A new generation of attribution registries, dataset licenses, and royalty flows is turning scraped training data into a paid relationship, and the working musicians testing them are shaping who gets a cut.
For a working session guitarist in Nashville, the question of how AI uses their sound has moved from abstract to administrative. After fifteen years of reading royalty statements, the same musician is now learning a new vocabulary: attribution registries, training-data licenses, and micro-royalty line items that did not exist three years ago. The IEEE Spectrum analysis on generative AI music attribution surveys a field in flux, where the plumbing of getting paid when AI learns a song is being built in real time, and the people closest to the music are trying to write the rules.
The mechanisms on the table are not one thing. They are a layered stack that pulls in different stages of the training and output process. Some proposals attach a royalty to the training run itself, treating every dataset as a paid input. Others attach payment to the model's output, splitting revenue when a generated track is streamed. A third family builds opt-in registries, where musicians can list catalogs they will license to AI labs and exclude those they will not. Collective licensing, modeled on the deals collecting societies already strike with radio and streaming services, is being adapted for training data. Companies like Sureel are building attribution formats that could let models disclose what they were trained on, and the Sureel/STIM partnership is testing what a licensing relationship between an attribution registry and a collecting society could look like.
What "getting paid" actually means varies by mechanism. A per-training-run payment rewards the catalog at the moment the model is built and ignores downstream use. A per-output payment, by contrast, can compound with the model's success but requires tracking that has not existed before. Upfront buyouts give musicians certainty but cap their upside. Micro-royalties attempt to thread the needle, paying small amounts to many rights-holders based on measurable influence. Each model has different politics. Per-output schemes favor catalog owners whose work becomes a hit inside the model. Per-training-run schemes favor established catalogs with leverage at the licensing table. Opt-in registries tilt the field toward musicians with the time and legal fluency to negotiate.
The infrastructure underneath is the story. STIM, the Swedish copyright agency, which for decades has paid songwriters when their songs were played on radio and streaming services, is extending that role to AI. Companies are drafting attribution formats. AI labs, facing lawsuits and public pressure, are signing their first data deals. The IEEE Spectrum piece documents this emerging architecture and frames the field as one that is "rethinking royalties" rather than settling them.
For working musicians and small rights-holders, the question is not whether any of this will happen. The question is which mechanisms will reach them. Opt-in registries risk excluding the musicians who most need the income, because registering requires time, legal fluency, and an organization behind you. Collective deals promise broader reach but require the collecting society to actually sign the lab. Training-data taxes, floated in some legislatures, would pool money and redistribute it, but the redistribution formulas are themselves a battleground. The most likely outcome is a stack, not a single answer, with musicians collecting from several sources for the same piece of work.
The copyright lawsuits that have dominated the generative AI music landscape are beginning to give way to an increasing number of privately negotiated agreements between major labels and AI companies. Whatever mechanisms emerge in the next few years will be shaped by how those deals develop and by the outcomes of the remaining legal fights. The IEEE Spectrum analysis treats the legal fights as one input to a market that is being designed under uncertainty, with companies, collecting societies, and labs all drafting in parallel.
What to watch next: the first announced training-data deal between a major lab and a collecting society, which would set a template. The first opt-in registry that reaches a critical mass of independent musicians. The first attribution standard a major model voluntarily adopts in its output disclosures. The first royalty statement a working musician actually receives from an AI training event rather than a streaming play. None of these will settle the field. Each will be a data point in the negotiation over who gets paid when a machine learns a song, and on whose terms.