A new investigation by The Atlantic has exposed the massive scale of music being used to train generative AI systems, raising fresh concerns about copyright, consent, and compensation in the entertainment industry.
The report, led by journalist Alex Reisner, identified four large datasets containing approximately 21.2 million music tracks. These collections are being used to train AI models capable of generating new songs that imitate human-created music. The largest dataset alone reportedly includes around 12 million tracks, while another holds roughly 9 million.
Searchable records within these datasets show that music from some of the world’s biggest artists was included. This includes Taylor Swift, Bad Bunny, Billie Eilish, and even legendary rock band Nirvana. In many cases, the music appears to have been scraped and used without direct permission from the rights holders.
The findings give artists and record labels a rare opportunity to trace how their work may have been used in AI training. They also provide concrete evidence in an ongoing legal and ethical battle over how artificial intelligence companies source their training data.
At the center of the debate is whether this kind of large-scale data use counts as “fair use” or whether it amounts to copyright infringement. AI companies often argue that training models on existing music is transformative, meaning it creates something new rather than copying original works. However, critics argue that the outputs of these systems can closely mimic real artists’ styles, raising concerns about originality and ownership.
The investigation comes at a time when the music industry is already taking legal action. Major record labels, including Universal Music Group and Sony Music Entertainment, are currently suing AI music platform Suno over alleged large-scale copyright violations.
In court filings, labels have accused AI developers of using massive libraries of copyrighted music to train their systems without proper licensing. In response, companies like Suno have argued that their models are trained on data in ways that comply with existing legal frameworks, though they have also acknowledged exposure to large volumes of recorded music during development.
Streaming platforms are also feeling the effects of this shift. Services like Spotify and Deezer have reported a growing flood of AI-generated content, with some platforms saying a significant portion of new uploads are now artificially created tracks rather than human-made music.
The legal implications are now being tested in courtrooms, where judges must decide whether training AI on copyrighted music is a form of innovation or infringement. Previous cases in the tech industry, including disputes involving other AI companies, have shown how complex and unsettled this area of law remains.
For artists, the findings mark a turning point. What was once hidden behind proprietary AI training systems is now becoming more visible, giving musicians and labels clearer insight into how their work may be shaping the future of machine-generated music.
