The Atlantic has published four searchable databases cataloging music used to train generative AI models, offering the most detailed public look yet at the scale of copyrighted content ingestion in this space. The largest index contains 12 million tracks; another holds 9 million. Two smaller sets each include roughly 100,000 songs. Combined, they span hits from Taylor Swift, Bad Bunny, and thousands of other artists.

Staff writer Alex Reisner’s accompanying analysis connects these datasets to ongoing legal battles against platforms like Suno and Udio. Those companies have invoked fair use defenses after scraping protected material at industrial scale. A parallel case in book publishing failed to convince a judge on copyright claims, but allegations of piracy proved more persuasive. That suit settled for $1.5 billion, though full payout terms remain unresolved. These new databases could give music industry plaintiffs similar leverage.

From an engineering perspective, the sheer volume here raises familiar questions about data provenance and consent at training time. Most streaming services now deploy detection or labeling systems for AI-generated content, but results are inconsistent. The databases also highlight an ongoing architectural challenge: how to build training pipelines that respect copyright without sacrificing model quality. Scammers continue to produce convincing imitations of existing bands, exploiting gaps in both legal frameworks and technical controls.

Source: Engadget