The Black Box Is a Choice
The short version
Anyone who claims they can look at an AI output and tell you what went into it, without ever seeing the training set, is wrong. Nobody outside the company has seen that set. I'm not going to tell you otherwise.
There is a better place to look. When a system reaches for specific licensed works while it is making something, that can be written down. Echora is built on that moment. It works like a Google search. Everything that comes back in the search is cleared, the terms come with it, and the use gets logged.
Once that record exists, rights holders can see which of their works were used, how the attribution was calculated and how the money gets split, and they can check the statement themselves. Every one of those has an answer you can build.
The deals being announced right now pay the catalog, in one check to the company that owns it. They don't say which works got used in which output, even though this business has always paid per use, by the play or the rerun. Paying the catalog is a start. It isn't an accounting.
The black box is a choice. So is opening it. I have spent four years building the way in.
The call
On my screen was a list of licensed songs. An AI workflow had pulled them out of a catalog and used them as references for a new piece of music. Next to each song you could see who owned it and what share of the result had been assigned to it.
I was on a call with two people from a generative music company, showing them Echora running live. Somebody had typed a prompt into our sandbox, and this was the ledger of what went into the output.
One of them asked whether all of this was unnecessary complexity. Then he made the bigger objection. Attribution, he said, is an unsolved problem in mechanistic interpretability.
He was right, and it had nothing to do with what was on the screen. He meant tracing an output back through everything a model absorbed in training: millions of songs, broken into weights and blended together.
Nobody outside the company even knows which songs were in that training set. Give me the set and the outputs and I can line each recording up against each output, element by element, with times on both sides, so a person can listen and check. Without the set, you are arguing about a list nobody has seen.
The most rigorous test anyone has written down is to remove one song and retrain the whole model to see what changes. The papers call it leave-one-out, and whole families of shortcuts exist because nobody can afford to run it at scale.1 It gets worse: these models don't give the same output twice from the same prompt, so you would have to run that test many times over to tell a real change from noise. If training is the only place you look, you are stuck.
I was showing him the other place to look: which songs the system had retrieved for that one generation. His company's own model was running inside the workflow I had connected, and every work that didn't match its owner's terms had been filtered out before the model ever saw it.
That keeps the liability off the AI company and keeps the decision with the people who own the work. I have never understood why that is a hard sell.
Think about how a chatbot like ChatGPT handles the news. It wasn't trained on this morning's articles. It runs a search, picks a few, reads them and cites the ones it used. That is a search engine workflow, and nobody calls it an unsolved problem.
I have had some version of this conversation more times than I can count. Because nobody can fully explain what happens inside a model, and the people who build them say so themselves,2 people act as though nothing around the model can be accounted for either.
It can. What a system deliberately reaches for can be written down. After that the questions are ordinary: what did the owners allow, and how do they get paid?
Rights holders get there fast. In those meetings nobody brings up interpretability. They ask which works were used, how the shares were calculated, how labels and publishers will split the money, and whether they can check the statement. Hard questions, and the right ones.
What stayed with me about that call is that the AI company was already out chasing blanket licensing deals. You would think it would want to know which works its users keep asking for, and which references its system keeps choosing, before it writes the next check.
Here is what worries me about those checks. Pay once for a catalog, keep no record of what got used, then train the next model on what the first one made. Do that a few times and the company doesn't need the humans anymore. That is how we seal our fate. There are two ways to fight something you are told is inevitable. One is to refuse it. The other is to look the people saying it in the face and tell them how it is going to run. Echora is built for the second one. If this is happening, and it is, it happens with consent and a record, and the people who made the work stay in it for good. Those are the people I built Echora to protect.
I got into music publishing eleven years ago because I wanted to know how a song got from a studio to the radio, or to a stage with thousands of people singing it back.
At Pulse Music Group, part of my job was reporting ownership splits to the performing rights organizations. A song could have several writers, each with a publisher, and one publisher might administer another's share. If I reported the numbers wrong, somebody didn't get paid right, and untangling it took months. So I was obsessive about it. I cross-referenced everything, twice, and I never sent a wrong report.
That job gave me my idea of what a creative business owes people. Saying the catalog made money is where the accounting starts. Then you account for each work, and for each person with a claim on it.
I paid for Echora's early years by working on music videos and commercials, and I mean that literally. Every dollar I earned on set went back into the company, and then some; I have gone into debt for it.
Earlier this year I was art directing a video for one of the biggest artists in the world, a fourteen-hour day, and that same morning I had a meeting with a music company about why attribution matters.
So I left home early, pulled over near the set, hotspotted my laptop off my phone, turned on a virtual background so nobody could tell I was in my car, and propped my phone's flashlight against the rearview mirror to light my face. I took the meeting from the driver's seat, then went and worked the fourteen hours. Four years of explaining this have looked a lot like that.
On those jobs I watched teams use AI to rough out ideas and then keep it far away from the finished piece. They were worried about clearance, about liability and about how the work would be received. They wanted the tools. Nobody wanted to be the one holding the copyright risk when the spot aired.
And on a music video set, the question is never only about music. Look at a single shot. There's a performance, a song under it, a photograph on the wall with someone's face in it, a scarf a sponsoring brand sent over.
All of it belongs to somebody, and each piece arrived with its own permissions. Music is where I learned to follow that trail, but the trail runs through every medium.
I want creators to have a better choice than refusing AI outright or signing terms that leave them blind to what happens to their work and what it is used to make. I want the people building tools to be able to license real human work, the detailed, complicated kind that someone poured years and love into, without negotiating a custom catalog deal before they have a product.
Most of the companies building these tools are startups without the capital for a blanket deal, and a blanket deal makes them pay for everything, including the works their product never touches. Paying per use is cheaper for them and fairer to everyone else.
Echora sits between the two. Think of it as a Google search for your AI model. Everything that comes back is cleared, the terms come with it, and the use gets logged. Anything that doesn't fit what the rights holder allowed, or what the buyer agreed to pay, never comes back at all.
Today we can show that in music, images and video, and we are building the licensing and reporting relationships that turn a logged use into a payment to the right people. Most of what we earn is a share of the money we help move to rights holders, the way a distributor earns. We also get paid by companies that want their own model built from material they control. None of it comes from a creator's work being used without a payment to that creator. I set it up that way on purpose.
Meanwhile the industry is making its decisions. On September 9, 2026, Suno introduced v6 and said the models were developed with partners including Warner Music Group, BMG and Believe, with opt-in experiences around individual artists still to come.3
Billboard reported that Warner licensed a subset of its catalog for training, that the partners are paid through a revenue share, and that it is still unclear how the companies decide which music is in.3 So the label gets a share of Suno's revenue.
What nobody has explained is how that share gets divided among the songs and the people who wrote and recorded them. Whether a label can license a recording for this comes from the contract behind it, and how much of the money reaches the artist comes from the statement.
What I actually want to know isn't in any of these announcements. When a specific piece of work helps make something valuable, does the person who made it keep sharing in that value? Do they even find out?
Two different questions
Training is about what went into a model. Live reference is about what a system reaches for while it is making one particular thing.
A 2025 paper by Fabio Morreale and colleagues, "Attribution-by-design," draws the same line between training-time and inference-time attribution and argues for recording the works supplied to condition an output, so payment can follow those uses.4
A work can shape a model in training without anyone ever being able to say how much it contributed to a given output. That is the part nobody can see, and a lot of models never fetch outside material at all when they generate.
If that is the only way AI touches creative work, then creative work becomes a training pool with a one-time check attached, and everything after that is a blind spot.
Nobody knows what got used. Nobody knows what it was used for. Nobody gets paid twice. That is the black box, and left alone it swallows the entertainment business one catalog at a time.
The alternative is to run it like a search engine. When a system reaches for identifiable works while it is generating, there is a moment where permission can be checked and the use written down. That moment is the only place anyone can see AI using creative work live, so that is where we built.
In the demo, the system had a catalog of licensed references. The prompt led it to choose certain works. Because the choosing happened inside our workflow, we kept the record.
ChatGPT's web search does the everyday version of this when it retrieves pages and cites them.5 We are not asking AI models to behave in ways they don't already behave today.
A list of what was retrieved is the start. Attribution is the part that turns it into money. Our method establishes that a reference was used and assigns it the share that royalties get calculated from. It says plainly what it measures, and rights holders get evidence to confirm it was applied as written.
The only thing we leave out is what nobody outside the lab can see: whatever the model carried in from a training set years ago. We have no window into that, and we don't pretend to. Everything the system touched to make this output, we account for, and we lock it the first time.
None of this lets anyone off the hook for training. A creator can fight over how a model was built and license new reference uses at the same time, because permission for tomorrow's use covers tomorrow's use and nothing earlier.
But I have sat in too many meetings that spent the full hour on how a model learned in the past, and never reached the terms on which it could use a licensed song next week.
We can't reconstruct everything a model learned years ago. We can write down what it used this afternoon, and once that record exists, it is worth something to everybody who touched the work.
Where the money stops
The thing I find hardest to accept is how little information the music business, and the entertainment business around it, is willing to settle for here, when it wants data about everything else that touches the bottom line. Data decides who gets signed and where the marketing money goes. I don't always love that.
I still think A&Rs should be able to hear a song, feel something and back it before the numbers agree. But a business that runs on performance data ought to be demanding evidence of AI use, and right now it is mostly shrugging.
If a service is searching a catalog for references all day, why would you not want to know how often your work comes up? A creator could find demand they never knew about. An AI lab could find out that the next license worth pursuing sits outside the catalogs it already knows.
Anyone from publishing has been waiting for me to say that we already have a black box. We do. It is what the business calls money that came in for songs nobody could match to an owner. The Mechanical Licensing Collective has to hold that money for at least three years.
After that it can pay it out by market share, which means it goes to the songs that did get matched, in proportion to what they already earned. The MLC says its first payout of that kind starts in April 2027, with what is left unmatched from January 2021.6
The writers of the unmatched songs aren't in that payout. Nobody knew who they were. I'm using the words for something newer in this essay, the part of an AI system nobody can see into. Hold on to the old meaning, though. It shows where money goes when nobody can say whose it is.
The Suno deals pay the label a share of revenue. The question is what happens next, because a revenue share still has to be split across the songs. If it gets split by each company's existing market share, the formula pays a company for the position it already holds, whether or not its songs were what AI users were asking for.
And the majors spend enormous money keeping that position, which is their job. If an AI model starts picking a kid from Missouri they never signed over the superstar they did, that is a problem for them, and I understand why.
Which is one reason the black box suits a lot of people fine, on both sides of the table. Look at how these deals pay: a large check up front, perhaps a piece of the AI company's equity. All of it lands in this year's numbers. Whoever closed it, at the label and at the AI company, gets the credit this year, and probably the bonus.
What the deal does to the catalog ten years out is somebody else's job by then. That incentive would pull on anybody. And the black box lets a label say "we can't tell whether this was your song," which means none of that money has to be credited against any artist's recoupable advance.
Plenty of labels will pay their artists regardless. Nothing in a black box requires it. That is what a black box is for.
Now consider the kid from Missouri. Playing local dive bars, a few hundred listeners, no marketing budget, and one song that fits a particular request better than anything in a superstar's catalog. That artist isn't even in the market-share deal.
Nobody negotiated on their behalf. If an AI system chooses that song, the person who wrote it should be paid for the choice, and under a market-share split no one would ever learn the choice was made.
AI isn't automatically blind to fame. What a system was fed, the metadata it gets and how it is designed all tilt what it selects. What we are building is a search where the fit of a licensed work counts for itself, so a reference can be worth something without being a hit first.
That kid also needs a way into that search, which means new work registering under clear terms and showing up next to everything else. Nobody buys a higher spot in the results. If we ever sell placement, it will be labeled and kept apart from the search, the way a search engine keeps ads apart from results, and it will never change which work the system picks on merit.
I do wonder how the biggest catalog owners would feel about results that don't match their market share. The old formula is predictable. A record of what actually got used might show the value sitting somewhere else. Everyone should get to see that record before agreeing to freeze the old distribution in place.
Photography got here first, and I'll give the stock agencies credit for paying contributors at all. Shutterstock pays its contributors a cut of data-licensing money based on how much of theirs sits in the dataset, and pays again when an AI image is downloaded.7 Getty said its generator would share revenue by each contributor's slice of the training set and how that content performs.8 Those are training formulas, and they do what they were built to do. Neither can answer the question I care about: which image did the system actually reach for this time?
Say one contributor has 220,000 images in a reference catalog and another has 200, and for a certain kind of request the system keeps choosing from the 200. Pay by volume and you would never know.
The same question follows the money inside the building. I'm not anti-label. I entered this industry through music publishing. Labels and publishers are rights holders, and they are who we build for.
But when people tell me "the rights holders were paid," I ask which rights, which people and under what agreement. A payment to the catalog is not an account to the creator.
And for most signed artists, the first thing any payment meets is their unrecouped advance. The label fronted money, the artist's own royalties pay it back, and until that balance hits zero the artist sees nothing new. Whether a catalog-level AI check even gets credited against that balance depends on the contract, and most of those contracts were written before anyone imagined the question.
Musicians who played on a record can be owed money when that recording is reused for something new, under American Federation of Musicians agreements.9 SoundExchange splits U.S. digital performance royalties 45 percent to featured artists, 50 percent to the owner of the recording and 5 percent to a fund for session players and vocalists.10
Each of those is its own pool of money with its own paperwork. In June 2026 the AFM sued over the licensing of its members' recordings to Suno and Udio, arguing that its recording agreement already requires payment and notice for those uses.
Universal, Warner Records and Atlantic dispute that reading, and no court has ruled.11 I don't know who wins. I do know both sides are arguing about the same recordings, and that "the catalog was licensed" didn't settle what the players are owed.
It matters on the artist's own statement as well. If AI income comes in at the catalog level, the artist needs to see whether any of it was allocated to their work.
If their contract applies it against an unrecouped advance, the statement should show that credit. From where the artist sits, the deal isn't done when the label gets paid.
Then there is the buyout. A complete buyout fixes the price on the day of signing. If the use turns out to be worth far more than anyone guessed, the creator has no share in the difference. I understand why people take them.
Imagine if Billie Eilish had handed her whole catalog to Spotify for one flat fee: no play counts, no statement, no share when a song took off. Nobody would call that fair, and it is exactly what a buyout of AI rights looks like. It has happened before, too. In 1938 the two men who created Superman sold every right to him for $130, and spent the rest of their lives watching him earn fortunes they never shared in.12
A check now can be what keeps the lights on. But a deal that looks good this quarter may govern uses for years, and whoever is offering it should have to say what future participation is being given up.
Music already has a better answer, and I worked inside it: the advance. A writer or an artist gets money up front, it recoups against what the work earns, and royalties keep coming after that.
Book deals work the same way, and so do most of the entertainment deals I have ever seen.13 A check at signing can start a relationship. It doesn't have to end one.
I'm not saying everyone needs an advance. Some rights holders will want nothing up front and a payment every time their work is used, and some developers can only afford that. An advance and a per-use payment both keep the relationship open. A buyout closes it.
One more comfort I hear from rights holders about training deals is that the source files have to be returned or deleted after some number of years. Fine, but what happens to the model? Deleting a folder doesn't remove what the model learned from it. Google's researchers made that point when they launched their machine unlearning challenge.14
A contract has to say which uses of the trained model continue, and whether any reporting or payment continues with them. Live reference at least gives everyone a place to decide the next use. Are these works available, for this, on these terms?
We already know how to do this
Anyone in publishing lives off cue sheets. When a song goes into a film or a TV show, someone logs every piece of music in the final cut: the title, the composer, the publisher, how many seconds it ran, what kind of use it was.
BMI calls the cue sheet the official record of music in a production, and performance royalties get calculated from it.15 Get a duration wrong and a composer's check is wrong. It is unglamorous, and it is how a lot of a publisher's film and TV money starts.
Nobody filling one out needs a theory of everything that inspired the film. They need to know what's in it. A finished recording dropped into a scene is simpler than several references shaping a generation, I know. But the habit is the one we need. Define the use, keep the evidence, pay from the record.
Film and television have also been through the economics of reuse. The Directors Guild's history of residuals starts in live radio, where performers did a show for the East Coast and then did it again three hours later for the West.
Once the networks could record the first performance and replay it, the second one wasn't needed, and the guilds negotiated payment for the replays.16 The work only had to happen once to be valuable twice, so the pay followed the use.
Residuals don't depend on whether the project turned a profit. They are triggered when the work is shown again, under formulas agreed in advance. That is the precedent I care about for AI. The first fee doesn't answer every later question, and the agreement says what does.
Performers gave us the third piece, which is consent. The 2023 SAG-AFTRA TV/Theatrical contract required clear, separate consent for a digital replica, along with a reasonably specific description of the intended use. A different use needs new consent, subject to the contract's exceptions. It covers voice as well as likeness, and the 2026 successor agreement builds on it.17
What I like is that the performer never has to decide whether "AI" is good or bad. They are asked about one use. They can say yes to this project and no to that one. And owning the footage doesn't give a production the performer's permission for a new digital portrayal of them.
Every other medium I've worked around asks the same four things. What was used? Who could authorize it? What use was allowed? What payment follows?
Commercials have a real clearance system: clearance reports, releases, licenses, an errors-and-omissions policy before anything airs. The paperwork is thorough. It just lives with the production company and its lawyers, not with the file. Hand that file to a new team a year later and what travels with it is the word "cleared" and nothing else. That's fine until the client calls back.
They want the spot to run another year and they want a new tagline. Now someone has to find the voice actor's contract, because the script changed. The photograph turns out to have been licensed for six months.
Nobody can find the email. Every one of those is a paid use to the person who supplied the work, and the commercials contracts already say so.18 The rules were never the problem. The record didn't travel with the file.
Products are where I like to start with people outside music. A brand launches a collection and puts approved digital versions of its bags and scarves into creative tools, so an agency can drop the real scarf into a campaign, or so the brand can invite its own fans to put the real scarf into what they make with AI.
The brand has one instruction: use this exact scarf, and don't change it. Keep the pattern, keep the shape, don't blend it with fifty others.
That is a different kind of license from "here's some material, go remix it." A system that only knows the word "licensed" can't tell the two apart, and that is the whole problem. Echora stores the no-alteration term with the asset and enforces it in the workflow. The scarf stays the scarf, and the result gets checked before release.
Adobe published a case study in January 2026 on LG Household & Healthcare working this way, building exact product twins in Substance 3D and generating only the backgrounds around them.19 That is the right instinct. What we add is the permission and the payment attached to the twin.
Games stretch it further, because a game can generate new material while someone is playing. An actor who approved recorded lines hasn't approved a system that speaks lines nobody reviewed.
Steam asks developers to disclose AI content made during development separately from content generated live, and SAG-AFTRA's 2025 interactive agreement treats real-time generation as its own use, with its own consent and pay.20 A first yes can't stand in for every later request.
Now put all of that into one thirty-second spot. A voice actor reads the script. A photographer's image is on the wall behind the product. The product is the scarf. A licensed song plays underneath.
A year later the client wants a fifteen-second radio cutdown with a new tagline the voice actor never read. In Echora, each of those four pieces carries its own terms, set in advance by the person who owns it. The owner already answered, and the answer travels with the work.
If the actor allowed new lines, the read is generated under their terms and they are paid for it. If they didn't, the system uses a voice whose owner did, and the actor's terms are honored without a phone call. Same for the photograph, the song and the scarf.
Without that, the team is starting over from an unlabeled video and calling around to find out who said yes to what. Every one of these industries worked out long ago that you keep the record. AI doesn't get excused from that because it's new.
It stays in their hands
Consent comes first, before anyone talks about money. It is remarkable that consent is still a conversation in 2026. When an artist's record leaks, no one calls that consensual. When a model uses it without permission, are we expected to shrug?
People tell me that if work is on the internet it is public, as if public meant public domain. Your car is in public every time you drive it. It doesn't belong to whoever sees it.
Consent also can't be a single yes or no across an entire catalog. Can my song be used to generate new stems, or new songs? Can it sit under a video? Can my video be used to make an image?
Is my likeness available? My voice? A rights holder might allow live reference and refuse training. A performer may accept one use of their voice and reject another. A photograph can be available for editorial use and off limits for promoting a brand.
Echora lets you answer at that level of detail, down to the individual stems of a song, and change your answers whenever you choose. In training, your work is fixed the day the model is built. In live reference, your terms stay live. That is real control, and it belongs to the person who made the work.
A publisher with a big catalog can't negotiate every use of every song by hand, so we built the bulk terms editor. An authorized rights holder sets permitted uses, pricing and restrictions across as many works as they choose, and carves out exceptions where a song, a co-writer or an older agreement needs different treatment.
Bulk doesn't mean unrestricted, and nobody can use the editor to grant rights they don't hold. A label or publisher can also hand these decisions to the artist, inside whatever rights each of them controls.
We also have negative tags, which are restrictions on association. You can exclude politics as a category, or one politician by name. You can approve some advertisers and refuse a product you don't want near your work. A tag only works inside Echora. It can't reach into a model we never touch, and I'm not going to pretend it can.
Those terms have to change what the AI can reach. A restriction that lives in a dashboard while the work stays selectable is decoration. Inside the licensed workflow, the terms decide whether a reference is available before it is used.
And before a rights holder changes their terms, they can see what stops, what continues and which commitments survive under licenses already granted. Nobody should learn that on the day they try to pull out.
Some people will still say no. Good. They should be able to, the same way you're not obligated to sell on Amazon just because you sell on Etsy or at the local flea market. I'm arguing for a market people choose to enter.
And one thing should be said plainly: registering your work never gives Echora or anyone else permission to train on it. That power stays with you. Think of us as your shield, and when you want one, your weapon. Nothing is used, and nothing is trained on, without your yes. Training happens only if you choose to build your own official model, and it uses only what you approve.
Then there is identification, which my publishing years made me stubborn about. Knowing which work was used doesn't tell you who to pay. There are two sets of percentages here and people mix them up constantly. An attribution share is what the method assigns to a licensed reference for one output, and royalties are calculated from it.
An ownership split is how the rights in the original work are divided among its owners. You need both. The attribution share says how much flows to the song, and the splits say where it goes from there. The recording and the composition are handled separately, and permission on one side never covers the other.
When rights holders pointed out the holes in ownership data, they were right. An attribution system can't quietly convert a shaky ownership claim into permission. So a work doesn't become available through Echora until all the relevant co-owners, or their authorized representatives, agree on the ownership information and on the permission.
If ownership is contested, the money for that work is held until it's settled; it doesn't go to whoever claimed it first. We can't fix every ownership database in music.
What we can do is refuse to treat incomplete data as complete, because the first co-owner to show up later with a claim would drag both the creator and the buyer into a dispute, and our numbers would be worth nothing.
People ask me what happens when the splits on a song don't match, or add up to more than 100 percent. It happens for three reasons, and two of them are as old as co-writing. One is plain error.
Somebody keyed a number wrong, or the same share got entered twice, and the total on file comes to 110. The second is disagreement. The writers never settled who gets what, and the label put the record out anyway. The third, frankly, is that registering a song is confusing, and a lot of writers never finish the form right. We can't settle the first two, and I won't pretend software can. The third is the one we can help with.
What we can do is say the discrepancy exists, tell a buyer what the owners we represent are prepared to offer, and put the problem in front of every writer and publisher on the song, with a date on it.
Now price. I'm tired of the assumption that a technology company gets to decide what everyone else's work is worth. The rights holder sets the terms for what they control. The buyer can take them or walk.
A buyer can say "I can't afford that" and get a version built without that work; the system swaps in a cleared reference that fits the budget, and the expensive one never comes back.
Works don't all need one price, and there is no universal fee for pressing Generate. What I insist on is that the compensation is agreed, the use is logged and the payment can be explained. Nobody gets to quietly pay less than the rights holder agreed to.
Continuing participation is optional on both sides. A rights holder can offer a reference fee and a commercial-use fee and stop there, the way a stock photo license works.
Another can require a share of whatever the output earns out in the world, the way an actor with a piece of the back end keeps earning when the film does. The buyer says which obligations they will accept, and the search only returns offers that match. It works as a filter, so the bargain exists before the use does.
Why would anyone accept ongoing payments? Because the work they want is offered on those terms and they can't get it another way. A license doesn't guarantee a better result. It does mean the creator names the price at which their work is available.
It stays in their hands. It is an open market for creative work, and the people who made the work set the prices.
The other half of that freedom is that the record is final. The references, the permissions, the method and the shares that governed a generation are locked when it happens. Locking the record doesn't lock your terms. Change them tomorrow and tomorrow's uses follow the new terms. Yesterday's record stays as it was. Ownership data is separate.
If a split on file turns out to be wrong, or the writers finally settle one, the fix changes who gets paid out of a song's share and leaves the share itself alone.
You set your terms knowing that. If the output becomes a hit, the shares don't get renegotiated after the fact. Nobody gets to say "now that it's big, I want more," or "now that it's big, you should get less," not the rights holder and not the person who typed the prompt. The split was set before the money existed.
A later price change doesn't rescore yesterday's use, and neither does a newer version of our method. Those were the terms you had at the moment it was generated, and we gave you a lot of power to set them.
So a year goes by and an artist calls. Their song was in a hit and they want more. I open the ledger with them. Here is what the model used. Here is the share. Here are the terms you set that morning.
The song doing well is exactly why you set terms in the first place. What we sell is an allocation that gets locked the first time, and that anyone entitled can check.
Two kinds of money follow, and both are already wired into the record. If a licensed song is used as a reference again next month, that is a new use under its terms, and it gets logged the same way.
If an output goes on to earn from a release or a placement, the license can give the owners of the referenced works a continuing share.
The agreement spells out which events count, what income is covered and what the share is, because a percentage of receipts before deductions is a very different offer from a percentage of profit after them. A percentage means nothing until the contract says a percentage of what.
To make that second kind payable, every output leaves us with an identifier tied to its ledger, and a watermark that carries it, so that when usage and revenue reports come in from wherever the output went, they match back to the references and terms recorded at generation.
The ledger lives with us, separate from the file, so the file isn't the only place the history survives. People argue about dilution, the idea that after enough generations nobody could possibly trace what came from where. That is true only if you start late. Put the record in from day one and nothing dilutes.
Get a system like this in place early enough and every AI work that references human work carries a watermark that leads back to the ledger of what went into it. If that output gets referenced again, the ledger follows. The human work keeps living inside whatever gets built on top of it.
I think of the result as a split sheet for the output: who participates, on what basis, for which uses, clear enough that a rights holder can tell a payment is theirs without reconstructing the generation. It makes no sense to do the hard work of finding and paying people at one step and then lose them at the next.
The developer on that call had a fair point about complexity. Unreliable rights data, slow searches and unpredictable bills would make his product worse. So the offer to developers has to be one they can afford to repeat.
A developer can start by paying for the uses its product actually needs, without buying broad catalog rights up front, and if the product grows, the rights holders' income grows with it. The developer doesn't take on a giant fixed commitment before it has revenue, and the rights holder gets paid every time something is actually used.
Anyone who thinks ongoing royalties to an IP supplier can't coexist with building a technology company should read Arm's 2023 prospectus. Arm collects license fees plus a royalty on chips that use its designs.
Some products were still generating royalties more than twenty-five years after they were developed, and about 46 percent of royalty revenue for the year ending March 2023 came from products released between 1990 and 2012.21
A chip is easier to count than a creative contribution, and the comparison doesn't set a price for music. It does show that technology companies already know how to live with IP that keeps earning.
Administration is the harder objection. No developer can stop every generation to negotiate with several owners. The Mechanical Licensing Collective shows that broad access and work-level accounting can coexist. U.S. digital services operate under a blanket mechanical license, report usage monthly, and the MLC matches that usage to individual musical works for payment.22
One agreement, many uses, and the songs don't disappear. We borrow the approach. Agree the terms. Define what triggers a charge, including previews, failed requests and extra takes. Log the events and settle them together. A thousand one-cent obligations are ten dollars, not a thousand bank transfers.
And it saves compute. We compress the understanding of a media file, something we call the barcode method, so a compatible system doesn't have to analyze the same song or the same footage from scratch every time it is used as a reference.23
People like saving money. If using authorized work is cheaper and easier than the alternative, developers will choose it because it's good for the product, and the creator gets paid on the way through.
Who checks the math
Suppose Universal, Sony and Warner each built their own AI attribution model. Every one of them would be accused of favoring its own catalog, and it wouldn't matter whether the accusation was true, because nobody outside could check it, and nobody outside would believe it.
Now add the model developers running theirs. You would have several systems, each with a different idea of what a percentage means, and none of them comparable. Even an honest disagreement would be impossible to resolve.
Other industries settled this a long time ago. Public companies don't audit themselves. There is an independence framework for their auditors, overseen by the PCAOB. The scale at the grocery store and the pump at the gas station are tested by weights-and-measures officials, so that neither the buyer nor the seller has to take the other's word.24
You need a referee who doesn't play for either team. That is what Echora is built to be. We don't own the catalogs. The rights holders set the terms.
Our job is to apply those terms and the attribution method the same way every time, keep the evidence and answer for the result, including when the result is inconvenient for us. The standard is the same for a catalog of ten songs and a catalog of ten million. A famous name doesn't get a different calculation. Fame still counts where it has always counted: a star with a huge audience gets asked for by name far more often, and every one of those asks is a use that pays. It just doesn't buy a higher rank when nobody asked for you. The work has to earn that on its own.
The best standard anyone ever gave me is that a songwriter should be able to understand why they earned a particular amount. So a statement names the work, the use and the terms that applied.
It shows the attribution share and the ownership splits separately, shows how they produced the payment, and lists any deductions. A manager can follow the arithmetic without reverse-engineering it from the total at the bottom of the page.
And there has to be a way to test it. "Auditable" means the records are kept, an authorized auditor can get to them, and they are detailed enough to confirm that the agreed terms and the recorded method were the ones applied.
An audit checks the agreement. It doesn't renegotiate it. Privacy and other parties' confidential information deserve protection, and that protection can't become a blanket reason to withhold the evidence behind someone's own payment.
Technology like this can stay proprietary while its results face real scrutiny. The audit rights, the deadlines for a response and the process for a dispute all belong in the contract. Auditable has to mean more than the right to ask. It has to mean a statement a person can read and follow to the last cent.
We also check what the model didn't admit to. After an output exists, we compare it against the works the search surfaced that the generator never declared it used.
If the output leans on one of them past a threshold, it goes to a person for review before anything is paid. A system that only counts what it was told is easy to fool, and I don't want a black box hiding inside the thing built to open one.
The same standard applies to what we don't know yet. Ownership data in music is full of holes, and some of the money in any system depends on reporting that arrives late or not at all. A statement that hides that is lying with a clean layout.
So when a split is contested, the statement says contested, and it says since when. When a payment is calculated but waiting on a report from somewhere else, it says so, and it says what it is waiting on.
A rights holder should be able to look at the page and know the difference between money that is theirs and money that might be. I have read too many statements in this business that looked certain and were not.
The date matters most for a share nobody has claimed. At the MLC a share in a formal dispute is held until the writers settle it, however long that takes. An unclaimed share doesn't get that. It sits for a few years and then gets paid out by market share.6
Claim it in time and the writer gets paid. Miss the window and they don't. A dated flag that keeps showing up in front of every writer and publisher on the song is a clock they can read.
The bet
Payment is the obvious reason to keep these records. I have another one. I want people to know what happens to their work once it's out in the world.
I got into music because I wanted to know where songs came from. Who wrote that line, and what happened in their life that brought it to them that day? Credits, liner notes, interviews and conversations with collaborators let me follow a record backward after it came out, and knowing who was behind it always made the record better.
A ledger of licensed references does some of that in a new setting. A musician finds out their recording was useful for something they would never have thought to make. A listener finds the earlier recording, where the permissions allow the connection to be shared.
A publisher sees demand its performance data missed completely. I will only ever claim what the record shows. It can't tell you what a model soaked up in training years ago. Nobody can.
It can tell you which licensed works were in the room, how much of each one shows up in the finished piece, and whether the output leans on something it was offered and never declared. That is more than anybody is being told right now.
Go back to the kid from Missouri. Say that dive-bar recording gets chosen for a much bigger project, on terms that include a share of what it earns. When that money is reported, the artist opens a statement and sees which project paid them and why the payment is theirs.
I'm not promising anybody fame. I'm saying a good piece of work should be able to reach a buyer far outside the artist's audience without the artist vanishing from the account. The alternative is telling them that their slice of some large pool reflects what they contributed, and asking them to trust it.
That is the flat-fee catalog again, with a different logo on the check.
The plan at a lot of AI companies, as far as I can tell, is to sign a handful of catalog deals, extract what they need from human work in a single transaction, and train the next generation of models on synthetic material derived from it.
If they can pay you once for everything they need, they never have to pay you again, and you will have no way to trace what comes next back to your work.
Every executive who says "it's too hard to track, let's treat it as a black box" is helping build exactly that. An AI company's job is to build the best model at the lowest cost.
That is a legitimate job, and I work with people doing it, but it is not the same job as protecting your catalog, and no contract should pretend it is. Hand a model builder a black box and it will outgrow its need for you before you notice it happening.
My bet runs the other way. Independent artists, people making things nobody asked for: that is what keeps AI output from turning generic.
Human work stays valuable to these systems for as long as the systems want to be any good, and the only way to keep being paid for it is to keep it visible every time it is used.
I do worry about AI competing with the people whose work made it useful. Paying them doesn't make that go away. A generated track can still beat a human one to a placement, and a performer can still decide that no amount of money makes a use acceptable.
So what does paying change? If the work is going to be in the room either way, the only question left is whether the person who made it has a say and a share. With a say, they can refuse. With a share, every success built on their work pays them too. Without either, they're just watching. With both, they can weigh it like any other deal. That beats being told the future is coming and a check of some size will arrive at some point.
People sometimes ask me why copyrighted work even matters to AI. Couldn't the companies just train on public-domain material and machine-generated data? Some of the best-funded companies in the world, with some of the best lawyers, trained on copyrighted work knowing they would probably be sued for it, and they were.
OpenAI told a House of Lords committee that it would be "impossible to train today's leading AI models without using copyrighted materials," and that public-domain material alone wouldn't produce systems that meet people's needs.25 OpenAI maintains the training was lawful, and the courts will decide.
My point is different. You don't take that kind of risk for something worthless. They knew what human work had in it, and they knew their products needed it. The people building these systems already settled that question with their own behavior.
The law is catching up to them now, and it always does. We spent these four years building the system the law can run on when it gets there. I pitched it to the Copyright Office in 2024, and it is a lot further along now.
If the work is that valuable, the person who made it shouldn't disappear from the economics.
That is the bet Echora makes. Artists will keep making things other people want to hear, see and build with, and we only earn when that value reaches them. Judge us on the unglamorous parts: the contracts, the integrations, the statements, the payments.
Rights holders have to decide which uses they will authorize and what participation they expect. Developers have to keep evidence of the licensed references their systems actually use. The agreements have to connect that evidence to reporting and payment. Auditors need enough access to check it.
I have watched people put years of their lives into this work. Right now we are building the most sophisticated creative technology anyone has ever made and keeping its books in a shoebox.
That is not a technical limitation. Any company signing these deals could ask what was used and where the money went. The tools to answer exist. The asking is a choice.
Anyone who has poured themselves into something they made deserves control over how it is used and represented. I don't only mean a painting or a song. If you put everything you had into a slide presentation, this applies to you too. Everything else in this essay is what consent looks like once you take it seriously.
Keeping the record is the only way I can see AI actually benefiting the people who make things, instead of competing with them for the same spots. It lets the AI companies build on the best material there is with consent instead of a lawsuit.
Both sides do better. And it is how every new technology has ever been accepted by the people whose work went into it: they were asked, they were told what happened, they were credited, and they were paid.
I have spent four years and everything I have building the other option, and it works. Songs, images, voices, faces, footage, products: the uses can be logged and the terms can be held.
Getting the money to the person who earned it takes licensing and reporting relationships, and we are building those now. So when someone tells you it can't be done, understand what they are actually telling you. They would rather not.
Sources and notes
These sources support the passages indicated. Company accounts and historical examples retain their scope; none is an independent validation of Echora.
1. Pang Wei Koh and Percy Liang, "Understanding Black-box Predictions via Influence Functions," ICML 2017; Amirata Ghorbani and James Zou, "Data Shapley," ICML 2019; Andrew Ilyas and colleagues, "Datamodels," 2022; Sung Min Park and colleagues, "TRAK," ICML 2023; OpenAI Cookbook, "Reproducible outputs with the seed parameter."
Each paper treats removing a training example and retraining as the reference point for training-data attribution and offers a cheaper estimate because retraining at scale is impractical. OpenAI's documentation describes its model outputs as non-deterministic by default, with a seed giving only best-effort repeatability. None of these sources evaluates Echora.
Influence functions · Data Shapley · Datamodels · TRAK · OpenAI on determinism
2. Anthropic, "Mapping the Mind of a Large Language Model," 21 May 2024.
Anthropic's interpretability team describes AI models as mostly treated as a black box, and the internal features it identified as a small subset of the concepts a model holds. Partial understanding, stated by the people who build the models.
3. Suno, "Introducing v6," 9 September 2026; Billboard, "As Suno Rolls Out Its New AI Models With Industry Licenses, What Does That Mean For the Music Business?" (Kristin Robinson and Dan Rys), September 2026; Billboard, "As Some Labels Strike Deals With Suno, Artist Reps Sound Off," September 2026.
Suno's announcement names partners and forthcoming artist opt-in experiences. Billboard reported, citing industry sources and Suno executives, that Warner licensed a subset of its catalog for training, that partners are compensated through a revenue share, and that how participating music is selected remains unclear. Neither discloses how revenue is allocated across works or to individual artists and writers.
Suno announcement · Billboard FAQ · Billboard, artist reps
4. Fabio Morreale, Wiebke Hutiri, Joan Serrà, Alice Xiang and Yuki Mitsufuji, "Attribution-by-design: Ensuring Inference-Time Provenance in Generative Music Systems," 9 October 2025.
Distinguishes training-time and inference-time attribution. Research context, not an evaluation of Echora.
5. OpenAI, "Searching the web with ChatGPT."
Supports the example of retrieving material and citing sources. Citations are not a complete account of internal model influence.
6. The Mechanical Licensing Collective, "Market Share Distributions"; press release on the annual membership meeting, 30 September 2026.
Unmatched and unclaimed blanket royalties are held for at least the three-year statutory minimum, then allocated pro rata to songs that were matched. First monthly distribution announced for April 2027, starting with January 2021 usage. A description of U.S. mechanical royalties, not of any AI license. Shares in a registered ownership dispute are held until resolved, under the MLC's dispute policy.
7. Shutterstock, "Shutterstock Data Licensing and the Contributor Fund," and contributor FAQ on AI-generated content.
Allocation by content and metadata in a licensed dataset; separate compensation tied to AI-generated downloads. Neither is a record of references supplied to an individual generation.
Contributor Fund · Contributor FAQ
8. Lauren Goode, "Getty Images Plunges Into the Generative AI Pool," WIRED, 25 September 2023.
Launch-era interview describing contributor revenue allocation based on the training collection and its performance.
9. American Federation of Musicians, "Recording & Digital Media."
New-use payments under relevant agreements, including recordings reused outside their original purpose. The obligation depends on the applicable agreement.
10. SoundExchange, "Digital Performance Royalties."
The 45/50/5 allocation concerns qualifying U.S. statutory digital sound-recording performances, not all streaming or composition income.
11. AFM v. Warner Records Inc., Atlantic Recording Corp. and Universal Music Group, Inc., No. 1:26-cv-04760-ER (S.D.N.Y.).
Opposition filed 4 September 2026; replies filed 11 September 2026 (ECF 50-52). The essay reports the parties' positions, not an adjudicated result. Reported by Music Business Worldwide, 17 September 2026.
AFM opposition · UMG reply · Warner/Atlantic reply · MBW report
12. Jerry Siegel, Wikipedia, "Superman" section.
In 1938 Jerry Siegel and Joe Shuster sold the rights to Superman to Detective Comics, Inc. for $130 and later contested the sale in court without success. The Billie Eilish sentence in the same passage is an imagined example, not a report of any agreement.
13. The Authors Guild, Model Trade Book Contract, "Advance Against Royalties."
An advance against royalties as immediate payment inside a continuing agreement. The particular contract governs timing, deductions and recoupment. The description of music advances is the author's own experience.
Model contract · Earn-out explanation
14. Google Research, "Announcing the first Machine Unlearning Challenge," 29 June 2023.
Explains the difference between deleting data and removing its influence from a trained model.
15. BMI, "What is a Cue Sheet?"
The production and music-use information used for performance-royalty administration.
16. Directors Guild of America, "Residuals" and "History of Residuals."
The move from repeated live-radio performances to recorded replay; residuals distinguished from profitability. The applicable agreement determines eligibility and calculation.
17. SAG-AFTRA, 2023 TV/Theatrical Artificial Intelligence Resources; 2026 TV/Theatrical Contracts.
Specific, separately acknowledged consent for digital-replica use, subject to exceptions; ownership of recorded material distinguished from consent for a new replica use. Contract-specific protections, not a universal likeness law.
2023 guidance · 2026 contract hub
18. SAG-AFTRA, Contract Bulletin: Artificial Intelligence and Commercials, June 2025.
Digital-replica consent, generated-performance payment and holding/use fees under the 2025 Commercials Contract. Exact rates and exceptions depend on the agreement.
19. Pierre Bosset / Adobe, "How LG Household & Healthcare scales digital twin use with Substance 3D & Firefly," 15 January 2026.
Company-published interview. An example of separating product fidelity from scene generation, not an Echora deployment. Depicting a product, manufacturing it, and using its trademark or protected design involve different rights (see USPTO and U.S. Copyright Office, "Useful Articles").
Adobe case study · USPTO · Useful articles
20. Valve / Steamworks, "Content Survey"; SAG-AFTRA, "Achievement Unlocked!," 11 August 2025.
Pre-generated and live-generated AI content disclosures; summary of the 2025 Interactive Media Agreement, including replica consent and specified real-time generation. The agreement governs covered work.
21. Arm Holdings plc, Form F-1, 2023.
Business section: some products generate royalties after 25 years; approximately 46 percent of FY2023 royalty revenue came from products released 1990-2012. Historical comparison, not an Echora forecast.
22. The Mechanical Licensing Collective, "Blanket Royalties."
Monthly service reporting, work matching and distribution under the U.S. blanket mechanical license. An administrative precedent, not a license for AI reference use.
23. Huiqiang Jiang and colleagues, "LLMLingua: Compressing Prompts for Accelerated Inference of Large Language Models," EMNLP 2023.
Context for reducing information processed in AI workflows. It does not establish Echora's savings, encoding, compatibility or patent scope.
24. PCAOB, "About"; NIST, "Office of Weights and Measures."
Examples of outside scrutiny and commercial measurement oversight. Neither body oversees Echora.
25. OpenAI, written evidence to the House of Lords Communications and Digital Committee inquiry into large language models, December 2023, as reported by Euronews, 9 January 2024.
OpenAI's position in the same filing is that copyright law does not forbid training.
