Skip to the content
Dead Writing Theory

The Court of Public Prose

From the clerk's desk

Spotify and AI Slop: What a Removal Count Can and Cannot Tell Us

Spotify says it removed over 75 million spammy tracks in 12 months. That is a self-reported enforcement count with no stated base, so it cannot show how much AI music there is or whether music sounds more alike.

By Jack Stovell · Published

Short answer: the 75 million figure is a count of removals that Spotify reported about its own platform, with no stated base. It shows the spam filter was busy. It cannot show what share of uploads was spam, how much of it was made with AI, or whether music is drifting towards one sound.

What Spotify has said it is doing

The headlines fold three separate measures into one.

Removal. On 25 September 2025 Spotify wrote: “In the past 12 months alone, a period marked by the explosion of generative AI tools, we’ve removed over 75 million spammy tracks from Spotify.” Note the word “spammy”. The post does not call them AI tracks. It lists the tactics its new spam filter targets as “mass uploads, duplicates, SEO hacks, artificially short track abuse”, and says the filter will “identify uploaders and tracks engaging in these tactics, tag them, and stop recommending them.” The same post says Spotify is helping develop an industry standard for AI disclosures in music credits through DDEX, covering uses such as “AI-generated vocals, instrumentation, or post-production”. There is no ban on AI music. Spotify says the disclosure work is not about “punishing artists who use AI responsibly”.

Attribution. On 24 March 2026 TechCrunch reported that Spotify was beta testing “Artist Profile Protection”. Artists in the beta can approve or decline releases delivered under their name before those releases reach their profile. Only approved releases count towards stats and recommendations.

Labelling. On 11 August 2026 Spotify announced an “AI Persona” label for artist identities it judges to be AI-generated, according to Engadget and TechCentral. The label is due to appear from mid-September on profiles, in search and in playlists. Artists can declare themselves. Spotify says it “won’t rely on self-disclosure alone” and will use “human review alongside AI investigative tools”. Labelled artists are left out of personalised recommendations by default. The label sits on the artist, and individual tracks are not marked by it.

Nothing I fetched mentions an independent audit of any of the three. Each one is the platform describing its own enforcement.

A removal count is a prevalence figure, and an unfinished one

Prevalence asks how much of something there is. A removal count gives a partial answer: how many items the platform’s filter caught. Three gaps stop it from going further.

First, there is no denominator. Spotify’s post gives no total for uploads in those 12 months and no catalogue size. Over 75 million out of how many? Without a base there is no rate, and without a rate you cannot say whether spam is a sliver of uploads or most of them, or whether it is growing.

Second, the platform owns the category. “Spammy” covers mass uploads, duplicates, keyword stuffing and very short tracks. Some of that will be machine-made. Some will be ordinary fraud using human-made or recycled audio. The count does not split the two, so it is not a count of AI music.

Third, it counts only what was caught. Tracks the filter missed are left out, and tracks it removed wrongly are included. Nobody outside Spotify can check either.

Compare a study that sets out to measure prevalence. The websites nobody wrote stated its sample of sites and estimated a share with a detector, which carries the usual false-positive caveat. You can argue with that number because you can see what it was measured against. Spotify’s figure gives you nothing to hold it against.

Why ‘the same songs again’ is a different claim

The common leap goes like this: huge removals, so AI music is everywhere, so everything sounds the same. That last step asks a different question.

Convergence asks whether output has become more alike, whether the range has narrowed compared with an earlier period. Heavy prevalence can come without convergence if the machine-made tracks vary widely. Convergence can come with modest prevalence if a few popular tools pull many makers towards the same sound.

A removal count says nothing about either. It does not tell you how similar the removed tracks were to each other, or to what came before. A filter also chose the removed pile, so studying it would tell you more about the filter’s habits than about music.

The court’s own convergence evidence looks different. The shrinking range fixes a question about whether writing became less varied and measures similarity against a baseline. A convergence claim needs that shape before it can be weighed.

What would make the claim checkable

Each measure needs something specific before an outsider could test it.

  • Removal: the total uploaded in the same 12 months, the “spammy” category broken down by tactic, and a way for an independent party to audit a sample of tracks removed and tracks left up.
  • Attribution control: how many releases artists declined, out of how many delivered under their names, once the tool leaves beta.
  • Labelling: how many profiles were labelled, how many by self-declaration and how many after review, and how often the “AI investigative tools” get it wrong. A label applied by a classifier inherits the classifier’s mistakes. The detector that convicts the innocent is the court’s reminder that text detectors have flagged human writing as machine-made. Nothing exempts a classifier of artist identities from the same problem.
  • Convergence: a fixed sample frame, a similarity measure chosen before anyone looks, and a comparison with an earlier period.

None of this is exotic. It is what the theory page asks of every case: fix the question and the sample first, then measure.

Where this court’s jurisdiction ends

The cases here test writing on the web. They do not test music, and nothing on this docket says anything about how songs sound. When I point to a case above, I am borrowing its method: fix the sample, name the denominator, measure against a baseline, and say what the result cannot show. The verdicts stay with written text. Whether a given similarity measure carries over to audio is a question for people who study audio.

I am also not arguing about whether AI music is good or bad. This post is only about what a number can support. For what people mean by the term itself, see what does AI slop mean. For the same exercise on another public claim, see the Microsoft CEO piece.

How to read the next platform announcement

Ask four questions, in this order.

  1. What was counted, and who defined the category?
  2. Out of how many?
  3. Could anyone outside the company check it?
  4. Does the claim in the headline match the thing that was counted?

Most enforcement announcements fail at least one. That is no scandal. It is simply what a self-report is. The trap is the size. “Over 75 million” sounds like proof of a flood. What it proves is that, over 12 months, a filter removed over 75 million things Spotify classed as spammy, from a total it did not give.

Sources

Questions people ask

How many AI tracks has Spotify removed, and what does that number tell you?

Spotify has not given a count of AI tracks. On 25 September 2025 it said it had "removed over 75 million spammy tracks" in "the past 12 months alone". Spam there means tactics such as mass uploads, duplicates, SEO hacks and artificially short tracks, and some of that may be AI-made. The figure tells you the filter removed a lot. It gives no total uploads to compare against, and Spotify mentions no independent audit.

Is Spotify banning AI music?

No. Spotify's September 2025 post announced no ban, and it said its disclosure work is not about "punishing artists who use AI responsibly". Its measures are a spam filter, an industry standard for AI disclosures in music credits, a beta tool reported on 24 March 2026 that lets artists approve or decline releases under their name, and an "AI Persona" label for artist identities, announced on 11 August 2026 to appear from mid-September.

Does a removal count show that music is converging on one sound?

No. A removal count measures how many items a filter took down. Convergence asks whether output has become more alike over time. That needs a fixed sample, a similarity measure and a comparison with an earlier period. A removal count supplies none of those.

What would a fair test of 'AI slop' on a streaming platform need to fix before measuring?

Four things, set before anyone looks at results: the population (all uploads in a stated window, or the whole catalogue), the definition of the category being counted, the time window, and, for any claim about sameness, the similarity measure and the baseline period. Without those, a figure can be read to fit whatever conclusion you started with.

Can I turn off AI content on Spotify?

None of the sources I fetched describe a listener setting for that. What they do report is that artists carrying the AI Persona label are left out of Spotify's personalised recommendations by default.