For years, musicians have said AI song generators fed on their work without asking. A hacker just opened the black box and showed them exactly how.
Leaked source code from Suno, one of the biggest AI music tools, tells the story. It trained its model by scraping millions of songs and lyrics from across the internet. 404 Media got the data from a hacker who breached the company.
The haul is enormous. One file, labelled “youtube_music,” had logged more than 2 million clips. Others list tens of thousands of hours pulled from Deezer, Genius, and the stock library Pond5. In all, it adds up to decades of recorded music.
Ripping the vocals
The code is specific about what it wanted. To capture clean voices, it hunted for a cappella versions of songs on YouTube. To slip past YouTube’s defences, Suno routed its scraping through a proxy firm called Bright Data. It also trawled 420,000 podcasts, chasing roughly a million hours of speech.
