Gerencia 360 asked a California federal court on August 31, 2026, to impound or destroy unauthorized copies it says remain inside Suno’s model weights. The request reaches beyond scraped audio files and training datasets. If granted as pleaded, it could reach part of the trained system that makes Suno's music generator function.
The complaint invokes Section 503 of the Copyright Act and describes unauthorized copies as existing in training datasets, internal repositories, model weights, and server infrastructure. That wording makes Gerencia 360’s requested destruction of Suno model weights one of the lawsuit’s more aggressive remedies, because deleting source files is very different from disabling or rebuilding a trained model.
Section 503 does give federal courts substantial power, but Gerencia still has to connect the property it wants destroyed to copyright infringement. The statute does not say that anything produced during an infringing process can automatically be seized simply because it sits downstream from the original copy.
Destruction comes later. As part of a final judgment, a court may order destruction or another reasonable disposition of copies found to have been made or used in violation of the copyright owner's rights, along with articles by which those copies may be reproduced. That creates a basic proof problem for Gerencia if it wants the remedy to reach weights rather than ordinary stored music files.
A scraped song sitting in a training repository fits the familiar idea of a digital copy much more comfortably. A trained model consists of numerical parameters adjusted through training, and the disputed question is whether protectable expression from particular works remains fixed within those parameters in a legally meaningful form.
The distinction is not just technical housekeeping. Gerencia would have a stronger Section 503 argument if it can show that Suno's model reliably retains and reproduces substantial protected expression from identified songs, rather than merely showing those songs were used somewhere in the training process.
That is why model memorization evidence could become central. If a model can reproduce a training example verbatim or substantially similarly without the protected material being supplied in the prompt, the Copyright Office says there is a strong argument that the expression exists in some form within the weights.
A Northern District of California ruling in Andersen v. Stability AI points in the same direction without deciding the ultimate issue. At the pleading stage, the court allowed allegations to proceed that protected works could remain inside Stable Diffusion as algorithmic or mathematical representations. It did not hold that every trained model is an infringing copy.
That limitation matters for Suno. Gerencia cannot safely jump from "our songs were copied for training" to "the entire model is an unauthorized copy of our songs." Evidence about reproducible melodies, lyrics, arrangements, memorization rates, model versions, and the relationship between those outputs and identified Gerencia works could determine whether the weights themselves fall within the remedy.
The UK High Court reached the opposite result with Stable Diffusion in Getty Images v. Stability AI in 2025. It held that the final Stable Diffusion weights were not infringing copies because they did not contain or store Getty's copyrighted images, even though copyrighted works had been reproduced during training. That ruling applies British law and does not control a California court, but the factual divide is useful.
Both outcomes turn attention toward what the model actually retains. A model proven to reproduce protected expression from its parameters presents a different remedial problem from a model whose weights record learned relationships without storing any legally cognizable copy of a claimant's work.
Gerencia's request therefore reaches further than deleting a dataset. To force destruction of Suno's trained weights under Section 503, the label would likely need a persuasive factual and legal showing that the weights themselves contain infringing copies, or qualify as articles by means of which infringing copies can be reproduced. If that showing fails, a court could still order narrower relief against source files, repositories, specific infringing copies, or future conduct without dismantling the trained model itself.
The complaint invokes Section 503 of the Copyright Act and describes unauthorized copies as existing in training datasets, internal repositories, model weights, and server infrastructure. That wording makes Gerencia 360’s requested destruction of Suno model weights one of the lawsuit’s more aggressive remedies, because deleting source files is very different from disabling or rebuilding a trained model.
Section 503 does give federal courts substantial power, but Gerencia still has to connect the property it wants destroyed to copyright infringement. The statute does not say that anything produced during an infringing process can automatically be seized simply because it sits downstream from the original copy.
Section 503 separates impoundment from destruction
During a pending copyright case, Section 503 allows a court to impound copies or phonorecords claimed to have been made or used unlawfully. It also reaches certain articles used to reproduce those copies. The court controls the terms, which means impoundment is discretionary rather than an automatic consequence of filing an infringement claim.Destruction comes later. As part of a final judgment, a court may order destruction or another reasonable disposition of copies found to have been made or used in violation of the copyright owner's rights, along with articles by which those copies may be reproduced. That creates a basic proof problem for Gerencia if it wants the remedy to reach weights rather than ordinary stored music files.
A scraped song sitting in a training repository fits the familiar idea of a digital copy much more comfortably. A trained model consists of numerical parameters adjusted through training, and the disputed question is whether protectable expression from particular works remains fixed within those parameters in a legally meaningful form.
The distinction is not just technical housekeeping. Gerencia would have a stronger Section 503 argument if it can show that Suno's model reliably retains and reproduces substantial protected expression from identified songs, rather than merely showing those songs were used somewhere in the training process.
Memorization could make the weights legally important
The U.S. Copyright Office has already drawn that line in its generative AI training report. It said model weights may contain copies when training causes a model to memorize protected material, and copying those weights can then implicate the reproduction right. The important condition is retention of protectable expression, not simply exposure to copyrighted works.That is why model memorization evidence could become central. If a model can reproduce a training example verbatim or substantially similarly without the protected material being supplied in the prompt, the Copyright Office says there is a strong argument that the expression exists in some form within the weights.
A Northern District of California ruling in Andersen v. Stability AI points in the same direction without deciding the ultimate issue. At the pleading stage, the court allowed allegations to proceed that protected works could remain inside Stable Diffusion as algorithmic or mathematical representations. It did not hold that every trained model is an infringing copy.
That limitation matters for Suno. Gerencia cannot safely jump from "our songs were copied for training" to "the entire model is an unauthorized copy of our songs." Evidence about reproducible melodies, lyrics, arrangements, memorization rates, model versions, and the relationship between those outputs and identified Gerencia works could determine whether the weights themselves fall within the remedy.
Foreign rulings show why the factual record matters
Courts outside the United States have already split sharply on closely related arguments. In July 2026, a Munich court ruled against Suno in GEMA's case and treated six protected musical works as stored in model parameters or other data structures, relying on evidence that the models could reproduce them. The ruling restrained specified reproduction and model-related uses under German law.The UK High Court reached the opposite result with Stable Diffusion in Getty Images v. Stability AI in 2025. It held that the final Stable Diffusion weights were not infringing copies because they did not contain or store Getty's copyrighted images, even though copyrighted works had been reproduced during training. That ruling applies British law and does not control a California court, but the factual divide is useful.
Both outcomes turn attention toward what the model actually retains. A model proven to reproduce protected expression from its parameters presents a different remedial problem from a model whose weights record learned relationships without storing any legally cognizable copy of a claimant's work.
Gerencia's request therefore reaches further than deleting a dataset. To force destruction of Suno's trained weights under Section 503, the label would likely need a persuasive factual and legal showing that the weights themselves contain infringing copies, or qualify as articles by means of which infringing copies can be reproduced. If that showing fails, a court could still order narrower relief against source files, repositories, specific infringing copies, or future conduct without dismantling the trained model itself.
