Where Did AI Companies Get the Music Used to Train Their Models?

Where did AI companies obtain the actual music recordings used to train their models?

In an earlier investigation, I discovered that some music databases described as containing stolen music actually contained metadata—song titles, artist names, streaming counts—not the recordings themselves.

That left me with a question I couldn't answer: Where did the actual audio files come from?

Recent copyright litigation involving Anthropic and OpenAI revealed how companies obtained millions of books from pirated sources. Could something similar have happened with music?

I explore several possibilities, including employees bringing collections of pirated MP3s to work and the use of tools such as yt-dlp to download audio from YouTube.

I also examine digital rights management (DRM), why copying protected streams can be more difficult, and how Garth Brooks' decision to limit his music's distribution offers an interesting perspective on protecting artists' recordings.

Some AI music copyright lawsuits have settled, while others continue. But important questions remain about where the training recordings came from, how they were acquired, and whether the companies had permission to use them.

If you know how AI companies obtained the actual recordings used to train their music models, I'd like to hear from you.