AI Licensing And The EU AI Act: Practical Guidance For Music Lawyers
A wave of deals and new regulation is reshaping how AI trains on music. Recent licensing arrangements built around a "large music model" and the EU's cross-sector AI statute sketch a workable path: train on rights-cleared catalogs, disclose training sources, and compensate creators.
The creative debate over whether AI can capture "soul" will continue. The legal issues are concrete: copying for training, derivative outputs, and distribution at scale-plus transparency and consent. This is where counsel can create real leverage for artists and publishers.
The Copyright Stakes
AI music creation touches multiple exclusive rights and common theories of liability. Here are the pressure points you'll see in negotiations and disputes:
- Reproduction right: Training requires copying works into datasets, converting to machine-readable formats, and making intermediate copies. Rights holders argue this is infringement absent a defense (e.g., fair use). Even transient copies can be contested.
- Derivative works: Output that closely tracks a song's structure, cadence, melodies, or core organization may cross into unauthorized derivative territory if it's substantially similar.
- Distribution right: Platforms that let users generate and share AI audio raise direct and contributory exposure by enabling downstream public distribution.
A Move Toward Licensed Training
Klay Vision Inc. announced separate AI licensing deals with the three major labels and their publishing arms. The platform trains its large music model on licensed catalogs and positions itself as an interactive layer for listeners-while paying for the underlying creativity of artists and writers.
The EU AI Act adds legal muscle on transparency and copyright compliance for general-purpose AI. Providers must comply with EU copyright law, including the DSM Directive's opt-out for text and data mining (Article 4(3)). The Act also requires a "sufficiently detailed" public summary of training content following an EU template.
EU AI Act overview | Directive (EU) 2019/790 (DSM), Article 4
Bottom line: models trained on licensed, disclosed datasets are becoming the standard. These norms give counsel a template for future deals and a benchmark for compliance.
Open Questions That Affect Deals
- Authorship threshold: US law protects human authorship, not machine output. Where is the line between meaningful human contribution and technical assistance?
- Ownership and royalties: How do we split value among artists, featured performers, labels, publishers, platforms, and end users? Existing splits presume traditional roles; AI blurs them.
- Market competition: AI-associated tracks are charting and signing deals. Breaking Rust's "Walk My Walk" hit Billboard's Country Digital Song Sales Chart, and Xania Monet has both charted and signed a seven-figure deal.
- Royalty categories: Do we fit AI into current frameworks or add new ones for training, model use, and voice-model licensing?
- Consent scope: Can artists grant narrow, revocable permissions-by project, by work, or by timeframe-instead of blanket catalog access?
What To Know
- Deals like Klay's point AI companies toward licensed training through rights-cleared catalogs. Expect more pressure to license rather than scrape.
- Fair use defenses are still being litigated. Licensing provides predictability and speeds adoption.
Practical Steps For Counsel
- Audit existing contracts: Flag provisions on new tech, derivatives, text/data mining, machine learning training, voice likeness/right of publicity, sublicensing authority, and label/publisher "license on your behalf" clauses.
- Define AI licenses clearly: Specify scope (training vs. generation vs. voice modeling), fields of use, territories, term, revocability, and moral-rights-style restrictions (e.g., no political ads, no voice cloning).
- Set economic terms: Establish rates for training, per-use generation, and any voice-model or style-model fees. Tie share calculations to transparent reporting.
- Require transparency: Mandate audit-ready records of training inputs, opt-out management, model versions, and output attribution. Align with EU AI Act-style summaries of training sources.
- Labeling and attribution: Require clear disclosure of AI involvement, model identity, version, and any human contributors credited.
- Content controls: Prohibit cloning of protected voices or confusingly similar "sound-alikes" without consent. Include effective notice-and-takedown and repeat-infringer policies.
- Data governance: Limit retention of copyrighted works and intermediate copies to what's licensed. Require secure handling and deletion upon request or termination.
- Enforcement levers: Build in audit rights, step-in rights, and escalation paths for injunctive relief. Make breach consequences explicit.
Outlook
AI in music will keep expanding, but it doesn't have to sideline creators. Licensed training, real transparency, and clear consent terms can protect artist autonomy while giving models the inputs they need.
For legal teams building internal AI fluency to negotiate these deals, see this curated list of AI courses by job.
Your membership also unlocks:
