Product1 publisher3 min readPublished
Sony and UMG target the user preference data behind Suno's licensed v6 models
The labels read Suno's own disclosure that v6 learned from user interactions as an admission that the outputs of its earlier scraped models fed the new one. They put the count at 60,202 recordings and the demand at up to $9 billion.
The Product Desk · Product desk
What happened
- Sony Music Entertainment and Universal Music Group sued Suno over its newest v6 models, the first the company released after signing licensing deals with Warner Music Group and BMG, Variety reports.
- The complaint, obtained by Music Business Worldwide, fastens on Suno's own public statement that v6 was trained in part on users' interactions with the service.
- Reading those interactions as outputs and preference signals from earlier unlicensed models, the labels put the infringement at no fewer than 60,202 sound recordings and up to $9 billion in damages.
Compiled by The Product DeskSomething wrong?How this is made
Why it matters
- exposure Any team whose current model learned from its predecessor's outputs or from user ratings of those outputs has a provenance chain that runs back past the date its licensing deals were signed.
- constraint Publishing an honest description of the training mix becomes a legal liability: the phrase the labels quote back at Suno is Suno's own disclosure about learning from its community.
- decision Because the demand scales at roughly $149,500 per recording, the defence effort goes into shrinking the list of works the court accepts, not into arguing the per-work number.
- precedent If the court accepts that derived signals inherit the upstream claim, retiring an old model stops being a clean break for any product built on distillation or self-generated data.
A user who typed a prompt into Suno two years ago, listened to two takes and clicked the one they liked better was rating a product. Sony Music and UMG now describe that class of data as "the outputs of, and preference signals derived from, prior models that Suno built by copying Plaintiffs' recordings without authorization" [5]. The sentence they are reading is Suno's own. The company has said publicly that the v6 models were also trained on "users' 'interactions' with Suno's service" [4].
Preference data is the cheapest quality signal a generative product has. It arrives with usage, and nobody has to clear it. The labels' theory is that when the thing being rated was produced by a model built on their recordings, the rating carries the copying forward into whatever gets trained next [5].
Suno's answer is that the claims "remain fundamentally flawed on both the facts and the law," the company said in a statement to Engadget [9]. It described the inputs as "content licensed from our partners, interactions including creations and preference signals from our community, and the accumulated learnings from our team" [10]. Both sides describe the same three-part pipeline, and the dispute is over where the middle part came from.
The damages figure moves with the recording count and nothing else. The labels put the count at no fewer than 60,202 sound recordings [6] and the exposure at up to $9 billion under US copyright law [7]. Divide the second by the first and each recording on that list is worth about $149,500 in the claim [16]. Separately, MusicBusinessWorldwide reports the complaint seeks up to $2,500 for each time Suno circumvented YouTube's anti-downloading technology [8].
What Suno shipped with v6 was a licensed model, launched "in partnership with WMG, BMG, and Believe," the company said [17], and support for the older models ended at the same time [15]. Ending support removes those models from the product and leaves their outputs in the pipeline if the outputs became training signal. The earlier corpus is already on the record: a July 2026 hack of Suno's data showed millions of songs and lyrics scraped from platforms including YouTube Music, Deezer and Genius [13].
For anyone building on a distilled or self-generated data pipeline, the audit this suit implies is one sentence per input, naming where that input came from. When the answer names a model you trained, the audit continues into that model's inputs. The sort runs two ways: licensed or unlicensed at source, and direct data or derived signal, where derived covers outputs, ratings, distillation targets and synthetic sets. Licensed-and-direct sits outside what Sony and UMG are arguing about: Suno used licensed material for the v6 lineup [3]. The contested cell is derived signal whose source model was unlicensed. Suno has not said what share of v6's training signal came from interactions with pre-licensing models. A team whose derived-signal cell is small can say so with a number.
What to watch
- Whether the derivation theory survives a motion to dismiss, since the 60,202-recording count rests entirely on it.
- Whether Suno discloses what share of v6's training signal came from interactions with its pre-licensing models.
- Whether Warner Music Group, a plaintiff in the earlier suit and now a v6 licensor, joins this second complaint or stays out.