Model Laundering: Can Retraining AI Reset Copyright?

Short answer: Probably not — but no court has decided yet. Sony Music Entertainment and Universal Music Group sued AI music platform Suno on September 18, 2026, arguing that its newest model, v6, was trained on the outputs of earlier models that learned from unlicensed recordings. The labels call that practice model laundering: retraining does not clean up what the first model took.

What Sony and Universal Allege

The complaint was filed in the U.S. District Court for the District of Massachusetts, the same court hearing the labels’ original 2024 case. It names 60,202 sound recordings that Sony and Universal say were copied without permission to develop Suno’s technology.

The timing is the story. Suno launched v6, v6 Wild, and v6 Mini on September 9, 2026 — nine days before the suit — having spent the previous year signing licenses with Warner Music Group (November 2025), BMG (August 2026), and Believe with its TuneCore platform (September 8, 2026). Sony and Universal never signed. Their filing is the answer to Suno’s argument that a newly licensed model draws a line under the past.

Two theories sit at the center of it:

• Inheritance through outputs. v6 was allegedly trained partly on music generated by Suno’s earlier models, so whatever those models absorbed from unlicensed recordings passed forward into v6.
• Knowledge distillation. v6 was allegedly trained as a “student” model to reproduce the behavior of earlier “teacher” models, letting it replicate what the teacher knew even without seeing the original recordings.

The complaint’s framing is blunt: training a new model on the outputs of an infringing model “does not eliminate the infringement; it launders it,” and v6 “is not a fresh start; it is the fruit of the same poisoned tree.”

On remedies, the labels rely on statutory damages of up to $150,000 per work for willful infringement — a theoretical ceiling above $9 billion across 60,202 recordings, a figure reported by industry press rather than awarded by any court. They also seek damages for circumventing YouTube’s anti-downloading protections, which under the DMCA can reach $2,500 per act.

What Model Laundering Means

Model laundering is an argument, not a statute or a doctrine. It borrows the logic of money laundering: pass the proceeds through enough intermediaries and the origin does not change. The proceeds here are musical expression, the intermediaries are successive model generations, and the origin is unlicensed copying.

The practical question is narrower than the rhetoric. A model is not a database — it does not simply store the songs it was trained on. The labels must show that something recognizably derived from the original recordings survives into v6, through training on generated audio, distillation, or retained copies of the source. That turns on evidence and model provenance, and it has not been litigated in the United States.

The stakes are symmetric. If retraining on a model’s own outputs resets liability, developers gain an obvious route around licensing. If it does not, a model’s entire lineage becomes relevant to infringement, and provenance becomes a record companies must be able to produce.

Can a Company Retrain Its Way Out of a Copyright Problem?

Under U.S. law the answer is unsettled, and this case is one of the first to test it directly. Four factors are likely to decide it:

• Whether copies persist. If earlier models retained unauthorized copies that remain recoverable in the new model, retraining does little. The Munich Regional Court took this approach in the GEMA v. Suno ruling, treating works a model can be prompted to reproduce as copies held inside it — see what that ruling means for artists.
• Purpose and substitution. Fair use under 17 U.S.C. § 107 turns heavily on whether the use serves the same market purpose as the original, and music generated as a substitute for existing recordings sits badly on that factor.
• A functioning licensing market. The labels point to Suno’s deals with Warner, BMG, and Believe as evidence that a market for training rights exists — an argument the complaint makes explicitly.
• Where training happened. Copies made during U.S. training are governed by U.S. law; outputs generated and stored in the EU fall under EU law. Munich applied both in one judgment.

Ownership is a separate question, covered in is AI music legal and who owns AI-generated music.

What Suno Says

Suno disputes the claims. In a September 22, 2026 filing it said v6 was built “from the ground up” on three inputs: content licensed from partners, community interactions including creations and preference signals, and accumulated internal learnings. It says uploaded audio was not part of the training data, while music created on the platform was.

Two details deserve attention. Suno’s terms of service, effective September 3, 2026, grant the company a broad license to use submitted and generated content to improve its models — without transferring ownership of a user’s original submissions. And Suno has not disclosed how much user-created music went into v6, or which recordings are involved.

The company calls the labels’ claims “fundamentally flawed on the facts and the law.” No court has ruled on any of it. Everything above is what each side alleges.

What This Means for Indie Artists

If your music could have been training data, whether a retrained model escapes liability decides whether rightsholders have leverage over a model’s later generations — and whether future licensing schemes have a legal foundation.

If you make music with AI tools, your exposure is more immediate. Platforms generally reserve broad rights over what you upload and generate, and a claim against the service does not disappear because a user pressed generate. EU disclosure duties under the AI Act apply regardless of who wins; see when and how to disclose AI use.

If you are evaluating a platform, ask three questions: does it license its training data, does it say what your uploads are used for, and can it document where its model came from?

What Comes Next

Suno has not answered the new complaint, and no trial date is set. The original 2024 case continues, with the central fair use fight not expected to resolve before 2027. The Munich ruling remains a first-instance judgment likely to travel through the German appellate system and possibly to the Court of Justice of the European Union.

The decisive issue is no longer only whether training on copyrighted music is infringement, but whether that question follows a model through every generation it spawns. For artists the practical answer holds: treat whether AI music is legal as an open question, and assume outputs from these systems carry unresolved legal risk.

Practical Steps

Practical Checklist

• Read the current platform terms before uploading stems, demos, or masters — Suno’s changed on September 3, 2026, and terms in this sector move quickly
• Register your copyrights with the U.S. Copyright Office; statutory damages and attorney’s fees generally require timely registration, which is what makes a claim worth bringing
• Keep an inventory of your catalog — titles, ISRCs, release dates, sample sources — so you can check it against any disclosed training dataset
• Separate the two questions you care about: whether your work was used for training, and whether you own what an AI tool outputs for you
• Watch your distribution agreement for clauses permitting AI training or synthetic use of your recordings

Common Misconception

“The labels signed licensing deals, so music coming out of the platform is cleared.” Those agreements cover specific catalogs under their own terms and dates. Sony and Universal did not sign, no deal addresses the earlier models’ training history, and a license to train is not a warranty that every output is non-infringing. Clearance attaches to specific rights, works, and uses — not to a platform.

When to Ask a Lawyer

• You believe your recordings were used to train an AI model and want to evaluate a claim
• You have released AI-generated music and received a claim, takedown, or distributor notice
• You are asked to license your catalog to an AI company and need to understand the training and output rights you would be granting
• Your distribution or label agreement covers synthetic media, AI training, or model inputs

Sources

1. Sony Music Entertainment and Universal Music Group v. Suno, complaint filed September 18, 2026, D. Mass. (copy published by Music Business Worldwide)
2. Munich Regional Court (LG München I), GEMA v. Suno, July 31, 2026 — case no. 42 O 763/25
3. 17 U.S.C. § 504(c) (statutory damages) and § 1203(c) (circumvention damages)
4. “Sony and UMG are suing Suno again,” The Verge, September 2026
5. Music Business Worldwide, September 2026 — the 60,202-recording tally, the v6 launch, and the Warner, BMG, and Believe agreements

Frequently Asked Questions

What is model laundering in AI copyright?

Model laundering is a label-industry term, not a legal doctrine: the claim that training a new AI model on the outputs of an older model built with unlicensed music carries the earlier copying forward instead of erasing it. No court has yet accepted or rejected it.

Can an AI company avoid copyright infringement by retraining its model?

Not automatically. Retraining may reduce exposure if the new model genuinely has no connection to unlicensed copies, but whether inheritance through model outputs counts as infringement is untested in U.S. courts.

Did Sony and Universal prove that Suno's v6 model used their music?

No. The complaint filed on September 18, 2026 is an allegation, not a finding. Suno disputes the claims as flawed on both the facts and the law, and no court has set a date to decide them.

Does using AI-generated music put me at risk if the model was trained on copyrighted songs?

It can. Commercial AI services have faced infringement claims over the music their models output, and platform terms usually reserve broad rights over what users upload and generate.

How much are Sony and Universal seeking from Suno?

The complaint names 60,202 recordings. At the U.S. statutory maximum of $150,000 per work for willful infringement, the theoretical ceiling exceeds $9 billion, though actual damages could be far lower and rest with the court.