Spotify has spent a year telling the music business that it takes AI-generated junk seriously. On 20 October the streaming platform put numbers behind the claim: more than 50 million tracks flagged as “slop and spam” have been pushed out of its recommendations during 2026. The verb is doing a lot of work. None of those tracks were deleted. They remain in the catalogue, searchable and playable on demand, simply barred from the algorithmic and editorial machinery that decides what listeners hear next.
- Removed from feeds, not from the library: the 50 million figure sits on top of a separate 75 million tracks deleted outright for serious or repeated rule breaches.
- No listener switch: users still cannot ask Spotify to keep AI-generated music out of their ears entirely.
- Detection is probabilistic: behaviour, upload patterns, metadata and third-party AI detectors, with no published definition of “slop” and no false-positive data.
- AI still ships as a product: the same day brought a Believe and TuneCore licensing deal for fan-made AI remixes and covers.
Escaping the feed, not the library
Spotify keeps its two enforcement numbers firmly apart. The 50 million tracks excluded from recommendations are additional to the more than 75 million removed from the platform entirely over “serious or repeated” violations, a tally the company characterises as a separate effort. Human editors, it adds, do not programme anything identified as slop or spam. The recommendation overhaul landed over the past six months, after engineers studied tracks that listeners skipped, ignored or explicitly said they did not want, and Spotify claims complaints about unwanted material in recommendations have fallen markedly since.
Critics have spotted the weak point: the 75 million deletion figure is identical to the number Spotify published in September 2025, untouched thirteen months later. Either enforcement has slowed or the company has stopped reporting fresh totals. Spotify does not say which.
A probabilistic filter with unpublished rules
No single signal decides a track’s fate. Spotify combines listener behaviour and feedback, upload patterns, metadata and several external AI-detection tools, then looks for the fingerprints of industrial gaming: mass uploads, duplicates, manipulated positioning, fraudulent playlist networks and artificially short recordings built to farm payouts. The company concedes that AI detection is imperfect and will get harder as generative models improve.
It also publishes no definition of slop, no thresholds, no false-positive count and no documented appeals process, so nobody outside the building knows how many human musicians have been caught in the net. Money sharpens the stakes. Since 2024 a track needs at least 1,000 streams a year to earn recording royalties, and Spotify says recordings with strong AI indicators are ten times more likely to fall short of that line. As Morgan Hayduk, chief executive of fraud-detection firm Beatdapp, framed the arithmetic:
“If the denominator has grown because of AI slop, the percentage that each legitimate rights holder or artist will receive is diminished.”
Legitimate AI stays, with terms attached
Spotify insists it is not at war with AI music, and it has built disclosure scaffolding to prove it. AI Credits, launched in April 2026, let artists declare which roles and instruments involved AI, with tens of thousands submitted daily. AI Personas — profiles whose entire identity is machine-generated, numbering in the thousands — are excluded from editorial and algorithmic recommendations unless a listener follows them. More than 300,000 artists carry a Verified by Spotify badge, accounting for over 99% of what listeners search for and more than 70% of them independent. Artist Profile Protection, in testing with thousands of musicians, lets them review releases attributed to them before those records appear on their page and is slated for a full rollout before the end of the year.
The other half of the story arrived the same afternoon: a licensing agreement with French music company Believe, covering its TuneCore distribution arm, for a tool that lets fans generate covers and remixes with AI. Participation is opt-in for artists and labels, royalties flow to the rights holders behind the original recording, each remix links back to its source track, and access is a paid add-on for Premium subscribers. Price, royalty split and launch date remain unpublished. Believe follows earlier deals with Universal Music Group, Merlin and Kobalt, with Sony Music and Warner Music still unsigned.
“Done well, fan creativity can deepen the bond between fans and the artists they love,”
said Charlie Hellman, Spotify’s senior vice president and global head of music.
The contrast with rivals is uncomfortable. Deezer has said it will let users filter AI-generated music out of search before the end of the year — precisely the toggle Spotify does not offer — after reporting roughly 90,000 fully AI-generated uploads a day, more than half of all new tracks on peak days in June.
The arithmetic behind the headline
Co-CEO Gustav Söderström put the catalogue at around 250 million tracks during the first-quarter 2026 earnings call, a figure that sits oddly beside the “more than 100 million” still quoted in the company’s public materials. Take the larger number and roughly one in five tracks has been shoved out of recommendations — not a moderation backlog, but a description of what the catalogue has become. Spotify has chosen to bury that content rather than clean it, a defensible call given how unreliable AI detectors remain and how damaging a false positive would be for a working musician. What is harder to defend is withholding the off switch and staying silent on error rates while launching a paid AI remix product on the same afternoon. An optional setting in the preferences would cost very little, and would say far more than a press release about prioritising human artistry.