For years the argument was whether a machine could make something that sounded like music. In 2026 the argument is whether the machine learned what music sounds like by listening to yours — and who gets paid when it answers.

Major catalogs have moved from blanket denial to structured licensing experiments: limited training windows, opt-out registries, "clean-room" stems derived under contract. Independent artists are split between teams that use AI as a compositional sketchpad and teams that treat every diffusion step as an existential threat to session work.

The feedback loop

The uncomfortable symmetry is that AI doesn't just influence music once. Models tuned on streaming behaviors reshape recommendation; recommendation reshapes what gets recorded; those recordings become training fodder for the next generation of generators. Music influences what the systems hear; the systems influence what gets heard next.

Distribution executives describe the effect bluntly: catalogue that performs well in embedding space gets surfaced more often — which means certain harmonic and structural fingerprints propagate faster than others. Whether you call that artistic convergence or homogenisation depends on which side of the royalty statement you're on.

"We're not scared of the synth patch. We're scared of the synth patch trained on our masters without a clause." — Rights counsel, mid-tier label group

Where the money is moving

Enterprise-facing deals — synchronisation libraries, production suites for broadcast, automated mastering chains — are ahead of consumer-facing regulation. That's where the measurable unit economics live: per-seat tooling, capped inference, audit trails that prove which training corpus was not touched.

The open question for the next eighteen months is whether public-market music multiples start discounting catalogues that haven't clarified their AI posture. Early signals from institutional buyers suggest they might.