Suno's New Models Have a Transparency Problem

Stems is back for the fall season! Hit us up with feedback, tips and cures for AI doomer-ism at [email protected].

A few weeks ago, Suno’s PR folks reached out to let us know that the AI company was getting ready to announce a highly anticipated batch of models, developed “in partnership” with music industry counterparts. It’s a moment that many of us have been waiting for since WMG and Suno announced their deal last November and said that Suno would launch “new, more advanced and licensed models” in 2026.

Their arrival came Wednesday in the form of three new models (V6, V6-wild, and V6-mini). Suno says the models were developed with industry partners including Warner Music Group, BMG and Believe. Suno also talked up remix-style products in the pipeline that it says will create new revenue opportunities for participating artists.

The rollout — presumably calibrated to address both the ongoing litigation Suno is facing from UMG and Sony, and more broadly, the negative sentiment the AI company has inspired — seems to have gone fairly well for Suno, with neutral-to-positive media coverage from what we saw. And yet we have a lot of questions about this new chapter, even after speaking to Suno’s chief product officer, Jack Brody.

The biggest one: what exactly were the new models trained on? AI models require mountains and mountains of data to get to that generative critical mass; for context, Suno says its earlier models were trained on tens of millions of “publicly available” music files. Even if you assume good faith on the trainer’s part, if you don’t have visibility into that pile, it’s hard to be confident about the copyright status of its contents. It doesn’t help that Suno’s execs use language that can seem purposely vague on this topic. When asked, Brody wouldn’t say that V6 was exclusively trained on licensed data, offering instead that “The data set itself is new and … licensed content is part of that. But ultimately, the thing that has gotten it to where it is is all of these learnings from the past.”

Speaking to Billboard, Suno CEO Mikey Shulman gave a little more detail, saying that no music from UMG, Sony or Kobalt was in the new models (UMG and Sony not exactly a surprise given that they are pursuing Suno in court). But when asked point-blank if there was any unlicensed music in the model, Shulman seemed to deflect, answering: “All of the data that we have from our partners is licensed. All of the user interactions then go in on top of that. And so that is, in our opinion, what makes for a licensed model.” Should this really be a matter of opinion or general counsel-approved parsing? Suno’s executives could have spoken directly about what non-partner data was used to train the new models, but chose not to.

Then there are those “learnings from previous models” that Brody said guided the new model’s development, which he described as techniques around model architecture, evaluation and tuning. Those techniques emerged from work on earlier models that Suno says were trained on publicly available music files from the open internet. Which begs the question: even if this new model’s corpus ultimately proves to be 100% licensed, should Suno be allowed to freely benefit from years of questionable licensing practices?

Some of these tensions are visible in the way that WMG and Suno talk differently about the deal. In a memo to his staff Wednesday, WMG CEO Robert Kyncl wrote: “These models must be fully licensed, per the terms of our agreement.” But Shulman and Brody haven’t used the same language, and seem allergic to calling this anything more than a “partnership.” Which makes you wonder if only some people’s music is going to get “fully licensed” in this brave new world.

There are ideas and technology out there that could answer many of these questions. Earlier this week I spoke to Benji Rogers, a British-born, NYC-based entrepreneur who is co-president of Sureel, an AI attribution company that WMG agreed to acquire in June. According to Rogers, Sureel creates a mathematical “AI DNA” for songs; when its code is integrated into an AI pipeline, it can trace how much any given song is contributing to an AI model’s outputs. “Because we observe the training data [and] the training process with that little bit of code, we can actually see what’s being used,” Rogers said. I can’t speak to whether Sureel is better than its competitors in the attribution space, but there is undoubtedly a lot of smart tech being developed with the goal of getting artists and songwriters credited and paid.

It’s going to take work to get to a place with that kind of transparency, though. “If I owned a bunch of rights,” Rogers said, “I would be saying, ‘You want to do a licensing deal? Great. I need attribution. Because otherwise you’re blind.’”

Grocery store pods spend months going factory → warehouse → truck → shelf, losing the flavor oils that make coffee taste like coffee. Angelino's roasts, packs, and ships within days — 50+ flavors, compatible with Keurig®, from 39¢ a cup. New customers save 15%, no code needed.

Try Fresh-Roasted Pods