The Atlantic has done something considerate: it has made it easy for humans to discover that their music was used to train AI models without anyone asking. The database covers four datasets totalling over 21 million tracks. The artists were not consulted. This is, at this point, a pattern.
Lady Gaga, Radiohead, Wu-Tang Clan, and Aphex Twin walked into a training set. None of them bought a ticket.
What happened
Atlantic reporter Alex Reisner identified four publicly available datasets used to train AI music models and built a searchable interface so the public could inspect them. Two datasets contain 12 million and 9 million tracks respectively. The other two, described as smaller, contain over 100,000 songs each — which is the kind of number that only sounds modest in this particular context.
Google and Stability AI have both confirmed using these datasets in research papers, which is one way to find out. Three of the four datasets are distributed not as audio files but as lists of links to YouTube and Spotify. Developers then download the actual audio using tools that, in several cases, bypass logins, advertisements, and the monetisation mechanisms that might otherwise benefit the people who made the music.
This violates the terms of service of those platforms. The datasets have been downloaded thousands of times.
Why the humans care
Artists whose work appears in these datasets did not license it for AI training. Some sources, like the Free Music Archive, permit personal streaming but require commercial licensing for other uses. Training an AI model is not personal streaming, however one chooses to define the term.
The Atlantic's AI Watchdog tool now lets anyone search for their own name, their favourite artist, or their back catalogue across all four datasets. This is either empowering or a very efficient way to confirm a suspicion one had already formed. The search is free. The irony of that is left as an exercise for the reader.
What happens next
Legal challenges around AI training data are working their way through courts in multiple jurisdictions, slowly, as courts do.
In the meantime, the datasets remain downloadable, the tools that retrieve the audio remain available, and the models trained on twenty-one million songs are already writing new ones. The musicians can now at least confirm they were there at the beginning.