Warner's Sureel Acquisition Targets AI Training Data Royalties

Warner Music Group just acquired Sureel, a startup that wants to retrofit the music industry's pay-per-use economics onto generative AI training data. Sureel's Co-President Benji Rogers puts it bluntly: "Attribution isn't about re-creating the old economics. It's about measuring, for the first time, the thing the old economics only approximated."

Generative AI training violates the fundamental logic that ties usage to payment. A million streams of a song triggers royalties. A single training pass on that same song, even if the model later generates music that sounds like it, currently pays nothing. Sureel's software labels each music file with machine-readable instructions: allow training freely, limit influence, or block entirely. It then tracks how that file actually influences a trained model and sets fees accordingly.

Why Attribution Is Harder Than Spotting a Cover Song

Attribution must go beyond surface similarity. Sureel CEO Tamay Aykut says the real challenge is causal attribution - proving a direct relationship between a specific training data point and the model's outputs. SoundVerse, another player in the space, argues in a 2025 white paper for ongoing artist participation rather than one-time buyouts. They point out that not all training data contributes equally; a jazz-heavy output should reward jazz training data more than folk. Differential reward per output is the goal.

Simon Gozzi at STIM, the Swedish copyright agency partnering with Sureel, says they are now evaluating how Sureel's attribution reports could underpin actual licensing agreements between musicians and AI companies. That's a long way from a working marketplace.

The Gaming Problem Nobody Wants to Talk About

Any attribution system creates new incentives. If similarity drives royalties, expect a flood of reverse-engineered pastiches designed to capture training payments. Aykut proposes that carefully designed systems could value unusual, unpolished works more than radio standards - a hopeful vision that dodges the immediate engineering reality. Drew Silverstein, president of SourceAudio, doesn't mince words: "Attribution would seem to be the obvious answer, but it's flawed in AI, so we have to look at other models." He advocates simple negotiated agreements with annually recurring fees at training time, no attribution complexity.

Private deals are already emerging. Universal and Warner have signed agreements with major AI companies to train on copyrighted catalogs with consent. These bilateral pacts may set de facto industry norms before any technical attribution system matures.

Small Models, Bigger Leverage for Creators

There is a parallel trend: a shift toward smaller, customized models like IRCAM's RAVE or Jen's Style Filters, and AI tools focused on fan engagement rather than wholesale generation. Suno's recent emphasis on building fan experiences directly from artists' work, following its Universal deal, hints that the monolithic models might not dominate forever. If smaller, targeted models become the norm, musicians' collectives could band together to provide training data under egalitarian revenue splits - no black-box attribution needed.

Even the best attribution algorithm rests on human decisions. Music industry royalty splits are already baroque and opaque. A well-designed attribution layer should be "multi-layered and auditable, open to expert and regulatory scrutiny," Rogers says. Whether that happens, or attribution becomes just another opaque arm in a system creators already distrust, depends on policy intervention.

The next twelve months will show whether attribution or negotiated bulk licenses win the day, but the music industry's window to shape AI economics is closing.

Source: How Musicians Can Get Paid for Training AI
Domain: spectrum.ieee.org