In an earlier article, I posed nine questions to music industry leadership about AI and music. The third question was: Can the industry move from litigation to deals?
Some artists may embrace authorized fan remixing, multilingual vocal adaptation, or avatar performance. Others will refuse. A trustworthy ecosystem must allow either decision to be enforced. The use of a person's voice, image, name, or likeness (VINL) in AI tools makes this challenge even more complicated. Permission to create one product or experience does not necessarily imply broader permissions, such as to train an AI language model, generate new performances, or enable other users to do so.
The early legal and commercial discussions around "AI music" were dominated by shock, demos, scraped datasets, deepfake voices, and litigation. Increasingly, it is now about deal architecture.
A pattern is beginning to emerge: first, license where possible; next, encourage legislation where gaps remain; and finally, litigate if markets cannot be built on consent. That sequence does not eliminate conflict, but it turns the central question from "Can AI do this?" into "On what terms should an AI system be allowed to do this?"
Different Uses Require Different Permissions
There are different permissions, operating at different layers and points in the ecosystem's architecture:
- Training permissions are concerned with what content may be used by another party to develop and grow an AI model, whether that be entire catalogs, recordings, compositions, lyrics, stems, session files, voices, images, or metadata
- Output permissions are concerned with what a user of an AI-based system may do with any AI-generated material: stream it, download it, synchronize it, sell it, remix it, train further on it, or keep it inside a bounded environment
- Prompting rights and guardrails are concerned with whether and how users of an AI system may request content from a living artist, a sound-alike, a protected title, a catalog-specific style, or a recognizable voice.
Each layer may require different levels of permission, disclosure, monitoring, and compensation. It also assumes that the questions are asked and adequately answered by those with authority and competence to do so.
Artist Consent
Consent is—or should be—the baseline. The U.S. Copyright Office's report on digital replicas concluded that existing legal protections do not fully address the harms posed by unauthorized digital replicas. The report recommended a federal minimum protection, including guardrails for licensing and informed consent.[1]
In music, this issue is particularly acute because voice is so intimately bound up with identity, brand, history, and emotional recognition.
Disputes over AI training are more complex because they implicate catalogs at scale. The Copyright Office's 2025 report on generative AI training emphasizes that copyrighted works used in training are not merely "data"; they contain creative expression protected by copyright.[2] The report also recognized that voluntary licensing markets for AI training are already developing, particularly in sectors such as music, while acknowledging that licensing may not be equally workable in every context. For now, it has recommended allowing those markets to continue developing rather than imposing a blanket statutory solution.
The industry needs workable market mechanisms, but those mechanisms must not erase control or bargaining power.
Licensing for Different Uses
"Walled gardens" are one possible mechanism. Fan tools and artist-branded experiences may work best inside a controlled environment where output cannot be exported, monetized, or misrepresented without review by the license holders.
However, many professional uses require downloadable assets, sync delivery, stems, editability, and commercial warranties. The next rights market may need tiered permissions—therefore closed fan play, personal export, professional production, commercial sync, derivative works, and model improvement should not all be bundled together.
The EU AI Act also encourages ecosystem transparency by requiring providers of general-purpose AI models to maintain technical documentation, implement a copyright policy, and publish a summary of training content.[3] Even where legal obligations differ by territory, the direction is clear: black-box rights practices will not sustain trust. Music companies, AI developers, and platforms should assume that creators will increasingly demand to know whether their work was used, under what license, and with what downstream limitations.
From Rights Debate to Rights Infrastructure
For the Music Technology Coalition, this area calls for a practical "rights lab." The Coalition could provide a neutral environment to compare components of new contract structures (for example, what types of metadata are required), develop plain-language explainers for artists, test disclosure workflows, convene discussions on voice consent, and model how permissions information should travel from creative origin, through training data and processing tools, to generated output. It could also help students and founders distinguish between legal permission, ethical legitimacy, and product trust. A license may answer one question, but not all three. To address all three requires an ecosystem-wide trust framework and possibly a trust mark.
Deals will decide much of the future of AI music. The task is not to slow innovation but instead to ensure the market being built around AI starts with informed consent, preserves choice, and distributes value to the people whose work makes the technology meaningful.
The move from litigation to licensing will become consequential if the deals themselves are worthy of trust.
[1] U.S. Copyright Office, Copyright and Artificial Intelligence, Part 1: Digital Replicas.
[2] U.S. Copyright Office, Copyright and Artificial Intelligence, Part 3: Generative AI Training (Pre-Publication Version).
[3] European Commission, General-purpose AI obligations under the AI Act.