A catalog’s share of undisclosed AI-generated audio is now part of what it is worth. Music catalog due diligence AI checks exist to measure that share before you close, because the usual checklist of chain of title, registrations and income history says nothing about how the audio was made.
Catalogs are commonly priced as a multiple of representative royalty earnings, so anything that makes those earnings less durable makes the price less safe. Deezer reported in July 2026 that it now receives nearly 90,000 fully AI-generated tracks every day.
Why AI audio changes what a catalog is worth
Undisclosed AI audio reaches a catalog’s price three ways: earnings can be struck out, ownership can be thin, and registrations can be refused. A royalty statement will not warn you about any of them.
Earnings can be struck out: Deezer says it excludes manipulated streams from royalty payments and reports far higher fraud on fully AI-generated tracks than across its catalog. It will also take down AI tracks used for fraud or left unstreamed for six months or more.
Ownership can be thin: the US Copyright Office concluded in January 2025 that generative AI output is protectable only where a human author has determined sufficient expressive elements, and that prompts alone do not count. The Supreme Court declined to hear Thaler v. Perlmutter in March 2026, leaving the human-authorship rule standing. AI-assisted tracks with real human input can still qualify, but this is US law only.
Registrations can be refused: ASCAP, BMI and SOCAN said in October 2025 that they will register partially AI-generated works but not compositions created entirely with AI tools.
Disclosure alone will not close the gap. Spotify says a missing AI credit does not mean AI was not used, and InterSpace Distribution’s mandatory AI disclosure policy covers only new submissions from 25 May 2026.
Music catalog due diligence AI checks: the order of work
Scan every master, join results to earnings, and have a person review flagged tracks, top earner first.
- Get the audio, not just the statements: the master file of every track in scope, listed by ISRC (International Standard Recording Code) beside trailing twelve-month earnings.
- Scan it, newest first. SEAM (Sonic Evidence of AI in Music), an audio recognition service from The InterSpace Group, runs in a web console and via API (application programming interface). Store each verdict and likelihood against the ISRC; the first 1,000 tracks are free.
- Weight by earnings, not track count. Five flagged tracks that carry the catalog matter more than five hundred that earn nothing.
- Have a person listen to the Likely AI and Unsure results, then ask the seller to document each one: contracts, stems, session files and any AI disclosure sent to the stores.
- Publishers: scan the recordings behind your songs, and have writers confirm in writing whether they used AI tools.
- Re-scan anything delivered between signing and closing.
How to turn a scan into deal terms
Convert flagged earnings into a price and a protection, not a vague discount.
Take invented round numbers, not market data: a catalog earns $1,000,000 a year and is priced at 10 times earnings. Review leaves tracks earning $60,000 a year flagged Likely AI or Unsure and undocumented. At the same multiple, $600,000 of the price rests on audio nobody has vouched for.
That is a concrete figure for the seller to answer: carve those tracks out, hold that sum in escrow until they are documented, or adjust the price.
In mergers and acquisitions generally, Haynes Boone lawyers advise AI-specific representations on ownership of AI-generated intellectual property, and say indemnification, special escrows or price adjustments should be considered for AI risks found in diligence. We found no reputable public source that prices AI exposure in catalog deals, so be wary of any quoted standard discount.
A read to inform the deal, not a verdict
SEAM reads the audio itself, not the metadata or the seller’s word, and returns a plain verdict per track (Likely human, Likely AI or Unsure) with a likelihood. It is advisory: it never rejects, holds or approves a track by itself, and a person decides.
A Likely AI verdict is a reason to look closer and ask for disclosure. It is not proof, and it is not something to publish about a seller or an artist.
A scan will not tell you who made a track, who owns it or how any model was trained; those answers stay with the seller’s records, the contracts and counsel. What it adds is a consistent first read across the whole catalog.
Try it first on tracks whose origin you already know, at seam.interspacemusic.com. For what SEAM returns, see our launch article on SEAM for labels and distributors.
