Internal data obtained by 404 Media through a breach of Suno, the controversial AI music platform that allows users to create full tracks by entering text prompts, reveals new details about how the company assembled the recordings behind its song-generating models.Ad 0:00 Click for sound 0:00 / 0:00 According to the report, published this week by 404 co-founder Jason Koebler, the files were provided by a hacker who claimed they infiltrated Suno’s systems.
The material, which appears to date to 2023 and 2024, outlines how the startup allegedly pulled audio from a wide array of sources, including YouTube Music, Deezer and Genius, and several stock audio libraries such as Pond5, Jamendo and Freesound.Suno has faced significant legal scrutiny over its training practices, including a recent lawsuit filed by Hagens Berman, the firm behind the largest litigation settlement in history.
It’s the company’s latest legal battle after a seminal lawsuit filed in the summer of 2024 by the RIAA alongside three major record labels (Sony, UMG and Warner), alleging that its models were developed using millions of copyrighted recordings without authorization.Suno has not disputed that it trained on copyrighted material, but has argued in court filings that doing so is protected under fair use.
404‘s reporting appears to corroborate the RIAA’s allegations, which accused Suno of extracting audio directly from YouTube in violation of the platform’s terms.Internal metadata identified in the leaked code references several of the datasets the company compiled, including references to files that had ingested more than two million individual clips from YouTube Music alone.
Together with tens of thousands of additional hours drawn from Genius and the sheet music archive IMSLP, among other sources, the haul amounts to what the outlet called “decades worth of music and podcasts from the internet to train its AI ...