Spotify cut 50 million AI tracks from recommendations

Spotify says it pulled 50 million AI-generated tracks out of its recommendations over the past year, and the wording matters. The tracks were not deleted. They remain on the service and a listener can still find them by searching, but nothing algorithmic or editorial will put them in front of anyone.

How big 50 million is

Put three minutes on each one and the pile becomes legible. Fifty million tracks is about 285 years of continuous audio, and that is only this year's AI tier. Spotify separately removed more than 75 million tracks outright as spam, which takes the combined figure past seven centuries of playback.

How the tracks were found

The detection is not a single trick. Spotify names listener engagement, upload patterns, metadata and several third-party AI detection tools. Since August it also runs an AI Persona badge: a profile either declares itself non-human or gets assessed and labelled, and a badged profile stays out of recommendations unless a user follows it on purpose.

The interesting part is who is refusing to make this music. A$AP Rocky, asked about AI, said he prefers to work practically and by hand. Grimes licensed her own voice to anyone who wanted it three years ago and still does not generate her own songs. The people most comfortable with the technology draw the line at authorship.

“I prefer to do things myself, practically. If somebody does a totally AI-generated song where all they did was prompt something, then I think that's their thing.”

— A$AP Rocky, Asked about AI in music, October 2026

Supply and demand have come apart here. The supply side is industrial: training sets of millions of songs, like the one at the centre of Australia's copyright fight, and generation that costs cents, as we found when we priced Suno's credits. The demand side is a person deciding whether to press play, and that side is not scaling with the other.

What is still not labelled

One gap is left open on purpose. Spotify badges the profile but does not require a label on the song, so a track made by a human using AI tools, and a track generated end to end, arrive at the listener looking identical. The platform has solved who is publishing, not what is playing. Behind the files sits an ordinary diffusion model trained on other people's recordings, and the service is labelling the shop window rather than the factory.

Two things settle this. Whether hiding fifty million files is enough when the next fifty million upload themselves, and whether per-track disclosure arrives before regulators ask for it. Spotify built the badge without being made to, which is usually a sign a company would rather set the rule than receive it.

Informational material, not investment advice. The tracks were removed from recommendations rather than deleted, and Spotify has published no breakdown by genre, label or country.