AI Music Training Exposed

Could the songs you stream every day be fueling the next generation of artificial intelligence without your knowledge? Atlantic reporter Alex Reisner has uncovered 4 massive datasets used to train AI models, exposing a historic tipping point in how creative work is harvested.

2 of these never-before-seen sets are absolutely enormous, containing 12 million and 9 million tracks, while two smaller sets contain over 100,000 songs each. Reisner revealed that three of these datasets are distributed simply as lists of links to songs on major platforms like YouTube and Spotify.

To build their models, developers use automated tools to scrape the actual audio, bypassing logins, advertisements, and creator monetization, in direct violation of platform terms of service, While these datasets are freely available on the internet in theory, using them legally is another story, especially for commercial applications. Major industry players Google and Stability have already confirmed using these exact datasets within their own research papers.

Now, a first-of-its-kind searchable database launched by The Atlantic allows the public to see exactly which artists, from Lady Gaga to Radiohead, have been swept into the training machine.