Master and publishing cleared. Music, lyrics, and — through the Marquee Artist Program — name, voice and likeness.
The regulatory and legal position moved decisively in one direction over twelve months — and it isn't the direction that rewards waiting.
The ceiling on an EU fine for a general-purpose model provider — €15 million or 3 % of worldwide annual turnover, whichever is higher. Enforceable since 2 August 2026.
Penalty per further breach imposed on an AI music generator in Munich — alongside an order to stop offering the model in Germany and to publish the judgment in a national newspaper.
Licence agreements struck between rights holders and AI companies in twelve months. Three of them settled a lawsuit first — priced after the leverage had already moved.
Publicly reported developments, summarised for orientation and not as legal advice. Judgments cited are first-instance and under appeal.
A licensing decision inside an AI company is never one decision. It is three, taken by people who don't share a vocabulary — and each of them holds a veto.
“What is our exposure — including retroactively?”
An unbroken chain from consent to output. Master and publishing cleared in the same agreement, permissions logged and revocable per work, and a documented rights chain per asset that holds up under disclosure. Training holdings you already have can be cleared rather than hidden.
“How good is the data, and how long until it's in the pipeline?”
Identified audio with stems, ISRC and ISWC, resolved splits and descriptor depth — signal rather than volume. Delivered by API in the shape your ingest already expects, with the permission state on every asset live rather than snapshotted. Nothing about your architecture is needed to start.
“Can we defend this publicly — and win deals with it?”
Cleared provenance has stopped being reputational decoration and become a condition of market access: with regulators, distribution partners, enterprise buyers and investors. Built with collecting societies and industry bodies rather than around them. Verifiable in private, which is what makes it defensible in public.
One agreement satisfies all three. That is the whole point of doing it in one place.
Talk to us →Consent at the source, and proof that survives contact with counsel.
Not volume — signal. Identified audio with the metadata depth that makes a set usable.
Beside your pipeline, at the depth you choose. No forced access to weights.
Being licensed has stopped being a nicety. It's a condition of market access.
Aggregated across distributors, labels, publishers, production libraries, collecting societies and individual rights holders — permission granted per work, never assumed in bulk.
Included because the owner said yes — not because it was reachable.
Recording and composition cleared together, splits resolved per co-writer.
ISRC, ISWC, ownership shares, provenance — not a folder of untitled WAVs.
Separated vocals, drums, bass and harmony — higher signal per file.
Tonal, temporal, spectral, structural, semantic and lyrical — conditioning data.
Genres, eras and regions beyond the obvious — repertoire hard to clear elsewhere.
Its own licensable right, through the Marquee Artist Program — never quietly bundled.
Permissions change. You see the current state — a withdrawal is a signal, not a surprise.
The repertoire every dataset already has is the repertoire that teaches a model to sound like everything else. What actually moves a genre is harder to license — which is exactly why we go after it.
Breadth is a model-quality argument before it is anything else: a corpus assembled from whatever happened to be easiest to clear will sound like it. Coverage differs by segment and is confirmed per agreement — we don't claim territory-wide catalogues.
Name, voice and likeness belong to a person, not to a catalogue — and they are the one thing a model can imitate without copying a single bar. So they are licensed explicitly, through their own programme, or not at all.
A model does not need to copy a bar to sound like someone. That is why these three sit outside the recording licence and are granted on their own terms.
Availability is per artist and confirmed in the agreement. We do not licence a likeness we have not been asked to licence.
Voice and name are usually the first thing a model can do and the last thing anyone cleared.
Ask about the roster →No research project and no bespoke plumbing. Three modes on the same contract — and a permission state that is always the current one.
The licensed set delivered as audio plus metadata, refreshed on an agreed schedule. The simplest way to start, and the easiest to audit.
Query by descriptor, genre, era or territory and pull only what you actually want. Permissions are checked at request time, not at snapshot time.
Your ingest calls us per asset. New grants appear, withdrawals propagate, and the set you hold stays the set you are licensed for.
What lands in your system per asset. state changes before anything else does, so you always know exactly what you still hold a licence for.
Four things a general-purpose model provider has to be able to produce in Europe. None of them is hard once the training data is identified. All of them are close to impossible once it isn't.
| The obligation | Without a licensing layer | With AIxchange |
|---|---|---|
| A copyright policy that actually bindsAI Act Art. 53(1)(c) | A policy drafted by counsel, applied to a corpus nobody can inspect. Nothing connects the document to the data, so nothing can be verified — by you or by anyone else. | Permission is recorded per work at the moment it enters the set. The policy and the data say the same thing, and the record is what shows it. |
| Rights reservations identified and respectedArt. 53(1)(c) · Art. 4(3) DSM | robots.txt at web scale — which tells you nothing about an audio file that has already been copied a hundred times and stripped of its metadata. | Reservations resolved at ISRC and ISWC level, applied at ingest, and logged when applied. A reservation that arrives later is a signal, not a surprise. |
| A public summary of training contentArt. 53(1)(d) · AI Office template | An archaeology project across scraped sources, most of which have no owner attached — and a template that expects you to name one. | A query against an identified corpus: title, ISRC, ISWC, owner, share. Filling in the template becomes a report rather than a project. |
| Documented data sourcesArt. 53(1)(a) · Annex XI | A data-sourcing description assembled after the fact out of crawler logs, for material whose provenance was never captured in the first place. | Provenance recorded at ingest — provider, agreement, consent reference, per asset. The documentation is a by-product of the licence rather than a project of its own. |
Which works you licensed
On what terms
Weights, architecture
Your other agreements
Article references are to Regulation (EU) 2024/1689 and Directive (EU) 2019/790, summarised for orientation and not as legal advice.
Schematic comparison of deal architecture, not of named agreements. What any individual deal covers depends on which rights the counterparty actually controls.
The licence gives you the right to train. Creative Weight Attribution is how what you pay gets divided between the people who granted it — measured, deterministic, published with Fraunhofer IDMT. If your data team wants the mechanism, it lives on that page rather than this one.
Use case, repertoire needs, territories, and where the exposure sits today. Under NDA if you prefer.
Delivery mode, metadata schema, and how a change in permission reaches your pipeline.
A bounded run on a defined catalogue segment. Real records before you commit at scale.
Every layer in scope cleared, delivery in production, payout via existing royalty rails.
Start with the scoping call. Everything else follows from it.
Book a scoping call →No — clearing existing holdings has become a normal contractual subject. The obstacle is usually identification rather than willingness: you cannot license what you cannot name. We start from what is actually in your set, agree what is covered, and license it forward from there.
Compared with what? Set it against provisions for legal risk, re-licensing after a ruling at a price you no longer control, and enterprise partnerships that now ask for cleared provenance. And one agreement covering four rights layers is cheaper than four partial ones that still leave a gap.
Very little. A licence needs to know the use, the territories and the term — not your architecture, not your weights and not your existing training set. Delivery runs beside your pipeline rather than inside it, and the shallowest mode is a scheduled hand-off of files.
Then you hold some of the rights, for some of the repertoire. What direct deals rarely cover is publishing at the same depth as the master, lyrics as licensed text, and name and likeness at all. We work alongside agreements you negotiated yourself, for exactly the layers they leave open.
The asset's state changes first, and you see it — a withdrawal is a signal in your pipeline, not a letter from a lawyer. What it means for a model already trained is settled in the agreement rather than after the fact, which is the point of granting consent per work in the first place.
Which of their works you licensed, and on what terms. Not your weights, your architecture, your training set or your other agreements. Keeping those two things apart is a design requirement, not a courtesy.
How rights holders grant consent, set their own terms and get paid — the supply side of the same marketplace.
Tell us where you are — pre-training, in production, or working through a licensing question you have already been asked.
info@allrights-aixchange.com · Technical conversations under NDA on request
