The technology to trace which recordings contributed to an AI model's training run exists and is being refined. What most performing rights organizations do not yet have is a workflow for what happens when that attribution data arrives. The gap between having the data and being able to act on it is primarily an operational one, not a technical one. PROs that close that operational gap now will be positioned to handle the first wave of AI-attribution claims when they arrive, rather than building the workflow under pressure.
This piece looks at the specific workflow components a PRO needs to address, based on how attribution data is structured and how it maps against existing royalty calculation infrastructure.
What Attribution Data Looks Like in Practice
An attribution analysis of an AI-generated track returns a set of catalog matches: recordings identified as having measurable influence on the model's training, each with a confidence score and a rights holder identifier. The output is not a single clean answer. It is a ranked list of contributing recordings with varying degrees of confidence, because training influence is probabilistic, not binary.
This probabilistic structure is different from what most PRO intake systems are designed to handle. Standard royalty distribution assumes a clear chain: a registered composition, a performer, a licensee, and a specific license type. Attribution claims for AI training sit outside those categories. The composition may not exist as a registered work in the PRO's database. The AI developer may not be a registered licensee in the traditional sense. The license type is being negotiated or litigated in parallel with the claim itself.
Understanding this mismatch is the first step. A PRO that treats AI attribution claims as a standard intake item will encounter failures at multiple workflow stages. A PRO that creates a dedicated intake path for attribution data can process that data without disrupting existing workflows.
Intake: Creating a Parallel Claims Path
The practical recommendation is to build a separate intake path for AI attribution claims rather than forcing them through existing licensing intake. This path needs to accommodate three things that standard intake does not handle well: probabilistic match confidence levels, catalog identifiers that may need resolution before a rights holder can be contacted, and the fact that the underlying legal basis for the claim may still be uncertain.
For the confidence level problem, the intake system should treat attribution data as a ranked provisional claim, not a confirmed right. A recording with a 0.87 confidence match against a training corpus is a strong candidate for follow-up. A recording with a 0.34 match is still worth logging but should be flagged differently. The intake system needs a confidence threshold that triggers active claim processing and a lower threshold that puts records into a holding queue for when the legal framework clarifies.
For the catalog resolution problem, attribution analysis returns identifiers. Those identifiers need to map to registered works and their associated rights holders in the PRO's database. Any organization that does not have clean ISRC-to-rights holder mappings will hit an immediate wall at this step. Catalog hygiene is the prerequisite, not an afterthought.
Rights Holder Notification: Before the Legal Framework Solidifies
An under-discussed part of the PRO's role in this space is notification. When attribution data identifies a member's recording as a contributor to a training run, there is a reasonable argument that the PRO has an obligation to inform that member, even if the claim pathway is not yet fully clear.
Notification does not require a complete legal framework. It requires a record that says: we have evidence your recording contributed to the training of model X with a confidence level of Y. What the rights holder does with that information, whether to pursue a licensing negotiation with the AI developer, hold the record as documentation for future litigation, or take no action, is their decision. The PRO's role is to ensure they have the information to make that decision.
PROs that have invested in member communication infrastructure around digital licensing are better positioned to handle this. Those that still rely heavily on manual notification processes will need to think about how attribution notification scales. The volume of records generated by training attribution analysis will be larger than what manual processes can handle efficiently.
Royalty Calculation: Where the Framework Gap Is Sharpest
The most contested operational question is how to calculate royalty weights when multiple recordings contributed to a single model's training. This is genuinely unsettled. There is no agreed methodology for translating a confidence score and a catalog contribution into a royalty amount. Different legal theories imply different calculation approaches.
We are not saying a definitive methodology exists and PROs are missing it. The methodology is still being developed through negotiation, litigation, and emerging regulatory guidance. What PROs can do is document their current thinking about how they would approach this calculation, so that when external guidance arrives, they can map it to an existing framework rather than starting from nothing.
The key variables in any calculation methodology are: the number of recordings from the PRO's catalog that contributed to a specific model's training, the relative weighting of each contribution based on confidence scores, the commercial value of the outputs generated by the model, and the applicable license rate if a rate has been set. PROs that have thought through how their systems would handle each of these variables are weeks ahead of those that have not.
Coordination with Publishers and the Mechanical Side
AI training attribution involves both the sound recording rights and the underlying composition rights. A PRO collecting on behalf of composers and lyricists has a different claim path than a label or distributor collecting on behalf of the recording. In practice, attribution analysis often identifies the recording first, and the underlying composition must be traced from there.
This creates a coordination need between PROs, the Mechanical Licensing Collective, and publishers. Where a recording match is identified, the PRO should have a mechanism for flagging the associated composition to the relevant licensing body. This is not currently a standard workflow step. Building it requires knowing which catalog recordings in the PRO's database have corresponding composition registrations, and having the relationships in place to share attribution data appropriately when a match is found.
The preparation work for AI attribution claims is largely the same work that improves general rights administration. Clean catalog data, clear intake paths for novel claim types, member notification infrastructure, and interoperability with related licensing bodies: these are all improvements to existing operations that happen to also be the prerequisites for handling AI attribution at scale. PROs that frame the work that way will find it easier to justify the investment now, before the claims volume makes it unavoidable.
Stay Informed
