The Consent Stack
Spotify's AI covers and remixes tool is licensing deals one by one, and the chain of title is the bottleneck.
The received wisdom on AI in the catalogue business is that the fight will be about training data. The litigation is, and it will be for a while. But the machinery that will actually decide how AI monetisation works is quieter, and it is already being assembled deal by deal.
Spotify has now signed its third rights-holder agreement for its covers and remixes tool: Universal Music Group in May, Merlin on 4 August, and Kobalt on 13 August. The three are not the same kind of deal: compositions, recordings, and in UMG's case both. The tool will launch as a paid add-on for Premium users, and every generation is licensed. The architecture is what matters, not the launch.
What has to be cleared
A cover and a remix are different clearance problems. A cover is a new recording of an existing song, so it touches one right: the composition, which belongs to the songwriters through their publishers and earns mechanical income. Nothing of the original recording is used, so nothing on the recording side needs clearing.
A remix is the harder case, because it uses the original recording itself. That brings in the master, which belongs to the recording owner and earns share of earnings, and the neighbouring rights position that attaches to that recording in the territories where it is collected. Three rights, three sets of counterparties. The composition is the only layer common to both cases.
Spotify is not signing one deal that clears all of it. It is signing separate agreements with separate counterparties, each covering a different layer of that stack. The UMG deal spans recorded music and publishing because UMG controls both. Merlin works at two levels, since its labels sit under the organisation's agreement and the artists on those labels then opt in individually. Kobalt, which is a publisher and a collection society but no longer owns masters, covers the composition side.
That is the honest shape of the thing. The "3 Cs" that Alex Norström described on the Q2 earnings call, consent, credit and compensation, are a product frame. The underlying reality is a chain of title.
The bottleneck is the chain
Every one of these deals is opt-in. Participation requires the songwriter or the label to consent, and the announcement of the Kobalt agreement is explicit that the mechanism for handling songs with multiple writers across multiple publishers has not yet been set out.
That is where the value is, and where the friction is. A song with three writers at three different publishers needs three consents before it can enter the tool's pool. The party that holds the most complete, most verifiable data on who holds what will be the party that sets the pace. This is why the deals are landing with publishers, not just labels. The publisher is the entity that already maintains the writer data, the splits, and the registration. It is the entity that can answer the consent question at scale.
Consider what this does to the economics. Publishing typically takes around a quarter of total streaming royalties, against three quarters for recorded music. Moving listening towards derivative works does not shift that split by itself, because a stream of a derivative recording still pays a composition side and a recording side. What decides the split is the licence terms, including who is treated as the owner of a recording the platform generated. Those terms are being written now, one negotiation at a time, and the publisher's control over consent is what buys a seat at that negotiation.
Kobalt's own position illustrates the point. Its existing agreement with ElevenLabs requires Kobalt to control 100% of a composition's mechanical rights before that composition can be licensed. That is a chain of title standard set by a rights holder, applied to a platform. The same standard is now being applied, in a different form, to Spotify's tool. The industry is converging on a data standard for AI, and the standard is being written by the people who already own the data.
The quiet part
The loudest story in the same week was Spotify's decision to label AI persona profiles and exclude them from editorial and algorithmic recommendation by default. That is a curation problem, and it is being solved with a badge.
The covers and remixes work is not a curation problem. It is a title problem. It will not be solved by a badge. It will be solved by whoever can produce the cleanest, fastest, most complete answer to the question of who holds what, for which song, in which territory.
The platforms will keep signing these deals in sequence. The publishers will keep raising the data standard. And the parties sitting in the middle of the stack, the songwriters and the managers who negotiate the consents, will find that their leverage has quietly shifted from the training side of the argument to the licensing side.
The consent stack is the real product.