The artifact at the center of this is worth examining first, since everything else follows from it. Harris told The Verge that alongside similar-sounding vocals, he hears a sharp hiss running the whole length of a suspected track, and he traces that hiss to models which begin with a big block of white noise and then guess at waveforms by referencing data taken from real songs [5][6]. Whatever the merits of that account of how the models work, the test runs on the audio and on nothing else, so it cannot separate a generated track from any other track that happens to share the artifact. The second strand of evidence in the piece is comparative: Harris and other fans noticed that "Take Me (To The Moon)" resembles another person's AI-generated track [11].
Set that against what he offers as the counterweight, which is his own session: Ableton Live, a Launchkey 49, Push 3, Launchpad X, a couple of analog synths, Serum, Diva, Kontakt, and what he describes as hundreds of decisions per song [12]. He calls the output of the tools "a kind of decoy art form" [7]. What he is describing is labour, not provenance, and it is the only thing he has to offer in place of a record.
Teams shipping these tools tend to build for a user who discloses, on the theory that the tool is a legitimate part of the workflow. What The Verge documents is messier: some producers own up straight away, and others denied it until public scrutiny made denial untenable [2]. In the two cases named in the story, three acts sit under public suspicion and not one has said anything either way [13]. That silence is standing in for the entire adjudication process. Every disclosure event described is something a person says, either an accusation formed by ear or an admission produced under pressure, and none of it comes from a platform or a file [14]. The Verge itself notes there is no definitive evidence in the MANSA case and that Harris may be wrong [10].
The grid worth drawing has two axes. First, whether your tool emits a disclosure that survives export, re-encode, and upload to somebody else's platform. Second, whether a user can disclose partial use, one hook or one vocal, without being read as having generated the entire record. No signal and a binary frame is where dance music is now, and honest users are tried by ear along with everyone else. A durable signal with a binary frame is worse for your best users, because the producer who used your tool for eight bars gets filed with the producer who typed a sentence, so they will route around the signal. A graded frame with no durable signal lets people explain themselves with nothing behind the explanation. Only the fourth cell makes candour cheaper than silence.
My recommendation is the fourth cell, and the cost is real: a signal that survives re-upload is also a permanent accusation that anyone can harvest into a list, and some users will simply pick the tool that emits nothing. The forcing question for whoever owns this on Monday is what an accused user can hold up on Friday, and whether it came from you. If the answer is a checkbox visible only inside your own app, the argument lands in your support queue, and your users keep answering it by ear.
