original model was allegedly built on unauthorized copies.

What Suno has said about v6

When v6 was introduced, Suno’s Jack Brody said the model had been trained “from the ground up” on a new dataset that included user data. The company did not publicly provide a detailed accounting of what that dataset contained or how much of it came from user contributions versus other sources.

Later, Suno confirmed that the data included user “creations,” but it stopped short of explaining how those creations were used or how the training pipeline avoided the concerns raised by the major labels. The company also did not offer a substantive public response to the latest complaint in the material shared with The Verge.

That gap in detail has become part of the controversy. In the absence of a transparent data audit, the labels are asking the court to infer that the company’s new system remains dependent on the same disputed foundation.

Who is suing Suno, and why does it matter?

Sony Music and Universal Music Group are two of the most powerful players in the global recording industry. Their decision to sue again matters because it shows that the dispute with Suno is not simply about one product launch or one technical method. It is about whether the largest labels believe AI music companies can be allowed to train on copyrighted recordings without a license and then iterate their way into legality.

The two companies are also notable because they have not signed licensing agreements with Suno. That makes them holdouts in a market where some AI music providers have tried to negotiate access rather than litigate for it.

For the labels, the stakes are straightforward: if they do not challenge this practice now, they risk establishing a business model in which AI tools can absorb the creative value of recorded music without paying the artists, songwriters, labels, or publishers who helped create it.

How the dispute fits into the wider AI music fight

This lawsuit is part of a much larger battle over generative AI and copyright. Across text, image, code, and music generation, rights holders have increasingly pushed back on the use of protected material in training data, especially when companies describe the process as fair, transformative, or technically difficult to trace.

Music may be particularly vulnerable to this kind of conflict because high-quality recordings are easy to ingest at scale, and because vocal style, composition, instrumentation, and production patterns can be captured in ways that make outputs sound eerily close to known artists.

That has left courts, regulators, and lawmakers with a difficult task. They must decide whether training on copyrighted works is a form of copy-making, whether model outputs can contain protected expression, and whether successive rounds of AI training can cleanse a model of earlier legal problems.

Why the outcome could shape the industry

If the labels succeed, AI music companies may be forced to show much more clearly where their training data comes from, how model generations are connected, and whether “new” systems are truly new. They may also face a stronger push toward licensing deals with major rightsholders.

If Suno prevails, the ruling could strengthen the argument that model evolution, synthetic outputs, or student-teacher training pipelines create enough distance from original works to weaken infringement claims.

Either way, the case is likely to be read closely by developers, labels, publishers, and streaming platforms trying to understand what kinds of AI music tools can survive under current copyright law.

Timeline of the Suno–major labels conflict

Key arguments in the complaint

The labels’ case rests on several interconnected claims. Together, they form a broader argument that Suno’s latest model cannot be treated as a clean, independent release.

• The original Suno models were allegedly trained on copyrighted music without licenses.
• Those models generated outputs that were then reused in training v6.
• That process, according to the labels, does not erase the original infringement.
• The company may have used distillation to replicate prior models’ behavior.
• Even indirect use of unauthorized copies can still benefit from the copyrighted works, the labels argue.

The complaint’s larger message is that data provenance matters. Once a model is built on contested material, the legal taint does not disappear simply because the company creates another model on top of it.

How might courts view AI model laundering?

Courts will likely have to decide whether the law treats model outputs as a meaningful break from the source material or as a continuation of the same copying chain. That question is especially difficult in generative AI, where the boundary between training data, intermediate outputs, and final content is often blurred.

Judges may also want more technical evidence about how Suno assembled v6, what percentage of the training set came from user creations, whether those creations were fully owned by users, and whether any earlier infringing material could still be reconstructed from the model’s behavior.

That evidence will matter because copyright cases involving AI often turn on details that are invisible to the public. The labels are essentially asking the court to look beneath the product launch language and trace the lineage of the model itself.

Why transparency is now part of the legal fight

One of the most contested issues in AI copyright cases is information asymmetry. Companies generally know what data they used, but outside parties often do not. That makes it difficult for labels, artists, and courts to assess whether a model is clean or contaminated.

In this case, the lack of detailed public explanation around v6 has become a key part of the story. Suno has said the model was newly trained, but the labels want more than a broad statement. They want proof that the system was not built on the residual output of earlier, allegedly infringing models.

That demand is likely to remain central as more AI companies try to market “new” foundation models while maintaining continuity with prior systems.

What happens next?

The next phase will likely focus on discovery, technical analysis, and legal arguments about whether Suno’s latest model is sufficiently distinct from its predecessors to avoid the original infringement claims. Depending on how the case develops, the court may also have to decide whether the use of model outputs as training data changes the legal analysis at all.

For now, Sony and UMG are signaling that they do not believe a technical reset is enough. Their position is that if the first model was built on unauthorized recordings, the later one still carries the imprint of that decision.

That could make this case one of the most important music-and-AI copyright disputes yet. It asks not only whether an AI company copied protected songs, but also whether it can ever wash those copied inputs out of the system by building another model on top of them.

In an industry where licensing negotiations, artist compensation, and AI innovation are all colliding at once, the answer may help determine how the next generation of music tools gets built.

Bottom line

Sony and Universal’s new lawsuit says Suno’s v6 music model is still tainted by earlier alleged infringement because it was trained on outputs from prior models that were themselves built on unlicensed recordings. The case could help set a major precedent for whether AI companies can legally reuse the outputs of disputed models to launch supposedly cleaner successors.

As the fight continues, the labels are trying to make one point unmistakably clear: in their view, you cannot wash copyright problems away by simply building a newer AI model on top of an older one.