A Searchable Database Lays Bare the Music Behind AI Models
An Atlantic reporter has made four AI music-training datasets - two holding millions of tracks - fully searchable. For India's catalogue owners, the receipts have arrived.
The News
Alex Reisner, a reporter at The Atlantic, has built a public, searchable database exposing four datasets of recorded music that have been used to train AI systems. The tool lets anyone check whether specific tracks sit inside the troves that model builders have quietly drawn on.
The scale is striking. Two of the four collections are vast, holding roughly 12 million and 9 million tracks respectively. The remaining two are smaller but far from trivial, each containing more than 100,000 songs. Together they represent one of the clearest public windows yet into the raw material feeding music-generation models.
Reisner reports that the datasets have been downloaded thousands of times. He is careful to note that it is impossible to say with certainty who has used each one, though his findings point to links involving Google and Stability AI. The work was surfaced in coverage published on 20 June 2026.
Why It Matters
This is a familiar move from Reisner, and that is precisely why it lands. In 2023 he did much the same for the written word, building a searchable index of the Books3 dataset that revealed how thousands of pirated titles had been swept into language-model training. The provenance reckoning that began with text and images is now arriving for audio.
The timing is pointed. Since 2024 the major record labels have been pursuing music-generation startups such as Suno and Udio over how their models were trained, turning an abstract argument about fair use into hard litigation. A public, searchable map of training tracks shifts that fight from suspicion to evidence. Rights holders no longer have to guess whether their work was ingested - they can look it up.
For model builders, transparency is becoming a liability rather than a virtue. The more legible a training set is, the easier it becomes to quantify exposure, calculate damages and demand licences. Expect the next round of AI music deals to be struck in lawyers' offices, not research labs.
Indian Angle
Nowhere is the stake higher than in India, a country whose recorded-music economy is built on deep, valuable back-catalogues. Saregama alone controls a film-music library stretching back decades, while T-Series and Sony Music India command enormous shares of contemporary Bollywood and regional output. If even a fraction of that material sits inside these datasets, it represents licensing revenue that has so far gone uncollected.
Indian rights bodies are positioned to act. The Indian Performing Right Society and the trade group Indian Music Industry have spent years tightening royalty enforcement against streaming platforms, and a searchable AI-training database hands them a new front. India is among the world's largest music-streaming markets, served by JioSaavn, Spotify and others, which makes the commercial logic of Indian catalogues appearing in global training sets entirely plausible.
The legal scaffolding is already being tested. The Copyright Act of 1957 contains no clear text-and-data-mining exemption, and the Delhi High Court is hearing ANI's closely watched case against OpenAI over training data. A music equivalent now looks less like a hypothetical and more like a matter of when. For Indian labels, the question has moved from whether their songs were used to how much that use is worth.
FAQ
How big are the exposed datasets?
Four datasets in total. Two are enormous, at roughly 12 million and 9 million tracks. The other two each hold more than 100,000 songs. Combined, they form one of the most detailed public views yet of the music feeding AI models.
Does this prove specific companies used the data?
Not conclusively. Reisner stresses it is impossible to know exactly who used each set, though his findings point to links involving Google and Stability AI. The datasets have been downloaded thousands of times.
What does this mean for Indian music labels?
It offers a way to check whether catalogues from Saregama, T-Series or others were ingested, strengthening any future claim for licences or damages as Indian courts and rights bodies sharpen their stance on AI training.
Where can I read the original report?
The Verge covered the database and its findings in detail, with a link through to the underlying Atlantic project.
This story was reported by The Verge. Read the full original coverage at The Verge.