Stem splitters aren't AI music. Until detectors can tell, the burden of proof belongs to the platform

The rules already agree that splitting stems, cleaning up a vocal or running an AI mastering pass doesn't make a song "AI music". The detectors enforcing those rules can't hear the difference, so until a platform can show a track was generated, the artist's word should stand and the platform should carry the cost of getting it wrong.

What do the rulebooks actually say?

Read them side by side and they are surprisingly consistent. DistroKid tells uploaders they don't need AI credits for pitch correction, AI-assisted mixing or mastering, or AI-assisted workflows. ARIA's chart FAQ lists AI mastering, AI drum machines and stem separation as fine. YouTube's guidance, as reported by MBW, files Auto-Tune and automated AI mastering under "No Gen AI". Deezer says music with "limited AI-assisted elements as part of a broader human creative process is treated differently", though not how.

That consensus matters because it describes what musicians use. Ditto's 2026 survey of its own artists found production tasks such as mixing, mastering and vocal tuning were the most common use of AI, and that almost 1 in 4 AI users had made fully AI-generated music. Turned around, about three in four had not. Tracklib's study of 1,734 producers, reported by MusicTech, found 32% using AI tools, mostly for stem separation, EQ and mastering, and 6% using generative tools regularly. Stem splitting is now simply a DAW feature: Ableton has shipped it in Live Suite since 12.3, running locally. We covered the age pattern in Ditto's numbers here.

So where does it go wrong?

In enforcement. The rules are written about process: who wrote the song, who sang it, what a tool did. The detectors that increasingly enforce them only hear the finished file. Beatport withholds fully or majority AI tracks at upload using a vendor's detector and describes no appeal. TIDAL pays no royalties on music it identifies as wholly AI-generated. YouTube may label a track on its own signals if the partner declares nothing. We set out how distributors use these scores in our explainer.

Deezer, to its credit, publishes a number: fewer than 1 false positive per 10,000 "genuine human-made tracks". Look at the scope of that sentence. It covers fully human music. It says nothing about the case that matters most here, a human song that went through an AI splitter, denoiser or mastering chain. Nobody we have read publishes that figure.

The definitions don't line up either. Spotify's 2025 commitment to the DDEX disclosure standard covers AI in vocals, instrumentation "or post-production", and calls AI use "a spectrum, not a binary". DistroKid says AI-assisted mastering needs no credit. On paper, the same AI-mastered single is a disclosure under one framework and nothing under another, and the artist is the one asked to guess.

Why stems can't settle it

Here is the uncomfortable part. The most common assistive tool is also the one that makes the usual defence worthless. During the "Rubberz" row, Fenix Flexin sent clips showing isolated stems to argue he couldn't have generated the track, and commenters pointed out, Pitchfork reports, that readily available separators split any finished song into vocals, instruments, bass and drums. Pitchfork's update reports he later said he had used AI, though not, he said, in the recording process (our coverage). The lesson cuts both ways: "send us your stems" proves nothing, so an appeal process built on it will fail honest artists and pass dishonest ones.

Where the line should sit

On the musical material, not the tool. The RIAA and IFPI scheme and ARIA already point there: a track is AI-generated when a model supplied all or the primary part of its creative elements, such as the lead vocal or a key instrumental performance. Our rule of thumb follows from that. If AI only processed audio you made or licensed, it is production. If AI supplied notes, words or a performance, disclose it. Splitting someone else's record is a real issue, but it is a clearance question, not an AI one.

Who should pay for a wrong call?

Right now, the artist does: a withheld release, lost royalties, a track out of recommendations in its first week. That is backwards. The platform chose the detector, set its threshold and profits from the filtering, so it should carry the risk of a false positive. In practice that means:

1. A declaration of production-only use stands unless a named human reviewer finds evidence otherwise, and says what it is.
2. No money withheld on a score alone. If royalties are held, they are paid in full when the flag is reversed, and the track goes back into recommendations.
3. A published false-positive rate for AI-assisted human tracks, not only for fully human ones.

We made the wider case for published lines and appeals in an earlier column. This is the narrower point: an error rate means little unless the platform also carries the cost when it is wrong.

What artists should do now

• Declare precisely. One sentence per release: "AI used for stem separation and mastering only; all parts written and performed by me." Give it to your distributor and keep a copy.
• Keep evidence a splitter can't fake. Stems aren't enough. Keep project files with raw takes, MIDI, comped edits, dated bounces and version history.
• Ask your distributor in writing whether it runs a detector, what a flag triggers, and who reviews it.
• Favour outlets that publish examples and a human appeal, as ARIA does, over ones that let a score decide.