Something changed in the licensing conversations at ASCAP and BMI over the past year. AI developers are approaching performing rights organizations not to license catalog for public performance, but to license it for training. The ask is different in kind: not "we want to play this music," but "we want to use this music to teach a system to generate music."
The performing rights organizations are being asked a question that their existing license structures were not built to answer. How they respond will shape the rights framework for AI-generated music for years.
The Structural Problem with Existing PRO Licenses
ASCAP and BMI issue licenses under the concept of public performance. The legal theory is that when music is performed publicly, the copyright holder has the exclusive right to authorize that performance, and PROs administer blanket licenses that allow licensees to perform the entire PRO repertoire in exchange for a fee. The fee is based on usage: how many performances, in what contexts, reaching how many listeners.
AI training does not fit this framework at any point in the description. Training use is not a public performance. The recordings are processed computationally during model training; they are not performed for an audience. The output of training is not the original recording; it is a mathematical model that has abstracted statistical patterns from many inputs. The usage metric of "performances reaching listeners" has no analog in the training context.
This is not a gap that can be bridged by interpreting an existing performance license expansively. It requires a different type of license, probably closer to a synchronization or mechanical license in structure, but covering a use case that neither sync nor mechanical was designed for. The legal category of "training data license" is being invented in real time by the parties who need it.
Consent at the Individual Work Level
The training consent problem has a specific technical dimension that distinguishes it from other licensing questions. A blanket performance license covers all uses within a defined territory and usage type, and the individual songwriter's consent is aggregated through their PRO membership agreement. The consent to licensing is structural, not work-by-work.
Training licenses present a harder problem because the question of consent is meaningful at the level of individual recordings. A songwriter who did not consent to their work being used to train an AI model that will compete directly with human-composed music is raising a different kind of objection than a songwriter who objects to a specific venue's blanket license rate. The objection is about the nature of the use, not the price of the license.
Whether individual songwriter consent should be required for AI training licensing, as opposed to aggregated consent through PRO membership, is a contested question. The major PROs have taken the position that their existing blanket license frameworks can accommodate AI training with appropriate modifications. Many songwriters and composer advocacy groups disagree, arguing that training licenses require affirmative individual consent and cannot be bundled into blanket agreements signed decades ago for different use cases.
This disagreement is not resolved. It is actively litigated and negotiated. The outcome will determine whether AI companies can access music catalogs for training through PRO blanket licenses or whether they need to negotiate with individual rights holders.
What the AI Companies Are Offering
The training license proposals that AI developers are bringing to PROs have a consistent structure. The AI company pays a lump sum or per-track fee for access to a defined catalog, and in exchange receives a license to use those recordings in training datasets. The license is typically time-limited to a specific training run or version of the model.
The fundamental challenge in these negotiations is that neither party has a clear basis for valuing the license. For public performance licensing, a decades-long history of rate setting, licensing precedent, and royalty distribution data provides anchoring for negotiations. For training licensing, there is no comparable history. The rate is being set from scratch, without clear understanding of what the training use is worth to the AI developer or what it costs the rights holder in terms of competitive displacement.
There is also the provenance problem: even if a PRO licenses a catalog for training, verifying that the AI company used only the licensed recordings, and not additional unlicensed catalog, requires some form of technical verification. Most current training license proposals rely on the AI company's self-reporting. Rights holders are increasingly skeptical that self-reporting is adequate given the scale and opacity of typical training dataset construction.
The Consent Question and Retroactive Training
There is a harder version of the consent question that licensing negotiations tend to avoid: what about models that were already trained, before any licensing framework existed, on scraped music datasets that included PRO-administered catalog without consent or payment?
Most major AI music platforms that launched before 2024 were trained on datasets assembled by scraping publicly accessible recordings. Some of those platforms have since entered licensing agreements for ongoing training. But the original training run, the one that produced the model's foundational capabilities, happened without authorization for most of the catalog it incorporated.
Retroactive consent is not a meaningful concept in copyright. Either the use was licensed at the time or it was not. The practical consequence is that rights holders who believe their recordings were included in unauthorized training datasets may have claims that are independent of whatever forward-looking licensing framework is negotiated. Those claims rest on establishing what was in the training dataset, which is why provenance documentation is not purely a forward-looking concern.
Where This Leaves PROs Today
ASCAP and BMI are navigating a situation where the existing license structures do not fit the use case, individual rights holders have conflicting views on consent requirements, the rate-setting basis is unclear, and historical training practices create potential retroactive liability that forward-looking licenses cannot extinguish.
We are not arguing that AI training licensing is impossible or that existing PROs are unable to adapt. The track record of the music industry's adaptation to digital streaming, while imperfect and slow, demonstrates that licensing frameworks can evolve to accommodate new use cases. The question is how long that evolution takes and how much value flows to rights holders during the period of uncertainty.
What is clear is that the consent question will be answered through a combination of litigation, legislative action, and negotiated framework agreements, probably in that order and over a period of years. The organizations that are most prepared, with clean catalog data, documented opt-out filings, and provenance records of which recordings appeared in which datasets, will be better positioned than those that wait for the framework to settle before preparing their documentation.
The licensing question and the technical attribution question are connected. A clear picture of which recordings contributed to AI model training, at the recording level, is both the evidentiary foundation for rights claims and the data layer that makes a functioning training data license framework possible.
Stay Informed
