AI Music Checks for Labels, Publishers and Sync

Why an A&R ear is no longer the control

Experienced listeners are good at spotting bad generated music and poor at spotting good generated music. That asymmetry is the entire problem. The tracks that reach a label inbox or a sync brief have usually been through a human pass — re-sung, re-arranged, re-mixed, mastered by someone competent — and every one of those stages removes the artefacts that made the first generation obvious. What arrives is not raw model output; it is a hybrid.

Hybrids break the binary question. A demo can be generated end to end, generated and re-recorded, human-written with a generated backing, or entirely human with a synthetic vocal double. Those are four different legal and commercial situations, and no amount of careful listening separates them reliably. Acoustic analysis narrows the field, but it cannot tell you which of the four you are looking at either.

So the control has to move upstream, from perception to procedure. What protects a catalogue is a documented intake process: a declaration you collect, a warranty you sign, records you retain, and a measurement you take as a sanity check rather than as a verdict. Labels that already handle sample clearance have most of this muscle memory; the new part is that the thing being declared is a tool rather than a recording.

• Polished hybrids are the common case, not raw generated output
• Listening detects incompetence, not synthesis
• Four different provenance situations produce similar audio
• Procedure and paperwork outperform ears and detectors

The intake questions that actually work

Ask narrow, factual questions with checkable answers. Broad ones invite a shrug: 'is this AI?' is answered honestly by 'partly, like most records made this year'. Narrow ones create a record you can rely on later, and they make an untruthful answer an explicit misrepresentation rather than a misunderstanding.

Word the questions per stage of production, because that is how the work was actually made. Composition, lyrics, lead vocal, backing vocals, instrumental performance, and mastering are separate lines in every credit sheet already, and generative involvement lands in some but rarely all of them. A declaration that resolves to a per-stage answer is directly transcribable into your credits and your delivery metadata.

Then ask for artefacts rather than assurances. A project file with a plausible edit history, stems that isolate the parts a person claims to have played, an unmastered rough, a phone video from the tracking session — any one of those is worth more than a signed statement, because a signed statement can be honest and still be wrong about what a collaborator did.

• Which generative tools were used, on which stage, and on which version
• Is any voice on the recording synthetic, converted, or modelled on a named person
• Who wrote the lyrics and the topline, and can they produce a draft
• Under whose account and licence tier was any generated material produced
• Can you supply stems, a project file, and an unmastered rough

Where a detector belongs in the workflow

Use acoustic analysis as a triage step, not as a gate. Run the best file you have — a lossless master, never a messaging-app re-encode — and treat the result as a prompt to ask better questions. A high probability on a submission where the declaration says 'no tools used' is not proof of a lie; it is a mismatch worth one more email. A low probability tells you very little either way, because a competent hybrid is designed to look like a record.

Be aware of the failure mode that matters most in this seat: fully synthesised, tightly quantised electronic production is the hardest case for every acoustic detector, including this one. In our own measurements, three of twelve human-authored synthesiser-only reference renders read above 70% AI probability. If your roster skews electronic, a detector used as a gate will reject real artists at a rate you would not accept.

Record what you ran and when. A dated note saying 'analysed the delivered master on this date, result inconclusive, declaration on file' is exactly the kind of diligence record that matters if a dispute arrives eighteen months later — and it is far more defensible than a screenshot of a percentage with no context.

• Analyse the delivered master, not a compressed copy
• Treat any figure as triage, never as a pass/fail gate
• Expect false positives on synthesiser-heavy human work
• Log the date, the file analysed and the outcome

Contract language that survives the question

Most existing warranties were drafted for sampling and ghostwriting, and they nearly work. The gap is that 'the work is original and does not infringe' does not address whether a machine produced part of it, nor whether the resulting material is protectable at all. Both matter commercially: in several jurisdictions purely machine-generated material may attract no copyright protection, which means there may be nothing exclusive for you to license or enforce.

Add three things. A representation that lists generative tool use per stage, with a schedule the artist completes rather than a blanket denial. A warranty that no voice on the recording imitates an identifiable performer without written consent. And a remedy that fits the risk — usually a right to require replacement of an offending element, plus indemnity, rather than immediate termination, because termination punishes you as much as them once a release is scheduled.

For sync and library work, mirror whatever your end clients demand and assume it tightens. Broadcasters and brands are the strictest link in the chain, and the exposure they fear is a performer's likeness complaint, not a detection score. A cue that cannot be warranted clean is a cue you cannot place, however good it sounds.

• Per-stage disclosure schedule rather than a single yes/no representation
• Explicit likeness warranty covering synthetic and converted vocals
• Acknowledgement that protectability may be limited for generated elements
• Replacement-and-indemnity remedy rather than automatic termination

What to do when the reading and the story disagree

The awkward case is a plausible artist, a clean declaration, and a measurement that leans synthetic. Handle it as a documentation request, never as an accusation. Ask for the stems and the project file, ask who played what, and listen to the unmastered rough. Almost every honest artist can answer within a day, and almost no volume uploader can.

Keep the language neutral in writing. 'Our routine intake check was inconclusive on this master, so we need the standard provenance pack before we proceed' is a sentence you can send to anyone without insult and without creating a defamation exposure. Naming a generator, or asserting that a track 'is AI', is exactly what an acoustic estimate cannot support and what a lawyer will quote back to you.

If provenance cannot be produced, decline on documentation grounds rather than on detection grounds. That reason is true, defensible, and consistent — and it is the same reason you would decline an uncleared sample, which is a decision your organisation already knows how to make.

Retrofitting an existing catalogue

The awkward question for any established catalogue is what to do about material already signed under older paperwork. Auditing every track acoustically is a poor use of time: the false-positive behaviour on electronic material would generate a queue of accusations you cannot substantiate, and a low reading on a hybrid would give you false comfort. Sort by exposure instead.

Prioritise anything with a synthetic or converted lead vocal, anything pitched into sync where a warranty has already been signed, and anything acquired as a bulk catalogue rather than signed artist by artist. Those three categories carry nearly all the real risk. For each, collect the declaration retrospectively — most artists will answer a direct, non-accusatory question — and record the answer against the asset.

Where a bulk acquisition cannot be documented at all, treat it as unwarrantable rather than as guilty: exclude it from sync pitches that require a clean declaration, keep it in distribution if rights are otherwise sound, and note the gap in your own records. An honest 'unknown' in a database is an asset. A guess is a liability.

• Do not mass-scan a catalogue; triage by commercial exposure
• Synthetic vocals, signed sync warranties and bulk acquisitions come first
• Collect declarations retrospectively, per asset, in writing
• Record unknown as unknown rather than resolving it with a guess

Synthetic and converted vocals: the risk that actually bites

Of everything in a submission, the voice is where commercial damage concentrates. A generated backing track that turns out to be generated is an embarrassment and a metadata correction. A lead vocal that a named performer believes imitates them is a complaint from a person with representation, and it arrives regardless of what any analysis says about the file. That asymmetry should shape how much scrutiny each element receives.

The technical landscape is worth understanding at a working level, because the three cases have different paperwork. A fully generated vocal comes from a model with no specific person behind it, and the question is licence terms and disclosure. A converted vocal — a real singer's performance transformed to sound like a different timbre — depends entirely on whose voice supplied the target, and consent for that target is the whole issue. A cloned vocal trained on an identifiable artist's recordings is the case nobody in the chain will accept, and it is also the one most likely to arrive already sounding professional.

Practically, ask two questions and insist on real answers: whose voice is the target, and where did the training or reference material come from. Then ask for the isolated dry vocal. Converted and cloned vocals are far easier to reason about without the reverb, doubling and mastering that a finished mix hides them behind, and a singer who really sang the part can supply the dry take in minutes.

Treat consent as documentation, not as reassurance. If the target voice belongs to a session singer who agreed to a conversion, get that in writing with the singer identified. Verbal assurance from the producer that 'she was fine with it' is precisely the evidence that evaporates when a placement earns money.

• Generated, converted and cloned vocals raise different questions — ask which
• Always request the isolated dry vocal for anything with a suspicious lead
• Get consent for a target voice in writing, naming the person
• Cloned vocals of identifiable artists are commercially unplaceable, full stop

How generative involvement changes the deal, not just the diligence

Once you know how a track was made, the interesting question is what that changes commercially, and the honest answer is that it mostly affects certainty rather than value. Audiences do not audit provenance; the market prices exclusivity, and exclusivity is what generative involvement can undermine. If part of a work may not be protectable, then your exclusivity claim over that part may be thinner than the contract implies, and every downstream promise you make inherits that thinness.

That argues for structuring around the uncertainty rather than pretending it away. Where a significant element is generated, consider narrower warranties matched by narrower promises to end clients, shorter terms so you can revisit as the legal position settles, and splits that reflect who actually contributed authorship rather than who operated the tool. None of that requires a moral position on generative music. It is the same logic used for a track built on an uncleared interpolation: price the risk, or remove it.

There is also a positive case to be explicit about. A well-documented hybrid — human topline, human vocal, generated bed, all declared, all licensed on a commercial tier — is a cleaner asset than an undocumented track whose producer will not answer questions. Documentation, not the absence of tools, is what makes a catalogue defensible. The organisations that get this right will end up with better records than their competitors on human material too, because the process improves paperwork across the board.

• Generative elements can weaken exclusivity, which is what the market pays for
• Match narrower warranties with narrower promises to end clients
• Attribute splits to authorship, not to whoever ran the tool
• A declared, licensed hybrid beats an undocumented human track

Making it operational without adding headcount

This only works if it lives in the tools your team already opens. Put the disclosure schedule into the submission form, so it arrives with the demo instead of being chased afterwards. Put the provenance pack into the delivery checklist that already collects artwork and split sheets. Put the analysis step into the same pass where someone checks the master for clipping and true-peak.

Name one owner. Sample clearance works in most organisations because a specific person is accountable for it; generative provenance fails when it is everybody's job. That person does not need to be technical — the job is chasing declarations and filing artefacts, not interpreting spectrograms.

Review the wording every couple of quarters. Tool names change, industry metadata fields for declaring generative involvement are arriving, and the questions worth asking in a year will be more specific than the ones worth asking today. A process that is reviewed on a schedule stays useful; one written once becomes a form nobody reads.

• Disclosure schedule inside the existing submission form
• Provenance pack inside the existing delivery checklist
• Acoustic check alongside the existing technical QC pass
• One named owner, and a quarterly wording review