A new searchable database is putting a clearer public face on a hard-to-see part of AI development: the music used to train models. The Atlantic reporter Alex Reisner recently identified four music datasets connected to AI training and made them available to search.
The finding matters because the scale is large, the sources are varied, and the rules around use are not always straightforward. The database gives artists, labels, researchers, and listeners a way to look for specific tracks instead of speaking about AI training data only in the abstract.
What The Atlantic made searchable
Reisner uncovered four datasets of music used to train AI models. Two of them are especially large, with 12 million and 9 million tracks. The other two are smaller by comparison, but still contain over 100,000 songs each.
That means the database is not a niche catalog. It represents a broad pool of music that has been gathered into datasets and made available online. The searchable format changes how people can interact with that information: instead of relying on descriptions of datasets, they can look for names and works directly.
The article says the datasets have been downloaded thousands of times. It is not possible to know exactly who has used them, but Google and Stability have both confirmed use in research papers.
Why licensing is central to the story
The fact that music is accessible online does not automatically mean it can be used for every purpose. One example in the source is the Free Music Archive dataset, which is free to stream for personal use but requires licensing for commercial applications.
That distinction is important for AI training because a dataset can sit in a gray area between availability and permission. A track may be easy to find, easy to stream, or easy to list in a public dataset while still carrying limits on how it can be used.
For musicians and rights holders, the practical question is not only whether a song appears somewhere online. It is whether the use matches the conditions attached to that music. The database gives people a starting point for that inquiry by making the contents easier to search.
How the datasets point to music
The source article also explains that training data is not always distributed as a bundle of audio files. In three of the datasets Reisner found, the dataset is a list of links to songs on YouTube or Spotify.
That structure matters because the links are only one step in the process. AI developers still need the audio, and the source describes tools that automate the download process. Some of those tools can bypass logins, advertisements, and mechanisms that might generate money or subscribers for creators.
According to the source, such tools violate the terms of service of those platforms. That adds another layer to the issue: the concern is not only about whether songs are named in a dataset, but also about the methods used to turn links into usable audio for AI training.
Whose music appears in the data
The searchable datasets include names from across popular, experimental, and electronic music. The source lists Lady Gaga, Fred Again.., Radiohead, Aphex Twin, Wu-Tang Clan, Bruce Springsteen, and experimental composer Hainbach among names that appear.
Those examples show that the material is not limited to obscure recordings or one type of music. It spans highly visible artists and more specialized creators. That range is part of why a searchable database is useful: different communities can check the same underlying collections from their own perspective.
For an artist, appearing in such a database does not by itself answer every question about use, permission, or impact. But it can provide evidence that a work is present in a dataset associated with AI training. That is a more concrete starting point than guessing whether a catalog may have been included.
What changes when training data becomes searchable
AI training data is often discussed at a high level, with attention on models, companies, and broad disputes over creative work. A searchable database shifts attention toward the individual works inside those systems.
That shift makes the issue easier to inspect. Users can search through songs, books, and other media on The Atlantic’s AI Watchdog site. The result is a more practical view of what has been gathered and how large these collections can be.
The database does not answer every open question. The source makes clear that it is impossible to know exactly who has used the datasets. It also does not resolve whether every use is permitted, licensed, or commercially relevant.
What it does provide is visibility. With millions of tracks in the largest datasets and over 100,000 songs in each of the smaller ones, the searchable tool gives creators and the public a way to examine music AI training data with more specificity than before.