The Atlantic exposes the music datasets fueling AI training
Millions of copyrighted songs, ranging from Bruce Springsteen anthems to Aphex Twin soundscapes, are being fed into AI models without clear licensing. Reporter Alex Reisner has pulled back the curtain, publishing a searchable database of four massive datasets that reveal exactly which artists are powering the industry's latest generative tools.
Two of these collections reach staggering proportions, housing 12 million and 9 million tracks respectively, while two additional sets contribute another 200,000 songs. Although these files exist in the public domain of the internet, they are not simple archives. Developers frequently scrape the audio from platforms like YouTube and Spotify using automated tools that bypass logins and advertisement blockers, directly violating the terms of service of those hosting sites.
Google and Stability AI have acknowledged utilizing similar data in their research papers, though the full scale of corporate adoption remains opaque. The Atlantic’s new AI Watchdog portal allows users to search the lists, confirming that high-profile musicians like Lady Gaga, Radiohead, and the Wu-Tang Clan are present in the training material. While some sources, such as the Free Music Archive, permit personal streaming, their inclusion in commercial AI training sets raises significant questions regarding intellectual property rights and the future of creative compensation.
Comments (0)
No comments yet. Be the first!