Beatport spent this month doing the thing every dance music platform has been asked to do since the first convincing machine-made loop turned up in a promo folder. It drew a line. Fully AI-generated tracks are now withheld at ingestion, the rightsholder gets told, and a detection tool from Beatdapp reads incoming audio and sorts it into human-made, AI-assisted, or fully synthetic. The policy itself is older than the news, dating to early 2025. What changed is that there is now machinery behind it.
The reporting has mostly stopped at the headline, which is understandable, because the headline is clean and the detail is not. Beatport did not ban AI. It banned tracks that are fully or majority AI-generated, and it explicitly kept the door open to AI-assisted work as long as the finished record stays majority human-made.
That word is doing an enormous amount of work, and nobody has defined it.
Every modern record is already part machine
Consider what sits in an ordinary production chain in 2026. Stem separation pulled a loop out of a record that was never released in parts. A mastering service trained on thousands of finished tracks set the final limiter. A plug-in suggested the compression settings and another one repaired a clipped vocal. A drum tool generated four variations of a pattern and the producer kept the third.
None of that is unusual and none of it is new. Some of it has been standard for a decade. Yet each step is a machine making a decision that a person used to make, and if you were auditing the record afterwards, you would struggle to say which side of a majority it fell on.
The question that matters is not whether a tool was involved. It is where authorship sits, and authorship is not a property of the audio file. It is a fact about how the file came to exist.
Detection reads the signal, not the process
This is the gap at the centre of the story. A detection model examines the finished waveform and estimates how likely it is to have been produced by a generative system. It is looking at output. The standard Beatport has written is about input, about how much of the work a person did.
Those two things correlate, and they are not the same measurement. A fully synthetic track that has been resampled, run through hardware, re-recorded off a mixer and edited by hand may read as human. A painstakingly played live take, recorded cleanly and mastered by an automated service, may read as suspiciously tidy.
Any classifier operating at catalogue scale has to pick a threshold, and a threshold produces two kinds of error. Refusing a human record is the one that will generate the noise, because the person on the other end knows exactly how long they spent on it and has no way to prove it. Beatport says rightsholders are notified when something is withheld, which is the right instinct and puts the burden of appeal on the artist.
Why the store is the place this was always going to happen
The argument about generative music has moved several times in three years. First it was about whether the output would be any good. Then it was about training data and whether the models had been built on records nobody licensed. Now it has arrived at the shelf.
That progression makes sense once you notice that the shelf is the only place with the means and the motive. A store has the audio, it has a commercial relationship with whoever uploaded it, and it has a customer base that has told it what it wants. Beatport's own survey found 60 percent of users would not play an AI-generated track in a set and 77 percent want to support human creators. Those numbers are the commercial case, and they are the reason this is a policy rather than a statement.
Meanwhile the other direction is being tested too. Suno has started offering to press machine-generated music onto vinyl, which takes the output and gives it the most human-coded format available. A record shop cannot run a classifier over a sleeve.
What this actually changes for a DJ
Very little this month and quite a lot over a couple of years. The crate has always been a curated space, filtered by distributors, promo pools, and whoever decides what a shop stocks. Adding a probabilistic filter to that list is a difference of kind rather than degree, because the previous filters were people making judgements they could explain.
Three practical consequences are worth watching:
- Label submissions get a new question. If you run a label, you will eventually have to ask artists what they used, and you will have to decide what you do with the answer.
- Promo pools become the soft edge. Anything withheld from a store still circulates in folders and on private links, so the ban shapes the shop rather than the record box.
- Provenance becomes a selling point. Expect more sleeve notes about how a record was made, for the same reason food packaging started listing origin.
The standard will get tested somewhere awkward
The interesting case is not the obvious one. Nobody will defend a fully generated track uploaded by a spam account, and the classifier will catch most of those.
The case that decides how this rule feels in practice is a producer with a real history who used a generative tool for one element of one record, disclosed nothing because nobody asked, and got withheld. What Beatport does the first time that happens in public will tell you more about the policy than the policy does.
Drawing the line was the easy part, and doing it was still better than not doing it. Working out where the line is will take years, and it will be settled case by case rather than by a definition.