AI music labels are the easy part. Streaming services should show how their filters decide

Between 13 and 20 Oct 2026, three big streaming services announced new rules for AI music. Deezer announced a toggle that keeps fully AI-generated tracks out of a listener's search results. YouTube Music started putting labels on songs "created primarily using generative AI". Spotify said it had taken more than 50 million "slop and spam" tracks out of its recommendations this year.

The labels and the toggle will get the attention. We think the filters matter more.

Why are filters a bigger deal than labels?

A label is information. You see it, you decide. A recommendation filter is a decision made for you, before you ever see the track. For an artist without an existing audience, being filtered out of recommendations can come close to not being released.

The platforms are already doing this at scale. Deezer says songs it detects as AI-generated "are automatically removed from algorithmic recommendations" and kept off editorial playlists. In its July update it said it receives nearly 90,000 fully AI-generated tracks a day, more than half of all new deliveries on peak days. Spotify's 50 million figure is about recommendations too. It also keeps AI Persona artists out of editorial and algorithmic recommendations by default.

We don't object to any of that. Deezer says up to 85% of streams on fully AI-generated tracks in 2025 were fraudulent, and that as of June 2026 this music makes up only 1–3% of streams on Deezer. Spam is a real cost to every artist who shares the royalty pool, and filtering it is reasonable.

Where does it go wrong?

At the edges, which is where most working producers now live. Plenty of people use AI for a drum idea, a stem split or a vocal cleanup and do everything else by hand. Deezer says music with "limited AI-assisted elements as part of a broader human creative process is treated differently", but doesn't say how. Spotify doesn't define "slop". It also admits that "AI detection technology in particular has limitations, and it will only get harder as generative AI evolves."

None of the four announcements we read describes a way for an artist to contest a decision, and Spotify's doesn't say how an artist would learn that a track was filtered. Deezer does publish one figure: its July post says its detector is 99.8% accurate, with fewer than 1 false positive per 10,000 human-made tracks. That is a start, but it is Deezer's own number for its own tool, it covers the detector rather than each filtering decision, and it doesn't say how often AI-assisted human work is flagged. YouTube says it is "using our own detection tools" alongside music partners, and doesn't say whether its labels affect recommendations at all.

What should platforms publish?

If you filter tens of millions of tracks, you owe the people who make music three things:

1. The line. A written definition of AI-assisted versus AI-generated, with examples, so a producer can know in advance which side a workflow falls on.
2. Independent error rates. Figures checked by someone other than the platform, reported for each decision the detector drives (search, recommendations, playlists), including the false-positive rate for AI-assisted human work, not only for fully human tracks.
3. A route back. A notice in the artist dashboard when a track is filtered, and an appeal that a human reviews.

Spotify's AI Credits show this is possible. Artists already tell Spotify, track by track, where AI was used. A platform that asks artists to be transparent should be transparent about what it does with that information.