A new search tool from The Atlantic lets you check if your music was used to train AI models.
The database relies on research by Atlantic reporter Alex Reisner. He recently identified four major music datasets used in AI development and made them searchable for the public.
Together, these collections contain millions of recordings spanning major artists, independent musicians and underground producers. Two of the datasets are particularly vast, containing around 12 million and 9 million tracks respectively, while the remaining two each include more than 100,000 songs.
READ MORE: “If you’re a musician and you support this degenerate shit, you’re disgusting”: SZA calls out the “vultures” training AI on tracks without permission
Reisner states the datasets have already been downloaded thousands of times. While it is impossible to pinpoint exactly who has used them, both Google and Stability AI have acknowledged drawing on at least some of the material in their research papers. Some sources, including the Free Music Archive dataset, are available for personal listening but require licensing for commercial use.
Reisner argues that what the datasets ultimately expose is a gap between how AI companies describe their training data and how it is actually sourced. While developers often claim to rely only on freely available material, the reality is more complicated.
Although the datasets are publicly accessible in theory, using them for AI training is not as straightforward as downloading files and feeding them into a model. As Reisner explains, many of the collections are distributed as lists of links to tracks hosted on platforms such as YouTube or Spotify.
From there, AI developers then use automated tools to download the audio at scale – in some cases bypassing logins, advertisements, or other mechanisms designed to generate revenue for creators. Such tools violate the terms of service of these platforms, Reisner notes.
The result is a training landscape where enormous quantities of commercially released music are technically accessible, but where consent, compensation, and visibility remain murky at best. Those tensions sit at the centre of ongoing legal and ethical debates around generative music tools – particularly the argument that training on copyrighted material can be justified as fair use, a position frequently cited by AI companies like Suno, and most recently, Anthropic, in court filings and public statements.
Credit: The Atlantic
The datasets also underline just how wide the net has been cast – with music from across genres and eras appearing inside them. Artists found within the collections include Daft Punk, Aphex Twin and Radiohead, alongside Wu-Tang Clan, Jack Antonoff and many more.
Antonoff, for one, has previously spoken out against the use of AI in music-making, saying, “the idea of optimising what we do is a complete miss of the entire point of what compels us in the first place.”
He added, “We (myself, the band and everyone I know) have never been looking for this work to become quicker or easier. We were never frustrated by the randomness and magic it takes. We do it for that exact reason – and without the process itself ::: nothingness. So to everyone who is gassed up about the new ways you can fake making art, by all means drive right off that cliff. We’re genuinely happy to see you go.”
In the meantime, users can explore the datasets via The Atlantic’s AI Watchdog tool, which allows searches across music, books and other media potentially used in AI training.
