AI training data for music models under scrutiny

The Atlantic released a freely accessible tool called AI Watchdog that lets users search published datasets to see which songs appear in material assembled for AI music model training. AI Watchdog reveals which songs were used to train AI music models and aims to provide transparency about the usage of music tracks in AI training datasets, including AI training data for music models.

The databases behind the tool are organized into four datasets that together contain more than 21 million tracks, and The Atlantic hosts the searchable interface on its homepage. The tool is intended to make visible the inclusion of individual recordings in those public datasets, while noting that presence in a dataset does not necessarily mean a specific model used the track.

Artists identified within the datasets include Taylor Swift, Billie Eilish, the Beatles, Miles Davis, and Nirvana, among others. The datasets also contain material collected from the Free Music Archive.

The AI Watchdog database is composed of four distinct datasets that together contain more than 21 million tracks. The collections include recordings by high-profile artists such as Taylor Swift, Billie Eilish, the Beatles, Miles Davis, and Nirvana, and they span a wide range of genres and eras.

Three of the four datasets are structured as collections of links that enable automatic download from platforms such as YouTube and Spotify, and those linked structures allow bulk retrieval of the listed recordings. The searchable interface exposes the scale and organization of these linked collections and lets users locate individual recordings within the aggregated datasets.

One dataset incorporates over 100,000 tracks drawn from the Free Music Archive that were originally collected in 2016 by the Swiss Federal Institute of Technology in Lausanne and later used by companies including Google and Stability AI. Materials from the Free Music Archive are subject to licensing terms that prohibit commercial use without permission.

The aggregated listings also show that some artists appear with dozens or hundreds of recordings in the collections. For example, SZA appears with 238 recordings and artists such as Skrillex, Peggy Gou, and Bicep are listed with dozens of tracks. The presence of a track in a dataset does not necessarily indicate that any specific AI music model used that track.

The provided material reports that GEMA has been suing Suno since March 2026. It also reports that Sony Music and Universal Music Group have filed parallel lawsuits against Suno and Udio. The provided material lists those legal actions but does not include specific claims, filing dates for the Sony and Universal actions, or the cases’ current procedural status. The available sources do not provide this information.

The provided material cites a GEMA study that estimates about 27% of music revenues could be lost by 2028 due to competition from AI-generated music. The material does not include the study’s methodology, the geographic markets covered by the estimate, or the baseline revenue figures used to calculate the percentage. The available material does not present additional, corroborating financial projections within the supplied content. The available material does not provide information on remedies, compensatory figures, or industry responses tied to the estimate within the supplied content.

The provided material includes direct reported language such as: “The tool shows which songs were used for AI music models – including tracks by Skrillex and Peggy Gou.” It also includes a reported characterization that “The company is stealing from countless musicians struggling to survive.” Paraphrased statements in the material convey concerns from artists and observers about the use of recordings in model training. The material presents these quoted and paraphrased remarks as part of the reporting.

The material additionally contains an explicit caution that “The presence of a track in a dataset does not necessarily mean that it was actually used for a specific model.” This caution is presented alongside the reported quotes to clarify the distinction between dataset inclusion and model usage. The available material does not attribute these exact lines to specific speakers within the provided excerpts. The available material does not provide further speaker attributions or contextual statements beyond the quoted and paraphrased language included in the material.

Named artists are listed throughout the AI Watchdog datasets and are identified by individual recordings and entry counts. SZA appears with 238 recordings in the training data, and the material records that Kenny Beats (Kenneth Blume) DM’d Suno on X. The listings include contemporary producers and performers as discrete entries rather than aggregated genre labels, showing specific artist-level inclusion across the collections.

The datasets also list artists such as Skrillex, Peggy Gou, and Bicep with dozens of tracks, alongside high-profile performers including Taylor Swift, Billie Eilish, the Beatles, Miles Davis, and Nirvana. These named entries appear within the aggregated datasets assembled for AI training and are presented as individual recordings or linked items in the compiled lists. The available material does not provide additional metrics such as per-artist percentage share of the total, play counts, or model-specific usage details.

The Atlantic released a freely accessible AI Watchdog tool that reveals which songs appear in datasets assembled for AI music model training and is presented as a transparency resource for AI training data for music models. Reporting ties the database findings to legal and ethical tensions, noting that GEMA has been suing Suno since March 2026 and that Sony Music and Universal Music Group have filed parallel lawsuits against Suno and Udio, while a GEMA study cited estimates about 27% of music revenues could be lost by 2028 due to competition from AI-generated music.

The material includes characterizations that the practices amount to stealing from musicians struggling to survive, cautions that a track’s presence in a dataset does not necessarily mean a specific model used it, and frames the investigation in a critical, rights-focused tone centered on musician rights and the industry impact of AI-generated music.