The Mechanical Licensing Collective launched in 2021 after years of lobbying and negotiation, with a mandate to address the unclaimed royalty problem in digital streaming. It was designed for a world where music distribution meant platforms and listeners. The design assumptions built into the Music Modernization Act, which created the MLC, did not contemplate a world where a meaningful portion of music consumption would involve tracks generated by models trained on existing catalog.
We are now in that world. AI-generated music is not a fringe phenomenon. It is appearing in content libraries, sync placements, background music for video, and increasingly in mainstream distribution channels. The rights infrastructure that processes royalties for those uses was not designed for a source type that has no traditional authorship chain and derives its character from training data rather than composition and performance.
What the MLC Was Designed to Handle
The MLC's core function is to administer blanket mechanical licenses for digital audio transmission and distribution. Under the Music Modernization Act, services that stream or distribute digital recordings can obtain a blanket license from the MLC rather than negotiating with each publisher individually. The MLC then distributes royalties to publishers and songwriters based on usage data submitted by licensees.
This system works well for its intended use case. A streaming service reports plays of specific ISRCs. The MLC matches those ISRCs to registered compositions and ownership data. Royalties flow to the appropriate rights holders. The data flow is linear and the identities are known: a service played a specific recording of a specific composition, and the rights holders of that composition receive a royalty.
AI-generated music disrupts this at the source. The track being played does not have a ISRC tied to a specific recording session. It does not have a registered composition in the traditional sense. The question of who is the songwriter is legally unresolved. The mechanical license that would normally apply to a composition does not have a clear subject to apply to. The MLC's matching infrastructure has nothing to match against.
Where PROs Face the Same Gap
Performing rights organizations face a structurally similar problem on the public performance side. When an AI-generated track is used in a commercial context, a performance license should apply. But the PRO's distribution logic depends on knowing which composition is being performed. Without a registered composition and its associated songwriter and publisher identifiers, there is no distribution path.
Some PROs have begun creating special categories or routing rules for AI-generated content. The approaches vary. Some default to treating unattributed AI-generated works as original compositions by the person who prompted the model, which is legally uncertain. Others hold the royalties in suspense pending legal clarification. Neither approach addresses the underlying question of what obligation exists toward the rights holders of recordings that trained the model.
This is the attribution gap in its institutional form. The money is flowing. Some of it is being collected. But the distribution path for training contributions does not exist yet, because the infrastructure for identifying training contributions was not built into the licensing system that currently operates.
The Technical Layer the Infrastructure Is Missing
The gap is not primarily a legal one, although legal clarification is needed. It is primarily a data layer problem. The MLC and PROs have distribution infrastructure that can operate once they receive the right inputs. The input they are missing is a reliable mapping of which catalog recordings contributed to the training of the model that generated a given track, along with rights holder identifiers for those recordings.
If that data were available in a standardized format compatible with existing rights management systems, the institutional infrastructure could act on it. The MLC could create a distribution path for training contributions. PROs could establish a member notification and royalty holding process. Publishers could receive documentation that their catalog was used in a specific training run. None of this requires new legislation at the operational level; it requires an attribution data layer that does not currently exist in a form that can feed the existing systems.
This is the specific gap that a provenance API is designed to fill. Not to replace the MLC or the PROs, but to provide the input layer that their existing infrastructure needs to function in an AI-generated music world. The API returns attribution data in a format designed for rights management workflow integration, not as a standalone finding.
Who Has the Most Exposure to This Gap
The exposure is not evenly distributed. Publishers with deep catalog in genres that heavily influenced AI music models, specifically popular music from the 1960s through the 2000s, have the most to gain from a functioning attribution layer. These are the catalogs most likely to appear in training data for general-purpose music generation models.
PROs whose membership includes a large proportion of songwriters and composers from those same eras have an advocacy interest in pushing for attribution infrastructure development. Their members have no current mechanism to claim what may be owed to them from AI training use of their work.
We are not saying the MLC or PROs are failing. They are operating the system they were designed to operate. The gap is at the boundary of their mandate, not inside it. Addressing it requires either extending the mandate through legislation, creating a parallel claims infrastructure, or establishing voluntary licensing programs with AI developers that use provenance data as the basis for distribution. Any of these paths requires the same starting point: a reliable way to identify which recordings trained which models.
Building Toward a Working System
The most productive near-term frame for the MLC and PROs is not to wait for legislative resolution before beginning operational preparation. The legal questions will be resolved with or without the infrastructure. If the infrastructure exists when the legal framework clarifies, distribution can begin quickly. If it does not exist, there will be a further delay while the operational layer is built under pressure.
Operational preparation means: mapping existing member catalog to available attribution analysis tools, establishing provisional holding accounts for AI-training royalties pending legal clarification, and creating intake procedures for attribution data from provenance API outputs. None of this commits an institution to a legal position. It is infrastructure work that will be needed regardless of which legal theory ultimately prevails.
The gap between what the current rights infrastructure handles and what the AI music market requires is real. It is also bridgeable with the right data layer. The institutions that start building toward that data layer now will close the gap faster when the legal framework finally gives them a clear distribution path.
Stay Informed
