Human or AI was never the question.Now it's how much, and who can prove it.

The binary died in a Munich courtroom

On July 31, 2026, the Munich Regional Court ruled against Suno in a case brought by GEMA, Germany's music rights society. The court found Suno liable on four counts: training on copyrighted works in the United States, memorizing those works inside the model itself, offering the model to the public, and generating outputs that echoed songs like "Rasputin" and "Mambo No. 5" closely enough to count as infringement. It also asserted German jurisdiction over training that happened entirely on US soil, reasoning that a collecting society can reach conduct with effects inside Germany. Suno says it's evaluating an appeal.

That ruling landed inside a wider settlement pattern. Warner closed its cases against both Suno and Udio last November and is now building a licensed generation platform on the back of those deals. Universal settled with Udio in October and has its own "walled garden" platform planned, one where AI output can't leave the site. Sony is still suing Udio in New York, with discovery running into late August. Universal and Sony are still suing Suno in Massachusetts, where the fair use ruling the industry expected this summer has been pushed to spring 2027.

None of this says AI generated music is illegal. It says training on an unlicensed catalog is a liability, and the way through is licensing, not scraping. That's a process distinction, not a taste distinction, and it's the same one the Recording Academy already made. Grammy eligible work can include AI, as long as a human's contribution to the writing or the performance is meaningful and more than incidental. Full synthesis with no credited human hand: ineligible, in court and at the Grammys, for close to the same reason.

A checkbox is not a signature

Since April 2026, Spotify has shown AI credits inside the app, built on a DDEX standard that splits disclosure into five separate fields: AI vocals, AI instrumentation, AI composition, AI post production, AI lyrics. It's a real improvement on a blind AI-or-not toggle. A track with a human vocal over an AI generated bed and a fully human mix can say exactly that, instead of getting lumped in with a one-click prompt.

In July, RIAA and IFPI backed a joint voluntary labeling push riding the same pipe: the DDEX ERN schema that already carries every other credit through the supply chain. Both organizations framed it as hybrid friendly on purpose. Artists who want to use AI in the process should be able to, as long as it's disclosed.

Here's what nobody's fixed yet. That field is self reported. If a distributor leaves it blank, by accident or on purpose, no tag travels downstream and no platform catches it automatically. Reporting has put the artist royalties lost every year to fraud actors exploiting exactly that gap at somewhere around two billion dollars. Deezer's own detection numbers show the scale: something like 44% of everything uploaded to the platform daily, on the order of 75,000 tracks, is AI generated. Only 1 to 3% of actual streams are AI tracks, and Deezer flags roughly 85% of those as fraudulent activity rather than real listening.

Disclosure metadata tells you what a label claims. It was never built to tell you what's actually true. That's a different job, and it needs a different tool.

What listeners actually reward

Producers got here first. Luminate's 2026 midyear data has 54% of working musicians reporting a positive view of AI production tools, against 35% of the general public: a 19 point gap. Musicians are also three times more likely to use AI to edit or remix existing material than non-musicians are. The people doing the work adopted the tools years before the people arguing about them online.

Listeners haven't followed at the same pace. Luminate's attitude tracking shows overall comfort with disclosed AI generated music getting worse, not better: net -13% in May 2025 to net -20% by that November, with the steepest drop among Gen Z and Gen Alpha. And despite the upload volume, no AI assisted track cracked the global Top 250 in the first half of 2026. The highest charting one landed at 282.

The clearest single data point might be Monet, an AI R&B project that reached Billboard's Adult R&B and Gospel charts and pulled more than 7 million weekly streams at launch in September 2025. By March 2026, that number had fallen under 3 million. Novelty spikes. Fandom doesn't build around a voice nobody's actually singing with.

Put it together and the picture isn't "listeners hate AI." It's listeners don't reward ambiguity, and they reward it even less once they find out after the fact. Disclosed, provable, human anchored hybrid work is the version of this with a shelf life. Undisclosed full synthesis is a spike, not a career.

What this means if you're making records right now

• Label your process honestly, even while the DDEX field is still voluntary. The version of this that isn't voluntary is closer than a full back-catalog re-tag is fun.
• Being outside the EU doesn't opt you out. Article 50 targets providers and professional deployers, but streaming doesn't respect borders, and audience facing disclosure duties can reach independent releases with EU listeners.
• A metadata checkbox isn't proof, and it wasn't built to survive a re-encode or a platform strip. If your royalties or your credibility depend on someone believing your disclosure, that's a cryptographic problem now, not a paperwork one.
• Full synthesis isn't where the volume or the money is heading. Deezer's own numbers make that plain. Hybrid work you can actually stand behind is the lane that survives contact with a real audience.
• Courts are now willing to hold AI providers liable for their training data, not just the person who typed the prompt. That changes which tools are safe to build a catalog on, and it makes reading a settlement worth your time before your next release.