Spotify's 50 Million Tracks Reveal a Fight Over Discovery

Spotify says it has pulled more than 50 million tracks it classifies as “slop and spam” out of recommendations in 2026. If you use Spotify to find your next song, that change concerns the part of the service making suggestions on your behalf. The tracks remain on Spotify; its recommendation system and editors no longer put them forward. Spotify has not said how many of the 50 million were AI-generated.

The number measures a decision about promotion across two categories Spotify has grouped together. It does not count AI songs deleted from the service. A musician whose track stays online can still be harder to discover if the service stops recommending it. For a listener, meanwhile, a song left in the catalog can remain reachable even if it no longer turns up while browsing recommendations. Those are different experiences of the same moderation decision.

Spotify also says it has deleted more than 75 million spam tracks for serious or repeated rule violations. That figure describes a separate, stronger action. Spotify had already given the same figure in September 2025, rather than disclosing it for the first time in October 2026. Adding the two numbers together would mix different actions and time frames, with no stated accounting for any overlap.

How a Track Loses Its Recommendation

Spotify says it examined songs listeners skipped, ignored or explicitly said they did not want to hear, and found a pattern among tracks its systems had flagged as slop or spam. Over roughly six months, it changed recommendations to exclude those tracks. Listener concerns about such tracks appearing in recommendations have fallen since the changes.

The machinery Spotify describes has more inputs than an AI detector. It combines listener engagement and feedback, upload patterns, metadata and third-party detection tools. It also says it looks for mass uploads, duplicates, manipulative playlist networks and artificially short tracks. Consider two songs made with the same generation tool: one might be uploaded as part of a flood of near-duplicates, while the other might be a single release that listeners choose to hear. A system weighing distribution behavior could treat those uploads differently. That is an explanation of the design Spotify describes, not a finding about how it classified either song.

The strongest case for this approach begins with a fairly ordinary nuisance: people do not need to settle the artistic status of AI music to want repetitive or unwanted uploads kept out of their suggestions. A service can respond to skip behavior and suspicious upload patterns without declaring every use of a model unacceptable. Spotify itself says AI-detection tools have limitations and will face a harder job as generative systems improve. Combining signals addresses a practical problem, though it gives Spotify considerable discretion over what gets heard.

That discretion has a cost for artists. A recommendation decision can affect the chance that someone encounters a track while leaving its page intact. Spotify’s October update, as described in contemporaneous accounts, gives neither a false-positive rate for this intervention nor a before-and-after comparison of listener exposure. Its 50-million count tells us how many tracks it says it acted on, rather than how accurately it identified spam or how much listening changed. For an artist who uses AI for one component of a song, the unresolved question is how the system separates that work from a suspicious upload campaign.

Spotify’s update fits into a longer sequence of controls. It announced AI-related policy changes in 2025, and its existing 1,000-stream threshold determines whether a track qualifies for recorded-music royalties. It has also put AI Persona badges on thousands of artist profiles and says tracks from those personas stay out of editor-curated playlists and recommendations by default, unless a listener follows them. Artists can identify AI use in song credits. Each measure touches a different point in a track’s life on the platform: whether it earns royalties, how its creator is identified, and whether Spotify helps a listener find it.

That sequence explains why “removed” needs a destination attached. Spotify can delete a track for a serious violation, leave it available while withholding promotion, or display information about its maker. Treating those actions as one purge would obscure what a listener can still play and what an artist can still earn or reach. It would also obscure the question behind each action: Did the track break a rule, does Spotify want to recommend it, or does the listener want to know how it was made?

Three Ways to Handle an AI Song

YouTube Music has added an AI label beneath a track’s name for songs generated with AI or using it for major components such as vocals. YouTube says the label applies where AI played a significant role; tapping it brings up more information about the use. The label gives a listener information at the point of encounter. It does not, by itself, prevent the encounter.

Deezer has announced a toggle intended to filter fully AI-generated music from a listener’s experience, with availability planned by the end of 2026. That design puts an explicit choice in the listener’s hands. It also makes classification consequential: if a fully human-made track is wrongly tagged, a listener using the filter could miss it. Deezer says the risk of false positives in its detection system is very low; that is the company’s assessment of a tool on which its planned filter will depend.

Spotify’s recommendation change, YouTube Music’s label and Deezer’s planned filter address related frustrations through different levers. Ranking acts before many listeners see a track. Labeling helps someone decide after the track appears. A filter lets a listener set a standing preference, provided the platform can identify the tracks it covers. A label cannot clean a recommendation feed by itself, while a recommendation adjustment does not give someone who wants to avoid all fully generated music an explicit switch. The services also draw their boundaries in different places: Spotify’s reported count covers slop and spam, YouTube’s label covers significant AI use, and Deezer’s proposed switch targets fully generated tracks.

Those boundaries affect artists as well as listeners. A heavily AI-assisted song could raise a disclosure question without showing spammy upload behavior. An unwanted flood of duplicates could warrant distribution controls regardless of the tool used to make it. A listener might welcome fewer repetitive suggestions while still wanting a label on an AI-assisted song they chose to play. Collapsing those judgments into a single “AI music” number makes it harder to see which decision a platform has made.

Spotify’s 50 million tracks tell you where it has chosen to exercise power: at the doorway to discovery. The count describes how many tracks it says it moved aside; it cannot tell a listener what arrived in their place, or an artist why their song did not.