Music publishers set out seven AI licensing principles
IMPF and IMPEL want clearer royalties and stronger recognition of songs
Independent music publisher groups IMPF and IMPEL issued seven licensing principles on September 30. This is an industry proposal, not legislation or a signed AI deal.
Why songs and recordings have separate value
The distinction matters because a composition and a recording are separate creative works. The US Copyright Office explains that a composition includes the music and accompanying lyrics, while a sound recording fixes a particular performance. Owning copyright in one does not substitute for ownership in the other. Its guidance gives the example of a songwriter whose composition is recorded by somebody else. The songwriter or publisher can claim the composition while the performer or label claims the recording.
What the framework asks for
1. Recognise the song’s value in every licensing model.
2. Value songs at least equally with recordings unless other relevant factors justify a different allocation.
3. Distinguish payments for earlier use, training, outputs and future exploitation.
4. Make deductions, costs and revenue calculations transparent, justified and fair.
5. Specify licence boundaries without unintentionally authorising wider future uses.
6. Agree attribution and valuation safeguards with rights holders.
7. Ensure downstream AI music uses yield appropriate royalties for songs used in training.
The groups say three attribution pilots are examining tracking across training, generation and outputs, and reporting for remuneration.
A controlled trial provides earlier context
On July 29, the organisations announced a sandbox with Sureel AI to study how its attribution technology traces musical works. Participating publishers would test selected repertoire in a closed environment, examine detection and measurement, and explore whether the resulting data could support compensation. The announcement restricted use to the trial’s defined scope and said works would not be made available for AI training outside it.
That announcement describes a research exercise. It does not provide a published accuracy benchmark or establish that attribution technology has solved the payment problem.
What would make the evidence useful
Our view is that a useful demonstration would let a publisher follow one known composition from a test input to an explanation of any proposed payment. A second test should deliberately use an unrelated work to reveal false matches. Publish failure cases alongside successful examples. Without that comparison, a convincing dashboard can hide uncertainty about what was actually measured. Commercial negotiations still need a decision about who carries the cost when the system is wrong.
Archival piano photograph titled Keys of music by Panda8pie2, taken in April 2016 and licensed under Creative Commons Attribution 4.0. Resized and converted to WebP. The image illustrates music making.
