Suno admits it kept the training recordings, and v6 learned from its users' creations

If you've made songs on Suno, you're now part of a federal copyright case. On 20 Oct 2026 Suno filed its answer to the second lawsuit from Universal and Sony's labels. Two things in it matter most. Suno admits it has not deleted all the recordings it trained its models on. And it repeats, several times, that its v6 model learned from "creations and preference signals" from its community.

The first point is about the labels' recordings. The second is about you.

What did Suno admit?

The answer runs to 44 pages and was filed in Boston federal court (UMG Recordings v. Suno, No. 1:26-cv-14275). Across several paragraphs Suno admits it "has not deleted all of the sound recordings it has used to train its AI models". It adds each time that this is consistent with the labels' own request, made before the first lawsuit, that Suno keep the "data sets used to train" its models. Music Business Worldwide counts the admission six times and notes that the answer doesn't say Suno has stopped using the recordings.

Suno also repeats its public line that v6 "was trained on content licensed from our partners, interactions including creations and preference signals from our community, and the accumulated learnings from our team." It admits it "uses information about user responses to its service to improve its service" and that it "maintains certain data concerning user-generated outputs and their preference signals". It denies that every one of its models was trained on the same corpus of recordings.

Its main defence hasn't changed. Suno argues that any copying is fair use: a "back-end technological process, invisible to the public" in service of a new, non-infringing product. It asks the court to deny the labels any relief.

What the labels are actually arguing

The complaint, filed on 18 Sep, asserts 60,202 recordings. Its new idea is lineage. Suno's older models were trained on unlicensed recordings, the labels say. Users generated songs with those models and rated pairs of outputs. If v6 learned from those songs and ratings, then v6 learned from the older models' copying. The complaint calls this laundering and says v6 "is not a fresh start; it is the fruit of the same poisoned tree."

Several of these allegations are made "on information and belief", including that v6 was trained on synthetic audio from earlier models and built with knowledge distillation, where a new model learns to copy an older one. Suno says it lacks the information to respond to some of them and denies them on that basis. None of it has been proven.

The complaint quotes Suno CEO Mikey Shulman telling Billboard how the preference loop works. Users get two songs, implicitly pick the better one, and Suno uses that signal to steer its models. That is ordinary product design. It is also, by the labels' reading, a way of passing what the old models learned into the new one.

What this means for people who make music on Suno

Suno told MBW in September that the user data it means is preference data, not uploaded audio. MBW pointed out that this answers the question about uploads, not about generated tracks. "Creations" sits next to "preference signals" in Suno's wording, not inside it, and Suno hasn't said publicly which of your songs, edits or ratings went into v6.

The legal permission is already in place. Suno's Terms of Service give the company a worldwide, sublicensable, "perpetual, irrevocable" licence to your content, meaning both what you submit and what the service generates. That licence covers improving Suno's AI and machine-learning models. If you upload audio, you also promise Suno that nobody else needs to give permission or be paid for it to be used to train the models. Suno's privacy notice says it uses submissions, activity and other content "to train and enhance the models that power our Services".

As we read both documents on 23 Oct, neither describes a setting to opt out of training. The privacy notice offers deletion requests, which Suno says it may not be able to carry out in full. It also gives EEA residents a general right to object to processing. Neither says what happens to data that has already been used to train a model.

The answer also confirms a feature that matters here. Suno's "Custom Models" lets users upload their own audio to build a personal version of the model. Whether those uploads also feed the main models isn't stated in the answer.

What we think

The retention admission will get the headlines, but it is the weaker story. Suno says keeping the data is consistent with the labels' earlier request to retain training data sets. That is a point worth weighing, though the answer doesn't say Suno has stopped using the recordings. The more important point is the one Suno makes with pride: v6 learned from its users.

The labels' complaint makes an argument that should bother Suno users more than it bothers Suno. It points out that Suno's terms require every user to grant a licence before Suno trains on their content, and says this shows Suno knows training needs permission. So in this case, the only people who have clearly licensed their work to Suno for training are its users. They did it through a click-through, for free and forever, and they can't take it back.

The lineage theory makes this worse for creators. If the court accepts that outputs of a tainted model can taint the next model, then user creations are the means by which the old model gets into the new one. Your songs become the evidence in someone else's copyright fight, and you have no say in it. If Suno wins, your creations stay training data anyway. If the labels win, they become contaminated training data. Users lose either way.

There is no clean restart in AI music, and pretending otherwise helps nobody. Suno's licensed-era pitch, with v6 launched alongside Warner, BMG and Believe, rests in part on a model shaped by choices users made on earlier models, the ones the labels say were built on unlicensed recordings. The labels have their lawyers. Users need two simple things now:

• A per-account, per-track setting to keep creations and ratings out of training, with an honest statement of what has already been used.
• A plain-language note on each model version listing which kinds of user data it learned from.

Until that exists, treat anything you make or upload on Suno, and every A/B pick you make, as a contribution to the company's next model. That matters most for vocals and stems you'd want to license elsewhere. It's the same identity-and-consent question that sits at the centre of Jason Isbell's class action, from the other side of the prompt box.

Sources

• Answer of Defendant Suno, Inc., UMG Recordings v. Suno, No. 1:26-cv-14275-FDS, Dkt. 11, filed 20 Oct 2026 (primary)
• Complaint, UMG Recordings v. Suno, No. 1:26-cv-14275, Dkt. 1, filed 18 Sep 2026 (primary)
• Suno Terms of Service, effective 3 Sep 2026 (primary)
• Suno Privacy Notice, last modified 23 Jun 2026 (primary)
• Music Business Worldwide: Suno admits it has not deleted all the recordings used to train its AI
• Music Business Worldwide: Suno confirms v6 model was trained on creations from users