The Atlantic publishes searchable database of music in AI training datasets

The Atlantic’s investigative reporter Alex Reisner has exposed a significant transparency gap in AI model training by uncovering and cataloging four datasets containing over 21 million musical recordings used to train commercial AI systems. According to The Verge, Reisner created a publicly searchable database documenting which artists’ work appears in these datasets, many without the artists’ knowledge or consent.

Scale and Composition of Discovered Datasets

The four datasets vary dramatically in size. Two datasets contain 12 million and 9 million tracks respectively, while the remaining two hold over 100,000 songs each. According to The Verge’s coverage of Reisner’s reporting, these datasets have been downloaded thousands of times across the industry. Major AI developers including Google and Stability AI have publicly confirmed their use of these training materials in published research papers, suggesting widespread adoption without proportional public disclosure.

The datasets span a broad cross-section of contemporary music. Artists whose work appears include mainstream pop acts like Lady Gaga and Fred again.., experimental musicians such as Aphex Twin and composer Hainbach, rock institutions including Radiohead and Bruce Springsteen, and hip-hop collectives like Wu-Tang Clan. The presence of such recognizable names underscores the datasets’ comprehensiveness and commercial value.

Technical Implementation and Legal Gray Area

Three of the four datasets operate as lists of URLs pointing to songs hosted on YouTube or Spotify rather than direct audio files. AI developers must download the actual audio using automated tools, which according to Reisner’s investigation, frequently bypass platform login requirements, advertisements, and revenue-sharing mechanisms. The Verge notes that these extraction tools directly violate the terms of service of both YouTube and Spotify, creating a tension between technical possibility and contractual obligation.

The Free Music Archive dataset, one of the sources, permits personal-use streaming without licensing but requires commercial licensing for derivative applications. Using such datasets for commercial AI model training likely exceeds the scope of personal-use licenses, creating potential copyright liability for model developers.

Why This Matters

The Atlantic’s searchable database addresses a critical asymmetry in AI training transparency. Most AI companies do not disclose which specific artists’ work trained their models, preventing creators from knowing whether their output influenced a system generating competing content. This database enables independent verification of training sources at a granular level—individual songs rather than aggregate dataset names. For recording artists, music labels, and policymakers evaluating AI regulation, the ability to search and identify whose work was used without permission raises urgent questions about fair compensation, attribution, and whether current copyright frameworks adequately protect creative workers in the generative AI era. The revelation that major vendors like Google and Stability openly used these datasets suggests industry normalization of a practice many musicians view as unauthorized use.