Browse the models and styles behind stories like this one — free account, instant gallery.
Atlantic reporter Alex Reisner has published a fully searchable public database of four music datasets used to train AI models — two of which contain 12 million and 9 million tracks respectively, making this the most comprehensive public accounting of AI music training data to date.
Reisner's investigation — published by The Atlantic — uncovered the four datasets through a combination of leaked information and public research. Two are enormous by any measure: one contains roughly 12 million tracks, the other around 9 million. The remaining two are smaller but still represent substantial training corpora. Together, they add up to a dataset that dwarfs most publicly disclosed AI training sets in any creative domain.
The searchable interface lets anyone enter an artist name or song title and see whether it appears. That's a meaningful shift from the usual situation, where training data is either undisclosed or buried in technical papers that most creators never read. Rights holders who have long suspected their work was used without permission now have a concrete tool to check.
This disclosure doesn't exist in a vacuum. Multiple ongoing lawsuits — from musicians, record labels, and visual artists — have pushed AI developers into an uncomfortable spotlight over where their training data comes from. Courts in the US and Europe have been wrestling with whether scraping copyrighted material for AI training constitutes fair use or infringement, and so far no definitive ruling has settled the question.
For AI developers, the timing is awkward. Several major music-focused AI tools have launched or expanded in 2024 and 2025, and the question of what those models were trained on is now harder to sidestep. A searchable database that any journalist, lawyer, or artist can query changes the dynamic entirely — it converts an abstract legal debate into a concrete, searchable fact.
This matters beyond music. The same legal logic that applies to training on copyrighted songs applies to training on copyrighted images, illustrations, and visual art. The Atlantic's database is a proof of concept: training data can be documented, made searchable, and used as evidence. If similar databases emerge for image training sets — something researchers have already begun building in academic contexts — the pressure on image-generation model providers will intensify.
For creators who use AI image tools, the practical consequence is model risk: platforms that can't demonstrate clean or licensed training data face a higher chance of legal challenge, forced model withdrawal, or settlement-driven restrictions on outputs. When evaluating which AI image generation tools to build a workflow around, training data provenance is no longer just an ethical consideration — it's a business continuity one.
It also raises a harder question about the future cost structure of AI models. Licensing music and visual art at scale is expensive. If courts or regulators push AI developers toward licensed training data, the economics of running cheap or free generative tools shift significantly. Checking out current model options and pricing before those pressures materialize is worth doing now, while the landscape is still relatively open.
Reisner's database won't resolve the underlying legal fights, but it sets a precedent for what training data accountability can look like in practice — and that precedent is now sitting in front of every AI developer's legal team.
