Sony and UMG sue Suno over 60,202 recordings

What this is about

Sony Music Entertainment and Universal Music Group filed a new lawsuit against Suno in federal court in Massachusetts on September 18, 2026. The 45-page complaint targets Suno's new v6 model family and identifies 60,202 sound recordings the labels say were infringed. Suno rejects the allegations as factually and legally flawed.

The dispute goes beyond the familiar question of whether copyrighted music may be used for training. Sony and UMG now argue that a new model can inherit tainted characteristics when it learns from outputs, user interactions, or preference signals produced by earlier models. That makes the case relevant to any company that refreshes an AI system through distillation or synthetic data.

What the lawsuit actually claims

Suno released v6 with licensed content from Warner Music Group, BMG, and Believe. Older models at the center of litigation dating from 2024 were due to lose support. According to the new complaint, that break is insufficient because v6 was built in part from outputs and signals originating in the earlier systems.

The labels describe this as a continuation of the original infringement. They also point to 60,202 identified recordings and say this is only a small portion of the affected catalog. Suno says v6 was trained on licensed partner content, community interactions, and accumulated learning from its team. The court has not determined which legal interpretation is correct.

Why it matters

If the plaintiffs' argument succeeds, signing new licenses may not be enough to cleanse a model lineage. Providers could have to document more precisely which data, model outputs, and preference signals entered every generation. That would affect not only music systems but also language, image, and coding models that inherit knowledge from predecessors.

The figure of 60,202 recordings is an allegation, not a judicial finding. It still raises financial pressure because US copyright law can permit statutory damages for established infringements. Simply multiplying the count by the maximum rate would be misleading: the court first has to decide liability, scope, and how individual works or acts should be grouped.

In plain language

Imagine a bakery that used someone else's recipe without permission for years. It later buys new ingredients under contract and makes a new dough, but mixes in sourdough starter from the old batch. The central question becomes whether the new bread is legally clean because the fresh ingredients are licensed, or whether the old starter carries the original violation forward. That chain is what the parties are fighting over.

A practical example

A music service trains Model A on one million tracks. It later licenses 600,000 tracks and builds Model B. Instead of starting from zero, Model B absorbs ten million rated outputs from Model A so that user preferences are preserved.

If 60,202 protected recordings were demonstrably used in Model A, a court would have to determine whether the transferred outputs or preference data carry protected features. Training logs, dataset versions, deletion records, and independent audits would therefore matter more to the operator than a broad claim that the new model was trained on licensed data.

Scope and limits

First, the 60,202 recordings and the alleged inheritance are still claims made by the plaintiffs. Second, Suno's exact v6 training method is not fully documented in public; terms such as accumulated learning can cover several technical processes. Third, a US judgment would not automatically control Germany or the EU, where copyright exceptions and transparency duties differ.

The case also will not decide in the abstract whether all generative training is lawful or unlawful. It concerns specific works, specific model generations, and whether a shift to licensed data removes earlier risks. Liability and damages remain unresolved until the court rules.

SEO & GEO keywords

Suno v6, Sony Music, Universal Music Group, UMG, AI music, copyright, model training, synthetic data, model distillation, music labels, Massachusetts, generative AI