Millions of songs were used to train the music AI.

Tens of millions of songs, from Taylor Swift to Bad Bunny, have been used to train artificial intelligence (AI) music applications like Suno and Udio, The Atlantic has revealed in a new investigation. The article comes in the form of several databases that can be searched. The biggest one contains 12 million songs, another one has 9 million songs, while two others have around 100,000 songs each. Other AI models used this data, such as Google’s.

This is similar to previous investigations by The Atlantic into book scraping, but this time, it’s at a much bigger scale and with a more popular dataset. It’s not only about random demo songs, but also about commercial music that gets sold daily on streaming platforms. Suno and Udio are currently facing lawsuits from big music labels for copyright infringement, and they’re using the ‘fair use’ defense (which is currently under litigation in many other AI cases).

1Why this matters beyond the lawsuits

This is not a theoretical threat. In the parallel case of book copyright, Anthropic paid $1.5 billion in a settlement, so you can imagine how much money music labels want to get out of this. Still, the part that concerns regular people is more subtle: AI songs are beginning to challenge their human counterparts in the same playlists you listen to. Streaming platforms are starting to react (Spotify is developing tools to help musicians control AI music, Deezer is working on AI music detection, and Apple Music warns distributors about AI songs if they choose to do so), but it’s still early days, and people are still able to clone any singer’s voice.

What’s interesting about this investigation is that it makes the problem tangible. Up until now, it was just a hypothetical concern for musicians whose songs could have been used to train an AI model, but now you can actually look it up. This is the kind of information that makes the difference between ‘we believe’ and ‘look, here it is!’ Honestly, the music industry has been arguing about this for two years, and now the question should be how to license this data rather than whether it’s already there or not (because it is, and it won’t be undone by any settlement).