Music generation AI Suno hacked, training data revealed through leaked code #117

The shocking reality of massive data collection.

The leaked code contained records indicating that data was scraped from YouTube Music—totaling approximately 110,000 hours—as well as vast amounts of data from stock music services and lyric websites. It is even more astonishing that the system included mechanisms to identify and collect acapella tracks specifically for vocal training.

Podcasts were also targeted for training.

Of particular concern is the evidence that they used the PodcastIndex to identify 420,000 programs and attempted to download roughly 1 million hours of audio. Personality YOSHIO shares his complex feelings about the reality that our very own speech was being treated as fodder for AI development.

The precarious relationship between AI development and copyright is highlighted.

Suno is currently being sued by major record industry associations, and this leak is likely to impact the court case. Learning the methods used to collect this data behind the scenes of such convenience forces us to rethink how we engage with AI. If you are interested in future developments, please check out the full episode.