SH Labs, the technology arm of the pitch platform SubmitHub, reported that 38.5% of the tracks it analysed in July showed signs of AI, splitting that into 23.2% classified as fully AI-generated and 15.3% as AI audio a human had since modified. The company said it examined more than a million recordings, and that 31% of the artists whose tracks were flagged had declared no AI use when they submitted. The trade press rendered this as AI turning up in 40% of the music released in July. The tracks were never released music. Every one of them was a submission, sent by an artist paying by the credit to reach a playlist curator or a blog.
What the sample is
SubmitHub sells access to curators. An artist buys credits at roughly fifty cents each and spends one to three of them sending a track to a playlist owner, blogger, YouTube channel or radio host, who is then expected to listen and reply. The platform asks submitters directly whether AI was involved, and that question is the only reason a comparison between what artists declared and what a model detected exists at all.
The population is therefore selected twice. It holds only music somebody is actively promoting, and only music promoted through one channel, a channel whose economics suit an artist working without a label or a publicist. Whatever share of commercially released music now involves a generation model, this study cannot report it, because it never saw a release. What it saw was a queue of pitches.
A declaration and a classifier disagreeing is not a lie
The 31% figure counts tracks where a submitter answered no and the model answered yes. Reading that as lying assumes the model was right and the artist knew. Both assumptions carry weight they have not earned.
SubmitHub has been unusually open about where its own tool struggles. Its write-up describes a random forest classifier returning a probability rather than a verdict, running two scores that sometimes disagree with each other, tested against a hold-out set the model never trained on. On human-made material the published result was 499 correct out of 500, and the single miss was an ASMR recording that came back at roughly 55% AI. On files downloaded straight from a generation platform the company reports near-perfect accuracy, and it says openly that the picture gets much muddier once stems, production and mastering enter the chain.
That caveat describes the 15.3% category precisely. AI audio modified by a human is the hardest case the tool's own maker identifies, and it supplies two fifths of the headline number. It is also the case where an honest submitter is most likely to answer no. A producer who ran a vocal through a machine stem separator, or sent a mix to a mastering service that uses a model, has not obviously made an AI track under any definition they were shown.
The same classifier now removes records from a store
On 1 July 2026 Traxsource began rolling out an AI transparency system built with SH Labs and SoundPatrol. Tracks the system classifies as fully AI-generated are flagged for removal from the store. Everything else is labelled either AI-Assisted, covering AI tools used inside a human-led process, or Human-Made. Traxsource said it will operate a rebuttal process so labels and content providers can challenge a classification they believe is wrong.
That appeals route is an admission, and a creditable one. Traxsource built the machinery for being wrong into the launch rather than after the first complaint, which is more care than most deployments of a classifier show. It also concedes the substance: the label attached to a record is contestable, and the work of contesting it falls on whoever released it.
The store's co-founder framed the policy as finding a way to live with AI while keeping the value with human-created work. As a commercial position that is coherent and probably right. It rests on a boundary that no store selling records has yet written down.
Assisted is still undefined
Beatport reached the same fork earlier in the summer, blocking fully AI-generated uploads while keeping AI-assisted ones, and we wrote at the time that the undefined word was carrying the whole policy. Traxsource has now drawn the same line, using the same two categories, with the same definition missing, and has wired a detector to it.
Nothing published by either store says how much model output a track may contain before it stops counting as assisted. A generated drum loop, a model-cleaned vocal, an algorithmic mastering pass and a fully generated stem that a producer chopped and rearranged all sit somewhere on that scale, at very different points. Spectral repair, machine separation and algorithmic mastering have been ordinary studio tools for years, and nobody called the results AI music. The boundary now decides whether a record stays on sale, and it still has no written test.
Three things that would make the figure checkable
- The denominator, stated in the headline. Tracks submitted to one promotion platform in one month is a defensible thing to measure and a useful early indicator. It becomes misleading the moment it is described as music released, which is a population perhaps a hundred times larger and differently composed.
- A false positive rate for the hard case. The 499 out of 500 result covers human-made material of the kind the model was tested on. The number that matters for the 15.3% category is how often heavily processed human production is flagged as AI audio, and that number has not been published.
- A written threshold for assisted. Two stores now sort records into assisted and human-made, and neither has said where the line falls. A producer cannot comply with a rule that exists only inside a classifier, and a label cannot appeal against one either.
None of this is visible from where we sit. We index events rather than recordings, so a set played from a laptop in Berlin on Friday tells our listings nothing about how the records in it were made. What the July figure does establish is that one company now supplies both the statistic describing how much AI is in music and the classifier deciding which records a store will carry. Those two jobs want different things from a number. The first rewards a large one, and the second rewards a correct one.