The Atlantic opens a search tool for 21 million tracks used to train AI

Alex Reisner, a journalist at The Atlantic, has published the results of an investigation: he identified four datasets of music that tech companies used to train generative AI models. He also made all four databases publicly searchable. Anyone can now check whether their tracks ended up in the training data.

What Reisner found

The two largest datasets are striking in scale: one contains 12 million tracks, the other 9 million. That adds up to 21 million music files in just two databases. The other two datasets are more modest but still significant: each includes more than 100,000 recordings. Together, this is a colossal volume of content, most of the musical heritage that can be collected automatically.

All four datasets have been downloaded thousands of times. It is impossible to identify exactly who downloaded them, but Google and Stability AI officially confirmed in their scientific publications that they worked with this data. This is documentary evidence: companies with multibillion-dollar valuations relied on the same sources that have now become public.

Where this music comes from

The sources of the datasets differ in legal status, and this is where the most important part begins:

• Free Music Archive is free for personal listening, but commercial use and the creation of derivative works are restricted
• Some tracks are published under Creative Commons licenses, but the specific terms differ from track to track
• Some material is protected by standard copyright, with no exceptions or caveats
• All the databases were technically available for download without any restrictions
• No AI company has publicly disclosed the exact composition of its music training sets

The gap between "technically available for download" and "legally usable for commercial AI training" is exactly the legal territory where lawsuits are now unfolding around the world.

A tool for rights holders

The Atlantic has launched a public search engine across all four databases. Any musician, producer, label or publisher can check their name or track titles and find out whether that content was part of the training corpus.

This matters in practical terms. Lawsuits against AI companies, including Suno, Udio, OpenAI, Stability AI and others, are already in court, but plaintiffs have so far lacked a reliable way to prove that specific works were used. The Atlantic's public database could become evidence in these cases.

Reisner's investigation continues a series of exposés over recent years. First came the mass use of books without permission (the Books3 dataset), then text from the open web (Common Crawl). Now it is music's turn. The logic is the same: AI companies collected everything that was technically available, without asking about its legal status.

What this means

The Atlantic's publication moves the AI copyright dispute from the abstract to the concrete: here is the data, here are the companies, here are the tracks. For musicians, it is the first public verification tool. For AI companies, it is a signal that opacity about training data is becoming harder to maintain.