Labels rolling out on 14 October depend on a three value declaration from rights holders and on YouTube's own detection when the field is blank. "Partly" still has no fixed definition for listeners.
YouTube Music began rolling out AI labels on 14 October. The label is decided long before a listener sees it, when a distributor fills in a spreadsheet column or an XML message, or leaves it blank.
The label is only as consistent as the declarations behind it.
YouTube's announcement describes a badge on the player page for songs "created primarily using generative AI". Tapping it opens a menu with more information, and the labels will also expand to the main YouTube app. The weak spot is "partly AI" tracks, which nobody has defined at the point where a listener sees them.
The blog says YouTube is "teaming up with music partners and using our own detection tools" to identify tracks. The detail sits in a separate YouTube Help page on disclosing Gen AI content. Partners are asked to declare each asset as Fully Gen AI, Partly Gen AI or No Gen AI. A blank is read as "unknown".
Partners deliver that declaration through channels that already carry credits and rights data: DDEX ERN 4.3.2 or later (the industry standard message format for music releases), CSV templates, or attributes set in YouTube Studio. In the CSV version the column is called contains_ai. It takes four values: all, partly, none and unknown.
The label a listener sees is the last link in a supply chain. A label, distributor or aggregator classifies the track. The classification travels in a feed built for royalty statements and territory rights, and YouTube reads that field alongside its own detection signals.
The classification carries the ambiguity. The blog speaks of songs made "primarily" with generative AI and gives composition and vocals as examples. The Help page uses three bands instead, and its scenarios do not line up neatly with that word.
The Help page's examples sort like this. A producer who typed a prompt into a generator and downloaded the result made a Fully Gen AI track. A human vocal over an AI bassline or string layer is Partly, as is a human recorded song with an AI generated video. Standard pitch correction and AI mastering sit in No Gen AI.
Only the first group matches the blog's "primarily" without interpretation. For the rest, the Help page says: "It is up to partners to interpret the Gen AI definitions provided by DDEX in relation to their content." Two distributors given the same stems and video could return different answers.
A track with an AI string layer might be Partly at one aggregator. Another might call the layer production polish and mark it None. The listener sees a label, or no label, and cannot tell which reading produced it.
YouTube does not depend on partners alone. Its own detection tools run alongside the declarations. When partners supply nothing, the Help page says YouTube "will rely solely on those signals and may designate content as fully or partly Gen AI". That is a backstop for catalogue that arrives blank.
When a field is blank, YouTube says it will treat the track as unknown unless its own signals decide otherwise, and those signals can return fully or partly. A blank is therefore a gamble on a detector whose threshold nobody outside YouTube has seen.
The delivery spec also lets partners correct themselves. In the CSV template, "unknown" clears a value that was previously misdeclared. A distributor that labelled a catalogue "partly" in good faith can clear it if the artist considers the record human.
For a delivery desk, the working rule is to set contains_ai explicitly on every asset. Leaving the field blank hands the decision to detection. The next step is a house rule for "partly", written down before the badge reaches the main app, where any dispute has a far larger audience.
That rule has to say whether AI stems, AI assisted mastering and AI generated visuals each count. The Help page already makes calls on two of them: mastering is No Gen AI and video visuals are Partly. A house rule that follows the Help page will be defensible. One that departs from it needs a reason.
Silence carries the chance that detection designates a track as fully or partly AI without the rights holder having made the case. An explicit declaration costs a decision up front and buys control over the outcome. Labels that commission visuals from generative tools for lyric videos and promotional clips face that decision most often, because the Help page already counts such visuals as Partly.
YouTube's general video label is the closest precedent, and it points the same way. The Help page on "Made with AI" disclosure says the label follows one of three things: a creator's disclosure in YouTube Studio, use of YouTube's own generative tools, or valid Content Credentials showing the whole video was made with AI.
YouTube also carries forward disclosures from C2PA 2.1 or later credentials, which are tamper evident records of how a file was made. It may proactively label content that was not disclosed. In each case, a provenance signal is declared upstream and the badge is the downstream effect.
The music version moves the same logic into the commercial feed. The people filling in the field are intermediaries with many thousands of records and little time to deliberate over each one.
As of 14 October 2026, the labels have only begun to roll out, and YouTube says they will also expand to the main app. When they do, the arguments will turn on a word, "partly", that the platform has described with examples but not defined.