A profile with no photo, an abstract sleeve, a 2:14 track whose bass grooves a little too cleanly to have been played by a person. You know that end-of-playlist moment when the algorithm offers up a stranger nobody asked for. The real question is no longer whether this exists, but who gets to decide what stays.

Sweeping and labelling

In a statement reported by Stereogum on 20 October 2026, Spotify says it has removed more than 75 million tracks classed as 'spam', and excluded tracks described as 'slop and spam' from recommendations. It has also added 'AI Persona' badges to thousands of artist profiles and an AI disclosure field in the credits, left to artists who release assisted tracks. The project was announced in the summer; it is now being rolled out.

A digital music library sorted by algorithms, amid mass removals and artist-profile labelling
Generic music interface, studio screen

Two moves, and they should not be confused. Removal is housekeeping: the track disappears and that is that. Labelling is something else — a public admission that the platform itself does not always know what it is hosting, and would rather say so on the artist page. Slop, the mass-generated waste made to fill playlists, is not treated like a fake account: it gets flagged, like a warning sign on a slippery stretch.

Detection is not a verdict

Spotify does not claim to wield a magic detector. The statement lists a mix of signals instead: listener engagement and feedback, upload patterns, metadata, and information from several third-party AI detection tools. None is enough on its own, the platform says in effect — which amounts to admitting that spam is identified by a cluster of clues, not a single proof.

Technical signals and listening data analysed in parallel to estimate a track's origin
Mixing desk and detection signals, generic studio

The next part is more interesting still: 'AI detection technology in particular has limitations,' Spotify acknowledges, adding that the task will only get more difficult as generative AI improves. In other words, the badge is not a verdict but a hypothesis made visible. The credit mention, meanwhile, rests on the goodwill of the person uploading. So you have two pieces of information of different kinds: one from the platform, the other from the artist.

The false positive blind spot

Where is the risk? In the grey area. A short track, functional production, a generic artist name, a faceless sleeve: that is the identikit picture of spam, and it is also the picture of a lot of real work — sound libraries, looping ambient, bedroom beat tapes, first releases put out without a label. Automated sorting does not measure intent. It measures patterns.

Dev Lemons is the counter-example. Her case, documented by Bandcamp Daily on 20 October 2026 in a piece by Hayden Merrick, says the opposite of the problem. A Los Angeles-based artist and comedian, she has played DJ sets opening for Skrillex and passed through a parody rap group, Queef Jerky, long filed under 'shitpost modernism'; she now makes PC-core indie rock tinged with nu metal and trip hop. Her own summary is one sentence: 'The music stands on its own more than it probably has in the past.' The gag and generative plagiarism look alike from a distance; up close, they are not sorted with the same tools.

Three checks before you accept a discovery: the 'AI Persona' badge on the artist page, the AI note in the track credits, and whether the page hangs together — dates, other releases, declared links. None of the three proves anything on its own; together, they at least tell you what the platform and the artist are prepared to stand behind.

Next time a track stops you, open the credits before adding it. Then go and check out Dev Lemons: here is a concrete case of music that looks like a joke and is not one, and a useful exercise in doing your own filtering.

Sources