A ground for the underground

On performance rights, the machines that learned without asking, and why paying for a sample is now the most strategic thing an artist can do

Jordan KatzCEO & Founder, bushido29 August 202611 min read

Tunnel I. The estate nobody can see

There is a second music industry sitting underneath the one you can see, and it is larger than the one you can.

The visible industry is streams, tickets, merch, playlists, charts. It has a scoreboard. Anyone can read it. Underneath it sits the performance rights ecosystem: the flow of money generated every time a composition is performed, broadcast, streamed, synced, covered, sampled, remixed, or played to a room of people who have no idea a ledger is being written. That flow is administered by collecting societies, publishers, sub-publishers, administrators, and the reciprocal agreements between them. It moves billions annually across the world. And it is functionally invisible, including — critically — to most of the people it is supposed to pay.

The invisibility is not a scandal. It is an architecture. Performance rights were built for a world of paper cue sheets, radio logs, and national monopolies, and that world had a workable assumption: that the number of performances was small enough to count, and the number of works was small enough to name. Both assumptions died. What survived is a system asked to reconcile hundreds of billions of usage events per year against a work registry that cannot reliably agree on what a work is.

The result is the black box. Royalties collected, attributable to no one, distributed by market share to the parties who already have the most market share. It is the closest thing modern capitalism has to an unclaimed estate — and like all unclaimed estates, it gets quietly divided among the people standing nearest the door.

Tunnel II. The matching problem is the whole problem

Here is the mechanical truth that almost every conversation about artist compensation skips.

A recording has an ISRC. A composition has an ISWC. A songwriter has a share percentage, expressed in the registration held by their society or publisher. Money can only reach a writer if all three resolve to each other: this sound → this work → these people → these percentages.

That resolution fails constantly. It fails because ISRCs are issued by whoever distributes, not by whoever owns. It fails because ISWCs are assigned late, duplicated, or never assigned at all. It fails because splits get agreed in a text message and registered eighteen months later, or never. It fails because two societies hold conflicting share data for the same work and neither has authority over the other. It fails because a live performance, a cover, a remix, an interpolation, and a sample are five different legal events that the data layer treats as one ambiguous blob.

Every failure is money that exists but cannot find its owner.

This is why the serious work in this sector is not marketing. It is normalisation and enrichment: reconciling society registries, DSP catalogues, and direct rightsholder feeds into a single resolvable identity graph, then treating rightsholder-supplied data as authoritative rather than inferring ownership from whoever shouted first. Matching ISRC to ISWC to writer share, across sources that disagree, is the unglamorous engineering that decides whether songwriters eat.

bushido's detection layer exists for exactly this reason: to find the works whose money is moving to the wrong account and force the correction. That is not an aspiration. It is a reconciliation problem with a defined answer.

Tunnel III. The artists performing other people's songs

Now overlay the creative reality.

A vast proportion of what audiences consume is one person performing, reworking, quoting, or building on another person's composition. Covers. Interpolations. Samples. Live sets. Flips. Edits. Beat packs. The modern producer's default working method is derivative in the precise legal sense of the word, and there is nothing shameful about that — Western music has always been built by inheritance.

The problem is that inheritance without registration is theft with extra steps.

When a producer samples a record and releases it unlicensed, they do not merely expose themselves to a claim. They inject an unresolvable object into the identity graph. A new recording exists, generating streams and performance income, with no relationship recorded to the underlying work it derives from. The original songwriter's share does not flow, because there is no path for it to flow along. Nobody stole a specific dollar. Something worse happened: a dollar was created that no system can route.

Multiply that by an entire generation of bedroom production and you have manufactured a permanent black box contribution. Not a leak. A structural intake valve.

The Machiavellian reading is not that unlicensed sampling is immoral. It is that unlicensed sampling is strategically incompetent. It converts a creative asset into an uninsurable liability, permanently, in exchange for saving a clearance fee. Every unlicensed release is a bet that no one will ever run audio fingerprint detection across the catalogue you are in. Fingerprinting is cheap now. That bet is already lost; the losers simply haven't been served yet.

Tunnel IV. The machines that trained without asking

Which brings us to the war.

On 31 July 2026, the Munich Regional Court decided the first European case against a generative music tool. The court ruled largely in favour of the German collecting society GEMA against Suno, prohibiting four distinct acts across six compositions: reproduction for training purposes in the United States, reproduction through memorisation inside the model in Germany, communication to the public by offering the model, and reproduction and communication to the public through its outputs. The court asserted jurisdiction over the US training conduct via a venue rule for collecting societies, applied US copyright law to those acts, and rejected fair use. Each future violation carries a penalty of up to €250,000. The court also found that the works had been extracted from YouTube by stream-ripping, circumventing the rolling cipher protection measure.

Read that list again, because the significance is not the win. It is the itemisation. The court did not treat "AI training" as one undifferentiated act. It broke the pipeline into discrete copyright events and ruled on each. That is the end of the industry's favourite rhetorical shield, the claim that a model is a single transformative process rather than a chain of reproductions.

Three weeks later, on 17 August 2026, the independent publisher Round Hill Music filed in the Northern District of California. It sued Suno and Anthropic, alleging both scraped hundreds of its songs to train models without permission, and attached 500 compositions as an initial exhibit with a stated plan to amend to ten thousand or more. Damages could exceed $1 billion per case, and Round Hill has said it will not settle. The Suno complaint also names the data-scraping vendor Bright Data as a contributory infringer. Round Hill's CEO framed the position with unusual clarity: licensing is not an obstacle to innovation, it exists to protect the owners of the raw material.

The strategic detail everyone is missing: every major AI copyright case so far has settled before producing binding precedent, which is why the central fair use question remains unresolved. Round Hill has chosen a jury and an attorney who won the Blurred Lines verdict, which signals an intention to produce precedent rather than convert pressure into a licensing cheque.

That is the war in one sentence. The majors monetised the threat. An independent decided to detonate it.

In their wordsJason RysCEO, Wixen Music Publishing

A word on AI
The unlicensed use of musical works to train and power AI systems remains a top enforcement priority for us. Our position is straightforward yet universally disrespected: songs are not free raw material like coal or ore, something to be freely dug out of the ground. Companies building AI products on the backs of your work and creativity need licenses. Full stop. The same as any other user of music.
We are actively monitoring AI platforms for unauthorized uses of our clients' music, pursuing takedowns and enforcement where we find them, and evaluating more formal legal enforcement. We'd prefer an opt-in (critically: not opt-out) licensing framework with reasonable terms but will not stand idly by and let them steal your music in the meantime.
State of the Industry
The NMPA reported that U.S. music publishing revenue reached $7.3 billion in 2025, up 3.7%. Growth of any sort is good, of course, but it's slowing across the industry compared to prior years. Why? Well, most immediately, look at bundling.
Here's the short version of a complex framework about why bundling matters. Streaming royalties are set by a formula with several backstops: songwriters and publishers are paid the greatest of a percentage of the service's revenue (the “headline rate” aka Prong 1, the one everyone negotiates over and is in the news from time to time), a percentage of what the service pays record labels (this is called the Total Cost of Content or “TCC” prong), or a per-subscriber minimum (“PSM”).
When a streaming service bundles music into a package with audiobooks or other content (like Spotify and Amazon did), it allocates only a slice of the subscription price to music. This shrinks the revenue pool before the headline rate is ever applied. So, despite these royalty “increases” over the last couple of years, the bundling effect has led to the NMPA estimating a loss of $480 million to songwriters and publishers since 2024.

US publishing revenue, 2025 (NMPA): $7.3bn
Year-on-year growth — and slowing: +3.7%
NMPA estimate of the bundling effect since 2024: −$480m

Tunnel V. What the courts are actually telling the industry

Strip the moral language away and the message from both benches is procedural, not ethical.

Provenance is now a legal asset. Munich did not ask whether Suno's outputs were beautiful or transformative. It asked what was copied, where, and under what authority. The defence collapsed because there was no record. There was never going to be a record, because the working assumption of the entire scraping era was that the record would never be demanded.

Licensing is not a cost centre. It is an evidentiary position. A cleared sample is not a receipt. It is a chain of title that survives contact with a court, an audit, a catalogue sale, and a distribution partner's compliance review. The unlicensed version is a contingent liability that compounds silently and detonates at the worst possible moment — usually the moment the work finally becomes valuable.

The registry is the battlefield. Every one of these disputes ultimately resolves to the same question the black box asks: who owns this, in what proportion, and can you prove it? AI did not create that question. It industrialised it.

Tunnel VI. The other way to build

bushido's position is deliberately unfashionable: do the clearance first, and make the registration automatic.

The licensing exchange exists so that an artist who wants to sample, interpolate, or rework another writer's composition can transact for it directly, at a price that reflects reality rather than a major label's opening posture. But the transaction is the smaller half of the design. The larger half is what happens in the seconds after the grant.

At the moment a licence is issued, the derivative work is registered. A new ISRC is assigned to the new recording. The relationship to the underlying composition is recorded. The share percentages of every contributing writer — original and new — are captured and pushed into the registration path, either through bushido acting as administrator for the licensee's share, or as an export package delivered to the licensee's own publisher or admin. Clearance status is exposed publicly and read-only per ISRC, so any downstream party — a label, a distributor, a sync agency, a DSP — can verify the chain without asking anyone's permission.

The effect is that the derivative work enters the world already resolvable. It is not an orphan. It cannot become black box income, because the routing table was written before the first stream. The original songwriter gets paid not out of goodwill but out of plumbing.

This is the Machiavellian case for doing the right thing, and it is not a moral case at all.

An artist who clears their samples and registers their splits is not being virtuous. They are being defensive with unusual foresight. They are making their catalogue auditable, saleable, syncable, and litigation-proof at the exact moment the entire industry is being forced to prove provenance under oath. They are choosing to be the party holding the documents when the documents start mattering.

Everyone else is holding a bet that no one will ever check.

Tunnel VII. The ending nobody in the room wants to say out loud

The AI cases will resolve into licensing regimes. They always do. Some will settle, one or two will produce precedent, and within a few years the training of models on commercial music will be a paid, metered, contractual activity — because that is the only equilibrium that lets both sides keep operating.

When that happens, the money will need to flow somewhere. It will flow through the same identity graph that already fails to route ordinary performance income: ISRC to ISWC to writer to share. If that graph is still broken — if splits are still unregistered, derivatives still unlinked, societies still disagreeing — then the largest new royalty pool in fifty years will pour directly into the black box and be distributed by market share to the parties standing nearest the door.

The war with AI is not really about whether machines may learn from music. That fight is being won. It is about whether the settlement money can find the songwriters when it arrives.

Registration is not paperwork. It is the difference between a victory and a payment.

Clear the sample. Register the split. Link the derivative. Do it now, while it is cheap and voluntary, rather than later, when it is expensive and compelled.

Jordan Katz is CEO and Founder of Bushido Way, Inc. bushido operates a licensing exchange for sample, interpolation, and derivative-work clearance, with automatic derivative registration and public clearance status at the point of licence grant.

Sources

The Munich judgment (Landgericht München I, case 42 O 763/25, 31 July 2026) is subject to appeal to the Munich Court of Appeal, and damages have not yet been quantified.