Suno built its AI music generator on the back of thousands of hours of audio it had no obvious right to use — and now there's leaked source code to prove it.
According to code that's surfaced publicly, Suno assembled significant chunks of its training library by pulling audio from Deezer, YouTube, and stock music platform Pond5. We're not talking about a few sample tracks scraped for testing. We're talking about a systematic, large-scale operation to hoover up content that real artists, producers, and rights holders put onto those platforms under very different assumptions about how it would be used.
What the Leak Actually Shows
The source code reveals how Suno's training pipeline was built — and specifically names Deezer, YouTube, and Pond5 as sources for the audio data fed into the model. This isn't speculation or a disgruntled ex-employee sounding off. It's in the code.
Pond5 is particularly telling here. That's a platform where musicians upload and licence their work. People put their music there specifically to control how it's used and to get paid when it is. The idea that it was being harvested en masse to train a competing AI product — one that can now produce passable music at the click of a button — is going to sting anyone who's ever uploaded there expecting their licencing terms to mean something.
YouTube and Deezer present their own complications. Both platforms sit on top of content they licence from labels and artists. Whether those licences extend to anyone else training AI models on that audio is a legal question that no one in the industry has properly settled yet, and frankly, that ambiguity has been doing a lot of heavy lifting for AI companies across the board. This leak suggests Suno wasn't waiting around for the legal framework to catch up.
The Bigger Picture for AI and Music
We've seen this pattern before. AI companies build, ship, monetise — and deal with the intellectual property fallout later, usually in court. The music industry has been watching this space with particular anxiety because, unlike text, audio has well-established licensing infrastructure. The rules exist. The question is whether anyone building these tools genuinely considered themselves bound by them.
This isn't entirely unlike what's been happening in other parts of the AI ecosystem — [OpenAI has faced its own scrutiny over training practices](/getohedz/crypto/openai-uses-ai-red-team-to-strengthen-gpt-56-against-prompt), and it speaks to a wider tension between how fast these models are being developed and how far behind the governance is lagging. Jonas Blue said it plainly — he called AI in music ["absolutely horrendous"](/getohedz/music/jonas-blue-rebranded-learned-an-instrument-called-ai-absolutely-horrendous) — and leaks like this give you a pretty clear window into why working musicians feel that way.
The irony is that Suno's whole pitch is democratising music creation. Making it accessible. Empowering people. But if the foundation of the product was built by taking from the very community it claims to empower, that pitch doesn't hold up.
Our take: The leak matters because it puts specifics where there used to be vague allegations. Names of platforms, shape of the pipeline, scale of the operation. Suno will almost certainly dispute the interpretation — that's standard procedure — but the code doesn't lie. The music industry's lawyers are going to have a field day with this, and rightly so.