Analysis
Publishers want the song valued at least as highly as the recording
IMPF and IMPEL have published seven principles for future AI licences. A central idea is that the song itself — lyrics and composition — should normally be valued at least as highly as the finished recording. This is a negotiating demand from independent music publishers, not an existing rule.

A song and a recording involve different rights. When AI is trained on music or used to create new works, publishers’ organisations want songwriters’ contributions reflected in payment and reporting.
What happened?
On 30 September, IMPF and IMPEL published a joint statement on how they believe generative AI should be licensed. They represent independent music publishers and their song rights. They want clear, transparent and economically sustainable licences.
The most prominent principle is that the song itself — lyrics and composition — should be valued at least as highly as the recording unless particular circumstances justify a different allocation. That is their negotiating position. The statement does not itself create a new law, a fixed percentage or an agreement already accepted by every AI company.
A song and a recording are not the same thing
Think of a song as two layers. The composition is the foundation: melody, lyrics and musical elements. The master is a particular recorded version of that song, with specific musicians, singers, sounds and production.
Different people and companies may hold rights in the two layers. A songwriter may own the composition or receive payments through a publisher. A record label may own the master. An independent artist may both write the song and pay for the recording. When an AI service uses music for training or offers song creation, several groups may therefore want to know what was used and how value should be shared.
What do the seven principles seek to change?
In brief, the organisations do not want agreements to lump together every type of AI use or every revenue stream. They distinguish past use, model training, AI output and future commercial use. They also want costs, deductions and calculations reported clearly.
- The song should have its own visible value in the licensing model.
- Payments for past use, training and new AI output should be reported separately.
- The licence should specify exactly which works and uses it covers.
- Rightsholders should be able to scrutinise how works are recognised, credited and valued.
- If AI music is later used commercially, compositions used to train the model should be able to receive appropriate compensation.
Why does it matter?
If a licence pays only for the finished recording, songwriters and publishers may feel their share of the value is missing. Conversely, a new model for compensating every work used in training faces a difficult question: how can anyone know which songs actually contributed to a particular result?
Without clear reporting, a songwriter cannot easily check whether compensation is reasonable. But technology that attempts to trace influence is not the same as proof that a particular song was copied. IMPF and IMPEL say they are testing attribution technology with three companies; they describe ongoing pilot projects, not a finished system.
What does it mean for hobby musicians?
If you write and record your own music, you may in practice be both the composer and owner of the master. Keep simple records of who wrote the lyrics and melody, who recorded it, which samples or loops were used and which licences apply.
- Record co-creators and their shares before release.
- Keep licences for samples, AI tools and recorded voices.
- Ask whether your distributor or publisher handles the master, composition or both.
- Check whether an AI licence covers training, generated results, commercial use or several of these.
A closer look: a framework for negotiation
IMPF and IMPEL list seven principles: recognise the composition’s value; balance the allocation between masters and publishing; distinguish payments for past use, training, output and future uses; make deductions and calculations transparent; define the scope of the licence; agree methods of recognising and valuing works with rightsholders; and enable commercial use of AI-generated music to compensate the songs used to train models.
There is a difference between a copyright rule and a negotiating demand. The organisations can argue for fairness and seek to include their principles in agreements, but the statement does not force a model developer or platform to accept their valuation. An actual agreement still needs to specify the catalogues covered, the models and features licensed, how money is calculated, which deductions apply and how reports can be audited.
“At least as much” is a valuation principle, not a ready-made calculation of every payment. Compositions and masters may contribute differently to different products. An AI melody generator may use compositions differently from a service making new sonic versions of a recording. The framework says differences should be assessed on their merits and the agreement’s scope must be clear.
The statement mentions three ongoing pilots exploring technology that can identify, track and attribute musical works through training, generation and output. Such tracking could support reporting, but it is not yet described as a complete solution to who inspired what. That is where the technical question meets the financial negotiation.
A numerical example: the percentage is only half the question
Imagine a purely hypothetical licensing pool of 100 kronor. An equal split gives 50 to the composition side and 50 to the master side if the whole amount is divided. If 20 kronor is deducted first, each side gets 40 instead. The split remains equal, but the money reaching the rights holders is lower. This is an illustration, not an IMPF tariff or an existing AI agreement.
A percentage therefore needs a denominator: a share of which revenue, after which deductions and over which period? Allocation within each group is another step. A particular share for publishers does not tell us what an individual songwriter ultimately receives.
Analysis: use and influence are different things
A training-material list can show that a work was included in a dataset. It does not automatically tell us how much a particular generated result depends on that work. If payment is to depend on that influence, it needs both an agreed definition and an auditable method.
One alternative to discuss is allocation from a shared pool under agreed rules. That could be simpler administratively but raises other questions: which catalogues count, how smaller participants are treated and who audits the records. This is our analysis of a possible approach, not a system the organisations are reported here to have adopted.
What a songwriter needs to be able to trace
Follow the chain from the work and ownership share to the licence scope, report and payment. If a link is unclear, it becomes difficult to assess the offer even when its headline promises a fair split. Concrete agreements and auditable reports are the decisive next steps. A statement of principles starts that discussion; it does not complete it.
Think about it: how should value be shared?
If an AI song draws on patterns from thousands of compositions but no individual song can be identified, should songwriters still receive money from a collective licence? If an AI work later becomes a hit, should compensation go to a recording, a composition, the training catalogue or all three? Which model would you call fair?
GO TO THE SOURCE
Sources & background
Expanded 13 October 2026. The reporting is followed by editorial analysis and clearly labelled examples. The original publication date is retained.
- IMPF and IMPEL: full joint statement (PDF)↗www.impforum.org
- IMPF: summary of the framework↗www.impforum.org

