A licensing deal between a publishers' trade body and an AI music company lands as a headline, and the headline reads like a ceasefire. The lawyers who were suing each other last quarter are signing terms this quarter. For songwriters and publishers, the takeaway looks clean: the catalog is being paid for, the training is being licensed, the AI music business is finally coming inside the tent.
Read the fine print and a different shape emerges. The AI music licensing deals announced so far mostly settle one question — can a model be trained on your songs, and on what terms — while leaving the harder one open: when an AI-assisted track racks up streams, who gets paid, and out of whose pool. That second question is where streaming fraud lives, and it is not a footnote. It is the part of the contract that will decide whether these deals were a win or an advance against money that never arrives.
I score games and short films for a living, and I have watched the AI tooling go from a novelty to a fixture in my own sessions. So I read these announcements the way a producer reads a sync contract: not for the press-release verbs, but for the clause that governs the back end. Here is what most people take from the news, what the structure of the business actually suggests, and what I do when an AI partnership crosses my desk.
What most people do
Most people read "licensing pact" and hear "problem solved." The litigation track and the licensing track running in parallel get read as a contradiction — weren't these the same companies in court? — and then resolved as good news: the grown-ups worked it out, the catalog got valued, the checks are coming.
That reading is not stupid. It is what the announcement is engineered to produce. A trade body signs with an AI firm, a quote about creators and AI working together gets pulled into the headline, and the framing settles into "this is what fair looks like now." Publishers nod. Songwriters assume their administrators are handling it. Policy watchers file it under precedent and move on to the next docket.
The litigation-and-licensing pairing is deliberate, not contradictory. Sue the companies that take without asking; sign the ones that ask. That is a coherent strategy, and it is the right one. The problem is what the comfortable reading skips over. It treats the deal as a destination when it is closer to a permission slip — the part that says the model may learn from these songs, on these terms, for this much. It says far less, usually nothing public, about the side of the ledger that matters most over time: what happens to royalties when the model's outputs go to market and start competing for plays.
Most people stop reading at the part that feels like victory. The structure of the deal is hiding in the part nobody quotes.
What the evidence suggests
Split the economics into two sides and the gap becomes obvious.
The input side is training. An AI company wants to use a catalog of compositions to teach a model. A publisher controls those compositions. They negotiate a price, a scope, an opt-in or opt-out mechanism, maybe a duration. This is the side the announced deals are good at. It is a recognizable transaction — you have something, they want to use it, you set terms. It maps onto licensing instincts the industry already has.
The output side is what happens after the model ships. A user generates a track. That track gets uploaded to streaming platforms. It accrues plays. Somewhere in that chain, money moves — and the question of whether the publishers whose work trained the model see any of that money, on terms comparable to what a human-written song earns, is the one the announcements tend not to address. When a deal is silent on output parity, that silence is the most important term in it.
Why does the output side matter more than the input side over a five-year horizon? Because of how streaming royalties are actually paid.
How the pool actually works
Most major streaming services pay rights holders out of a pro-rata pool, not per individual stream at a fixed rate. The service takes its subscription and ad revenue for a period, sets aside its cut, and divides the remainder according to each track's share of total streams. If your song is 0.001% of all streams that month, you get roughly 0.001% of the payout pool. The per-stream "rate" everyone quotes is a result of that division, not an input to it — it floats based on how many streams there were in total.
That design has a structural vulnerability, and it has nothing to do with AI specifically. The pool is fixed; the supply of tracks is not. Add more tracks competing for plays and you do not grow the pool — you slice it thinner. Every new stream that goes to a new track is a fraction of a cent that did not go to an existing one. This is true for human uploads too; it is why the catalog has been getting more crowded for a decade. AI changes the math because it collapses the cost of producing a plausible, uploadable track toward zero.
Here is the part that turns a crowding problem into a fraud problem. If producing a track costs almost nothing, and a track can earn even a trickle from the pool, the incentive to manufacture tracks at industrial scale and farm streams against them — with bot networks, with stream-inflation services, with thousands of near-identical uploads — gets sharper. This is not hypothetical. Stream-manipulation cases have already produced criminal convictions in the United States involving artificially generated tracks and automated playback at scale. The mechanism is dull and effective: flood the supply, simulate the demand, siphon the pool.
So the warning that often closes these announcements — that bringing AI music to market could intensify streaming fraud — is not a vague gesture at risk. It is a description of the exact pressure the deals create on the side they did not negotiate. Licensing the input legitimizes a pipeline whose output competes for the same fixed pool that pays the people who did the licensing. If the output side is not addressed, the deal can be a genuine payment for training data and a net transfer of future royalties away from the catalog that got paid. Both can be true. They are not in tension.
What "output-side parity" would even mean
It is worth being precise, because the phrase gets used loosely. Output-side parity does not mean "AI tracks should earn the same as human tracks." It means the publishing rights embedded in an AI-assisted output — the composition share — should be recognized and paid on terms comparable to the recording side, and that the publishers whose work trained the model have a defined claim when those outputs generate revenue. Sound-recording royalties and publishing royalties have rarely been at parity in streaming to begin with; the worry is that AI outputs inherit that imbalance and scale it, with publishing once again the side that gets recognized last.
None of this is resolved by an input-side deal. It is a separate negotiation, and as of writing, the public terms of the headline pacts say little about it. When the announcement is silent, assume the question is still open — not that it was answered favorably and left out of the press release.
What we genuinely do not know
I want to flag the limits of this analysis honestly, because the temptation in this corner of the press is to project confidence the facts do not support.
- We do not know the actual royalty mechanics inside most of these deals. They are private. Public statements describe intent, not the schedule of who gets paid what.
- We do not know how platforms will classify AI-assisted tracks, or whether they will at all. Some have floated detection and disclosure requirements; enforcement and definitions are unsettled.
- We do not know how large the fraud effect is relative to organic crowding. Both compress per-stream payouts; separating them is hard, and the numbers that circulate are often estimates with wide error bars.
What we do know is the structure: fixed pool, near-zero marginal supply cost, input-side deals that do not bind the output side. That structure is enough to take the warning seriously without needing a precise figure attached to it.
What I actually do
When an AI partnership or platform crosses my desk — as a creator deciding what tools to build with, or as someone advising rights holders who ask — I stop reading the verbs and go straight to four things. None of this is legal advice; it is how I price the risk before I let a tool near a paying project.
1. I read the deal for the output side first. The input terms are usually clear and usually fine. I go looking for whatever governs revenue after a track is generated and distributed. If the partnership is silent on output economics, I treat the input payment as the entire deal and assume the back end is unprotected. Silence is a term.
2. I ask what fraud controls the platform actually runs. Not "do you take fraud seriously" — every platform says yes. The specific questions: Is there upload-rate limiting? Stream-pattern detection that demonetizes inflated plays rather than paying out and clawing back later? A disclosure flag for AI-assisted tracks that platforms can act on? A controlled set of clear answers tells me whether the company understands it is operating inside a shared pool, or whether it plans to externalize the cost onto everyone else in that pool.
3. I separate "commercially safe to use" from "fairly paid." For my own work — a game build that needs adaptive loops, a video edit due Friday — what I need is original output I can license cleanly with no sample-clearance landmines. That is the input-and-rights question, and tools that handle it well (City of Punk is built around exactly that: original, clearable audio with the license terms stated up front) solve my immediate problem. But "I can safely ship this track" is a different question from "the publishers whose catalogs trained this are being paid fairly downstream." A tool can pass the first test and the industry can still be failing the second. I keep those two assessments in separate columns.
4. I read the license like a sync contract, because it is one. Format, scope, term, territory, exclusivity, and the back end. Below is the comparison I actually fill out.
| Question | Input side (usually covered) | Output side (often not) |
|---|---|---|
| Is there a payment? | Yes — for training/catalog use | Frequently unstated |
| Who is paid? | The publisher/rights holder | Unclear once a track is generated |
| Out of what pool? | A negotiated fee | The shared pro-rata streaming pool |
| Does volume help or hurt me? | Neutral | More AI output thins the pool |
| Fraud exposure? | Low | High and largely unpriced |
| Parity with sound-recording side? | N/A | Rarely addressed |
If the right-hand column is mostly blank, the deal is an input-side deal wearing an industry-wide headline. That is not a reason to reject it. It is a reason to call it what it is and keep negotiating the part that was left out.
A note for the songwriter who is not in the room
If you are a writer whose administrator signs these deals on your behalf, the practical move is unglamorous: ask your publisher or PRO in writing how AI-assisted outputs are accounted for in your royalty statements, and whether any portion of an input-side licensing fee flows through to you. You may get a non-answer. The non-answer is information. It tells you the output side is unsettled for your catalog too, and that the place to apply pressure is not the AI company — it is your own chain of administration, before the precedent hardens.
The deals are real progress on a real problem. Training on creative work without permission was untenable, and getting it onto a licensed footing matters. I do not want to talk anyone out of taking the input-side money; refusing it does not protect the pool. I want the celebration to stop one clause short of where it usually stops, at the line that says what happens when the outputs go to market.
Because that is the line where the bill comes due. A fixed royalty pool can absorb a lot of crowding before anyone notices the per-stream number sliding. It absorbs it quietly, month over month, until the writers ask why their statements shrank in a year their plays held steady. By then the supply curve has already moved, and you cannot renegotiate a pool you have already diluted.
So here is the question I cannot resolve, and neither, honestly, can anyone quoting hard numbers at you yet: can a pro-rata streaming model — built for an era when making a release-quality track cost time, money, and a room full of gear — survive a supply of tracks whose marginal cost is approaching nothing? If it cannot, the input-side deals are not the settlement. They are the last clean payment before the pool stops being worth dividing.
Try it yourself, free
Generate your first royalty-free track in seconds. No card, no catch — type a prompt and hit render.
