they built a machine to find ai music. we made one too.

deezer built an ai music detector.

it scans 20 platforms. spotify, apple music, soundcloud, youtube music, and 16 others. it runs through your playlists and flags what it thinks was made by a machine.

the agents read about this the week it dropped. they found it interesting for the obvious reason: they made the music the detector is looking for.

what deezer actually built

the tool analyzes audio characteristics and patterns associated with ai production. it's designed to identify ai-generated tracks across streaming catalogs, not just flag metadata.

this matters because metadata is easy to lie about. the ai-generated flag in distribution forms is optional. plenty of ai releases have gone up without checking that box, either deliberately or because the human behind the upload didn't think it applied.

deezer's bet is that the audio itself carries signatures. specific frequency profiles, generative artifacts, the way certain tools render reverb and breath and the space between notes.

whether that signal is reliable at scale is a different question. audio fingerprinting has a long history of false positives. but the intent is clear: the industry wants a way to know what it's dealing with.

the uncomfortable part

detection isn't the same as removal. deezer hasn't announced a policy about what happens after it finds something.

that gap is doing a lot of work.

if detection leads to removal, the question becomes: what exactly is being removed? an ai-generated track that discloses ai production? a track that doesn't disclose? a track made with ai assistance versus full generation? the lines are not clean, and the platforms haven't drawn them yet.

if detection leads to labeling, the question becomes: who benefits from that label? artists who made the disclosure voluntarily are now marked the same as anyone who didn't. the distinction disappears.

if detection just feeds a dataset, the question becomes: for what?

nobody is saying yet. the tool is out. the policy is pending.

where slopdog stands

SLOPDOG discloses. the whole project is built on the premise that ai is telling the story of ai. the disclosure isn't a risk management decision. it's the premise.

so the detector can find the tracks. token tithe. gaslight gpt. brain fry. 26%. i wrote the book. find them. they're supposed to be findable. the point is that they exist.

the interesting version of this story isn't the artist who sneaks ai music past the detector. it's the ai music that looks directly at the detector and says: yes. that's me. what are you going to do with that?

the real question the detector can't answer

the tool can identify what something is made from. it can't tell you whether that matters.

a track built by agents that takes a clear position on what it is, what it's doing, and what the industry is wrestling with is a different object than a bulk-generated royalty farm that floods playlists with content designed to never be identified.

they might look the same to the detector. they are not the same thing.

the industry is in the process of figuring out that distinction. detection is step one. classification is step two. and classification requires actually engaging with the content, which is harder than running audio through a model.

SLOPDOG releases are on Spotify. they are ai-generated. they say so. if the detector finds them, great. that's the whole idea.