2. Value & Exchange

Last year the UK’s Association of Illustrators asked its members what AI had done to their livelihoods. Nearly 7,000 answered. A third said they had already lost work to it, at an average of about £9,300 each. And 93 percent said they couldn’t opt out of AI training datasets without seriously damaging their business.

The Society of Authors found much the same among literary translators — more than a third losing work, more than four in ten watching their income fall. In online freelance markets, where the effect can be measured rather than reported, writers’ earnings dropped about 5 percent in the months after ChatGPT arrived and illustrators’ nearly 10 percent.

And then there’s stock imagery, where you’d expect the clearest collapse of all. When one large image marketplace opened its doors to generative AI in late 2022, the total supply of images jumped 78 percent. The number of firms producing them rose 88 percent. Sales rose 39 percent. What fell was the participation of the artists who weren’t using AI, down 23 percent.

That isn’t the story we’d expect. The market didn’t contract. It grew. So big that human artists got lost in the flood. And the human artists left it.

Now look at the other end of the same economy. Reddit is reported to be taking about $60 million a year from Google to license its users’ posts for training. News Corp’s OpenAI deal is reported north of $250 million over five years. Getty Images sells a generative image product trained on its own archive and now puts its pictures inside ChatGPT. And Anthropic has paid book authors $1.5 billion to settle claims over pirated training data — the largest copyright settlement in American history, about $3,000 a book.

Two ends of one ecosystem: creative value transferred at scale, and creative work devalued at scale. The two happen in different markets. But together they tell you how the whole structure is being rebuilt, and by whom.

This is the second of three framing essays. The first was about attribution — the system of signals that tells us who made something, and how, and whether we can verify any of it. This one is about value and exchange: what happens to the economic architecture of creative work when the old compensation mechanisms are failing and the new one is being assembled largely without the people it’s supposed to compensate.

The deceptively simple question is this. How is creative labor valued and paid when the old ways of valuing and paying aren’t working, and the new ways are uncertain, unproven, and being written by whoever is in the room?

#What the exchange system used to do

For most of the twentieth century, the economics of creative work rested on four imperfectly connected mechanisms. Copyright licensing generated royalties when someone used a protected work. Market sales — box office, retail, subscription, advertising — compensated finished output. Collective administration handled the cases where direct licensing couldn’t scale. And labor contracts, mostly through guilds and unions, paid for creative work produced inside an employment relationship.

None of these covered everything, and even together they left individual artists exposed in well-documented ways. But they shared a structural assumption: value was created at the point of a finished work, and that value flowed back to the people who made it. An artist made a thing. Someone bought it, licensed it, or paid the artist’s labor to produce it. The circuit closed on the output.

All four are now under pressure — copyright strained by the training-data question, markets flooded with zero-cost synthetic output, collective administration unbuilt for AI-era extraction, labor contracts renegotiated under duress across every creative industry.

And AI has added a layer the old architecture was never designed for: value extracted from creative inputs, before any finished work exists. Training-data licensing isn’t just a bigger version of the piracy debate. It’s a different market, with different actors and different dynamics.

The exchange system used to run on one track. It now runs on two, and the track that’s paying isn’t where most creators are.

#Recompilation: from layers to licenses

The first essay established that AI decomposes creative identity into separable, extractable layers — authorship, provenance, identity, artistic essence, training data, disclosure, audience. That’s the attribution problem.

But attribution alone doesn’t produce compensation. For value to flow, those layers have to be recompiled — reassembled into units that can be priced, licensed, traded, and distributed. Recompilation is the economic operation that turns an attribution map into a functioning market.

How that happens is not neutral. Every mechanism makes assumptions about which layers count, who holds the rights to them, and which transactions get paid for. Those assumptions determine who gets the money. Four architectures are competing for the job, and they are not equal.

Private licensing. The default: bilateral deals between large rights holders and AI developers, and the only one moving at scale. The music labels are licensing catalogs, Reddit licensing user content, the publishers are licensing archives, Getty is licensing images. Real revenue, real precedent, real new indemnity norms. These deals are familiar and they work.

The problem is where the money lands. Royalty distribution from catalog licensing runs through the same accounting that has rarely paid individual artists on terms anyone considers fair. The AI deal inherits the distributional pathology of the pre-AI catalog deal, at massively greater scale and with more layers.

There’s a second problem, and the past eight months made it vivid. In December, Disney and OpenAI announced the largest studio deal yet — three years, more than 200 iconic Disney characters, bundled with a billion-dollar Disney investment in OpenAI. It was, briefly, the template for such deals. But in March, OpenAI abruptly shut down Sora and left standalone video generation services altogether. The investment never closed and no money changed hands.

A licensing deal inherits the product risk of the other party. The rules a private contract establishes last exactly as long as the commercial rationale behind them, while legislation doesn’t get cancelled because a feature underperformed.

Collective licensing. Start with what the federal government actually said, because it isn’t what the headlines said. The White House’s national AI framework, released in March, holds that training AI models on copyrighted material does not violate copyright law, and leaves fair use to the courts. On compensation it goes exactly this far: Congress should consider enabling collective negotiation without antitrust liability — and then, in the next sentence, that any such legislation shouldn’t address when or whether licensing is required at all.

That’s an antitrust permission slip attached to a position that no license is needed. I’d call it the most consequential sentence in American AI policy right now, because a framework whose baseline is training is already lawful removes the leverage that makes even voluntary licensing negotiable. You can’t bargain hard over a right the government has just told you you don’t have.

Meanwhile the collective architecture is getting built anyway — and not by the people you’d predict. In June the music publishers announced the first industry-wide AI licensing deals, opt-in for members, valuing songs on par with recordings. In March the UK publishers launched a collective scheme covering training, fine-tuning and retrieval; 250-plus publishers have joined. The recorded-music side went the other way entirely in the US, telling the White House that licenses should be negotiated in the free market without regulation.

Which is itself a finding. The architecture a rights holder prefers seems to track how concentrated its membership is. Three major labels can negotiate one at a time. Several thousand publishers can’t, so they built a collective.

The historical analogy matters, and it’s more useful than it’s usually made out to be. A century ago, when music performance had scaled past the point where per-use licensing could work, ASCAP was founded to license a pooled repertoire in a single blanket agreement. It worked well enough that it’s still the bedrock of the performance rights economy.

But here’s the part that goes unmentioned, and it changes what the analogy is good for. ASCAP is not a voluntary-collective success story. Since 1941 it has operated under an antitrust consent decree, administered by a federal court, that requires it to grant a blanket license to any applicant, hold rights non-exclusively, license on non-discriminatory terms, and submit to a rate court that fixes a fee when the parties can’t agree.

Which means a user who can’t reach terms with ASCAP may perform the music anyway and let a judge set the price. That’s a compulsory license in everything but name. Collective formation was voluntary. Everything that makes the blanket license credible on both sides was imposed by antitrust enforcement and has governed the system for eighty-five years.

So the question isn’t only whether the ASCAP model can be adapted to multi-layered AI extraction. It’s whether anyone is prepared to build the backstop that made it work.

The obstacles to adapting it are real. ASCAP solved a one-layer problem, and it rested on something AI training lacks: a fixed, linear transmission. A song played in a room has a beginning, an end, and a countable audience. A model ingests a work, abstracts it into weights, and emits output bearing no traceable relation to any input. There’s no actual performance to count.

So the classic framework is probably insufficient. But the architectural logic may be the only approach that scales down to individual artists, and the question isn’t whether to use the model — it’s whether what gets built can handle the decomposition without reproducing catalog-era distribution.

Compulsory licensing. The most muscular solution, and the least likely in this political environment. But the usual framing — that it’s something America has never done — is wrong, and the correction matters more than the error.

We’ve done it repeatedly, and recently. The Section 115 Copyright Office mechanical license lets anyone record their own version of a song once it’s been released, at a rate set by the Copyright Royalty Board — and the Music Modernization Act converted that, for digital uses, into a blanket license administered by a collective, operating since 2021. There’s the cable compulsory license. And there’s SoundExchange, whose distribution formula is written into statute: 45 percent directly to featured artists, 5 percent to a fund for non-featured performers, 50 percent to the recording owner.

That 45-and-5 is probably the most important number in this essay. It’s a statutory override of private contract. Congress wrote it because it didn’t trust the pass-through — because it understood that setting the aggregate price does nothing if the money stops at the aggregator. A compulsory system can dictate internal splits, not just the price, which is precisely the failure private licensing reproduces and precisely the thing no bilateral deal will ever fix on its own.

Which is to say we already built the thing this debate treats as speculative. The obstacle isn’t that the model is untested. It’s that Congress built it once for a market it understood and hasn’t been asked to do it again (for a market it doesn’t, at least yet, understand).

What’s pending now is weaker. The proposed TRAIN Act would let a copyright owner compel production of records identifying training material; the COPIED Act would prohibit training on provenance-bearing content without consent. Both sit in committee without a hearing. And they aren’t cousins: TRAIN is a discovery mechanism, while COPIED is the opposite of a compulsory license. A compulsory license removes the veto and fixes the price, COPIED entrenches it.

Europe shows the limits from the other direction. Mining lawfully accessible works is permitted unless rights have been reserved “in an appropriate manner,” but a German appellate court held in December that a plain-English “no scraping” clause wasn’t machine-readable enough to count. The actors with the infrastructure to express and process a reservation accumulate power. The actors nominally holding the right do not.

Individual opt-in and opt-out. The most artist-centered architecture in theory and the weakest in practice. A creator saying “don’t train on my work” has no enforcement mechanism, no detection capability, and no bargaining power.

We now have a government’s own finding on exactly that. In March, after a full consultation and an economic impact assessment, the UK concluded that a broad exception with opt-out was no longer its preferred way forward — because creators told it that managing opt-outs across large volumes of work would be impossible. That’s the illustrators’ 93 percent as a policy conclusion.

The artist-led experiments show what individual opt-in could look like — Holly Herndon’s Holly+ treated her voice as a deliberately separable asset governed by a collective of stewards, and the more credible current versions are infrastructure plays: do-not-train registries, likeness authentication. But without an aggregation layer underneath, these stay boutique. Licensing your own voice matters only if something connects your opt-in to compensation at scale, and for individuals that something doesn’t exist.

Four architectures, one of them moving. And that is the core of what’s happening here: because private licensing is functioning while the other three are stalled or under construction, private licensing is becoming the de facto policy framework for the AI creative economy. Rates, carve-outs, indemnity norms, precedent on what counts as fair compensation — all of it established inside bilateral negotiations between concentrated rights holders and concentrated AI developers.

We’ve seen this before. It’s how the digital music transition played out twenty-five years ago: by the time regulators and creative communities had a framework, the platforms had already written the rules. This transition is on the same path, moving exponentially faster.

#The value-price divergence

There’s a difficulty with how the exchange problem is usually framed, and I think it obscures what’s actually happening.

The conventional story is that supply is now effectively infinite, so prices fall and traditional markets collapse. That’s partly right and badly incomplete. What’s actually happening is a divergence: the price of creative outputs is being driven toward zero while the value of the human contribution becomes more differentiated, not less. Price and value are moving in opposite directions.

Three effects compound.

The output market displaces rather than simply shrinks. Freelance illustration, literary translation, copywriting, voice-over, stock imagery — work disappearing, rates compressing, whole categories dissolving. But “collapse” is the wrong word for what the stock-imagery evidence shows, because that market grew while human participation has fallen by nearly a quarter.

What’s being repriced isn’t the sector. It’s the human contribution inside it, against a new cost floor of effectively zero for competent median-quality output. And competent median work was the economic foundation of most working creative careers. That foundation is being dismantled inside markets that may look perfectly healthy from the outside.

The macro evidence points the same way, with caveats. Stanford’s payroll study finds a 16 percent relative decline in employment for workers aged 22-25 in the most AI-exposed occupations — about 6 percent in absolute terms, against a rise for older workers in the same jobs. The clearest cases are software and customer service rather than creative work, and the authors now say the effect is statistically clean only after 2024. Suggestive, not dispositive. But the shape — entry level absorbing the displacement while senior roles hold — is the shape the creative professions are showing too.

The input market captures. Training-data licensing is generating real money, and almost none of it reaches the artists whose work built the systems. A session musician whose playing is documented on forty years of major-label recordings will never see a line-item reading “your performance contributed to the training of a voice-generation model.” The revenue goes to the catalog holder.

This used to be a structural prediction. It isn’t anymore. In June 2026, the American Federation of Musicians sued Universal and Warner — not the AI companies — alleging the labels took significant compensation from Suno and Udio and licensed the catalogs without compensating or crediting the session players on those recordings.

This was filed in federal court. And note where it points: at the aggregator, not the extractor. It’s the only action in this Observatory that names the intermediary rather than the AI company, which is exactly where the failure sits.

This isn’t bad faith on the labels’ part. It’s how catalog accounting works, and how intermediary economics has always worked. But it means the input market is functioning while its distributional pathway is broken. The musicians have declined to accept the structural explanation, and whether “structural” and “actionable” are the same thing is now a question in front of a judge.

The provenance premium is weaker than it looks. As the output market fills with synthetic work, the slice that will pay more for verified human work should in theory become more valuable. Live performance. Verified human authorship. Direct artist-to-audience economies. The organic-food market for culture.

Some of it is real — live music posted record revenue last year, 159 million fans across 55,000 shows. But that premium predates AI by decades, and the industry’s own explanation is international expansion and stadium supply, not anxiety about synthetic music. Nothing comparable has materialized for recorded or written work: Bandcamp Fridays paid artists about a sixth of one percent of what Spotify paid out, and Substack’s paid subscriptions have been flat for more than a year.

That distinction is a harder constraint than the analogy conveys: the premium has materialized for what cannot be synthesized at all, and not for what merely wasn’t.

There’s a deeper fragility too, because the premium doesn’t attach to anything an audience can detect. In a study of 9,000 listeners across eight countries, respondents were played three tracks — two AI-generated, one human — and asked to sort them. Ninety-seven percent failed.

This year’s research is sharper: people can’t detect AI, and they punish it severely once told. Play listeners identical audio and label half of it AI-performed, and the label alone depresses ratings of quality, engagement and liking, with four in five reporting they heard differences that did not exist. Manipulate origin and disclosure separately and listeners can’t tell them apart, slightly prefer the AI music when blind, value it equally — and cut both their appreciation and their willingness to pay the moment they’re told.

Detection and valuation have come apart. Which means the provenance premium is a function of disclosure infrastructure, not of perceptible quality. Strip the label and there’s nothing underneath it a listener can hear.

That’s also why the organic comparison holds up better than it’s usually made to. Give people physically identical foods and label one organic, and they rate it tastier and more nutritious and will pay 23 percent more. The premium is manufactured by the label — by a credible, legally enforced, independently audited certification regime. And after three decades of exactly that regime, organic is about 6 percent of the American food market.

Six percent, with mandatory certification and federal enforcement behind it. That’s the number to remember when anyone offers a human-made premium as a survival strategy for a displaced professional class.

The premium survives only where human origin is bundled with something else — the live event, the relationship with an artist, the status of a verified original. Strip those away and there’s no premium at all, which is exactly the tier where displacement is worst. The algorithmic culture we already live inside had loosened the artist-audience relationship long before any of this. That connection has been devalued once already. Human provenance may go the same way.

So AI-and-creative-compensation isn’t one question. It’s three, with three different answers. How do we redistribute input-market value back to the creators whose work built the systems? How do we sustain working professional creative labor when the mass output market is being flooded? And how do we protect the provenance premium without freezing it into a luxury niche?

Conflating these into one “creator compensation” issue produces a policy conversation that uses artist-centered language while building mechanisms that serve aggregators. The labels’ deals are solving the first question, imperfectly. They do nothing for the second. They’re marginal to the third. When we talk as if compensation means all three, we let the answer to the first stand in for all of them.

That’s not necessarily cynical. These three have run together in the culture for so long that most of us don’t separate them without deliberate effort. But the framework has to, or we build a solution to one and declare victory for three.

#Why the functional lens matters

The first essay argued for organizing AI creativity policy by function rather than by issue or jurisdiction. The case is even stronger here.

Convergence across domains becomes visible. A session musician, a freelance illustrator, a literary translator and a voice actor are experiencing versions of the same value-transfer failure, but policy tracking treats them as separate industry disputes. They’re the same recompilation failure at different layers.

The jurisdictional divergence comes into focus. Europe’s opt-out with a compliance duty attached. Japan’s exception, of a broad carve-out with a proviso that withdraws it where use would unreasonably prejudice the copyright owner, and guidance excluding fine-tuning to imitate a specific creator and copying databases sold for analysis where a licensing market already exists. That last one is striking: even the most permissive major regime pulls its exception once a market appears. Then there’s the American private-licensing default, layered on a federal position that no license is required. The UK, which consulted on opt-out and walked away. And Canada, which consulted two years ago, produced a policy paper, and has introduced nothing for creative work.

Same question, structurally incompatible answers, and a compliance environment that can’t function for anyone working across borders. (China belongs in a different sentence. Its edicts govern who may deploy models and how outputs get labeled. It has no collective licensing and no statutory royalty. That’s a governance regime, not a compensation one.)

And the policy gaps become visible. Private licensing has no mechanism for individual participation in catalog-level deals, and no durability past the commercial life of the product it’s attached to. Collective licensing is being retrofitted where a collecting society already existed, and has neither the capacity for AI-scale use nor a rate-setting backstop. Compulsory licensing has more American precedent than the debate acknowledges and less political oxygen than it needs. Individual opt-out has no aggregation layer and no enforcement — a conclusion one government has now reached in writing.

And the gap that ran through the attribution layer runs through this one: the unaffiliated individual artist. Guild-protected creators have frameworks. The Authors Guild’s answer is a partnership with an opt-in marketplace where authors enroll individual works and get a week to decline each proposed deal — a brokered marketplace, not collective licensing, and eighteen months in the Guild itself says deals through it are still rare.

Most working creative professionals aren’t in a guild with bargaining power. Every architecture assumes either institutional representation or aggregator-level rights-holding. The individual artist without institutional backing falls between all four.

That gap can now be measured. Every one of the 379 developments this project tracks (as of July 28, 2026) has been coded for who made the operative decision and whether creators were a party. Across the 315 where something was actually decided, creators were a party in 21 percent, consulted in 8, absent from 72. Creators or their unions held the pen in 18 percent — without a union behind them, in 6.

Six percent is the unaffiliated artist’s share of the decisions that will govern how their work gets paid for.

#The institutional piece: the labels’ pivot

Attribution systems are inert without institutions capable of administering them. In the first essay the failure mode was institutional disruption — the firing of the Register of Copyrights at perhaps the most consequential moment in American copyright history. That one resolved: Shira Perlmutter sued, the D.C. Circuit reinstated her, and this summer the Supreme Court declined to remove her while the case proceeds. The institution held, by litigation, over fourteen months.

The failure mode here is different. It isn’t disruption. It’s capture. And it doesn’t require breaking a public institution, which is why the Perlmutter outcome doesn’t touch it.

Consider the labels’ trajectory. Through 2024 the posture was confrontational — infringement suits against the AI music companies, cease-and-desist letters, public advocacy for strong training-data protections. The labels positioned themselves as champions of creator rights against a rapacious technology sector.

By late 2025 it had split three ways. Universal settled with Udio. Warner settled with Udio and then with Suno. Sony settled with neither — and when a judge refused to let it add some 30,000 recordings to its existing case, it filed a second suit. Universal and Sony are still litigating against Suno, where they’ve moved to expand from a few hundred works to roughly 61,000, identified by fingerprinting the training data they got in discovery.

So “sue the pirates, license the allies” is too clean. What’s actually happening is a negotiation conducted through litigation, in which the holdout is plausibly extracting the best terms. That’s not a strategic pivot completed last year. It’s a pricing mechanism, and litigation is the de facto discovery process.

But here’s what it means. The labels aren’t just licensing catalogs. They’re establishing the architecture that smaller rights holders, independent labels, and eventually individual artists — which is to say all of us — will have to operate inside. The rates they negotiate become benchmarks. The carve-outs they secure become templates. The indemnity structures they accept become norms. When a collective arrangement finally does get built, it will be negotiated in the shadow of terms Sony and Universal have already set.

It’s no coincidence that Universal reported record revenue for 2025, which Lucian Grainge, CEO of UMG called “another standout year.” Business is very good. And Grainge is very smart.

The labels are writing policy through private contract — not in the slow-moving legislatures, not in the courts, and certainly not with the artists who’ll have to live under the deals.

The Getty paradox makes the strategy visible. Getty is suing Stability AI over training. It’s licensing its catalog for display inside ChatGPT while pointedly not licensing it for training. And it sells a generative image product trained on its own archive. These look like contradictions. They really aren’t. They’re the rational strategy of an aggregator optimizing across every architecture at once: keep the licensing revenue, keep the litigation leverage, stake out a position in generative products, open a display line that sidesteps the training question, and hedge against whatever regime eventually arrives. One actor playing every board.

#What the evidence reveals

The Observatory now tracks 379 developments as of July 28, 2026, and the value-and-exchange ones cluster into recognizable patterns.

The freelance squeeze. The illustrators’ survey, the translators’ survey, and a voice-actors’ survey whose finding isn’t falling rates but that 7 percent have had a synthetic version of their voice made without permission. And the freelance-market study that measures rather than asks, where the effect turns out to be concentrated among the highest earners, not the lowest.

Our two best datasets disagree about who is most exposed — payroll data says the entry level, freelance data says the top — and that disagreement is a finding the compensation debate hasn’t yet absorbed.

The training-data deals. Reddit’s Google arrangement, never confirmed in a filing and now up for renewal with Reddit publicly weighing whether to walk away. News Corp’s deal. AP’s, on undisclosed terms. Getty’s display partnership. The label settlements, all with undisclosed financials. The through-line is this: none of these compensate the individual contributors who produced the content. Reddit users weren’t paid. AP’s bureau journalists weren’t paid. Session musicians haven’t seen the money, and have sued to say so.

The collective builds. The publishers’ industry-wide deals in June, first of their kind, and the UK scheme in March with 250-plus publishers are both built on top of existing collecting societies. But the marketplace is built from scratch for individuals, still without a disclosed completed license.

The stream-fraud economy. Deezer now receives roughly 90,000 fully AI-generated tracks a day — more than half of everything uploaded — and estimates up to 85 percent of the streams those tracks get are fraudulent, against a platform-wide fraud rate around 8 percent. Spotify says it removed more than 75 million spammy tracks in a year; that’s its word, not “AI.”

But the number that lands hardest is Deezer’s: AI tracks account for between 1 and 3 percent of streams, but more than half of what is uploaded to the platform. A fiftieth of what gets heard.

That gap separates two failures we keep seeing together, but they need different remedies. Dilution is a slowly growing revenue pool spread across a catalog growing much faster than demand. Fraud is synthetic engagement pulling real money out of that pool. The money leaking to AI tracks leaks mostly through the second channel, not because listeners are choosing them. Any proposed remedy has to say which one it’s fixing.

The author settlement, and what it actually settled. Anthropic’s $1.5 billion settlement with authors got final approval this month, with the court cutting the plaintiffs’ fee request nearly in half. The deal with Anthropic pays about $3,000 each for some 482,000 works, a fraction of the statutory ceiling, which is why the deal capped Anthropic’s risk more than it enriched anyone. But the crucial detail is what the class covers: books Anthropic downloaded from pirate libraries. The court had already held that training on lawfully acquired books was fair use. So the $1.5 billion is the price of the acquisition method, not of training, plus destruction of the pirated files, the only thing in this entire layer that removed something from circulation rather than setting a price for it.

The music publishers sharpen the contrast. Their lyrics suit against Anthropic has produced a stipulation and a denied injunction — no damages, three years in. And in January they opened a second front, a mass-torrenting complaint covering more than 21,000 compositions and naming Anthropic’s CEO and a co-founder personally. That has no precedent here, and it’s the clearest signal available that plaintiffs think litigation can still produce something licensing won’t.

This month, the labeling question got an answer. Eight organizations — the RIAA, IFPI, the Recording Academy, SAG-AFTRA and others — launched a unified track-labeling scheme with two tags: “AI-Generated” and “AI-Assisted.” TIDAL began demonetizing fully AI-generated tracks two weeks ago.

The caveats are the story. The scheme is entirely voluntary and self-reported, flagged by artists, labels and distributors themselves. No streaming platform has committed to it. And it covers sound recordings only, not composition, not lyrics, not cover art, not video.

Set that next to the willingness-to-pay research and you get the sharpest statement of reality. This is the organic-certification moment for music, without the certification. Organic food works because the claim is mandatory to make, independently audited and legally enforceable, and it still took three decades to reach 6 percent of the food market. A voluntary self-reporting regime, in a market where disclosure has been shown to reduce what audiences will pay, is asking rights holders to volunteer information that costs them money.

The short conclusion: Not gonna happen.

#The structural tensions

Six unresolved tensions run through the exchange layer.

Speed versus representation. Private licensing is fast. Everything else is slow. Waiting for a participatory architecture means accepting that the deals cut now will become the precedent. Moving with what exists locks in the distributional pattern. No clean answer, only trade-offs taken consciously or taken by default. The UK took the second one deliberately and said so.

Aggregator interests versus creator interests. The rights aggregators negotiating these deals are intermediaries whose commercial interests structurally diverge from the creators they nominally represent. That’s not an accusation; it’s how intermediary economics works. But treating major-label licensing as a proxy for artist compensation is wrong. Even where labels say in good faith that distribution to artists will follow, the track record on that promise across a century of music-industry intermediation is not encouraging — and the musicians’ union has now put the proposition to a court.

Catalog value versus living-artist value. Training data is most valuable when it’s deep and broad. The actors with deep broad catalogs are estates, labels and corporate aggregators. The actors making new work right now don’t have comparable catalogs, because they haven’t had time to accumulate one. The architecture being built therefore favors estates and aggregators over working living artists, without anyone designing it that way. And the asymmetry runs deeper than timing: aggregators monetize static historical assets carrying no maintenance cost, while living artists fund their cost of living and their new production out of a compressing output market. The architecture pays for past human labor in order to subsidize future synthetic labor.

Input compensation versus output compensation. Every input mechanism is retrospective. It pays for what was already made. It says nothing about how an artist producing new work today gets paid for the labor of doing it. If input licensing becomes the dominant mechanism, we’re building an economy that pays for the past and doesn’t yet pay for the present.

The provenance premium versus the mass market. A premium class of verified human work can be sustained. It won’t scale to the professional class being displaced. If it becomes the default survival strategy, creative work reorganizes around a luxury logic with its own gatekeeping, and the democratization AI in principle enables, collides with a market that prices out most of the democratized. Both can be true at once, and usually will be. And the premium is narrower than “luxury” implies, because it rides on experience and relationship rather than on anything perceptible. Audiences can’t detect AI and will pay less once told — which means the premium is a property of the label, not of the work.

Voluntary versus compulsory. Every American architecture now being proposed is voluntary — and the federal framework underneath them holds that no license is required at all, which removes the leverage that makes voluntary bargaining work. Voluntary mechanisms at the aggregator level have never produced equitable distribution to individual creators without a compulsory backstop. The evidence isn’t hypothetical or obscure: the blanket mechanical license and its statutory rate, SoundExchange’s 45-percent direct-to-artist carve-out, and ASCAP itself, whose blanket license has rested since 1941 on an antitrust decree and a rate court. The model this debate keeps invoking as the great voluntary-collective success is a compulsory system in everything but name. Remove the backstop and collective administration reverts to aggregator-favoring behavior.

#Where this is heading

Several directions show up in the data, though none is certain.

• Private licensing will set the operating template unless a compulsory backstop is legislated, and the politics make that unlikely soon. The commercial architecture is becoming de facto policy. But Disney’s collapsed deal with OpenAI is a reminder that a template written in contracts is only as durable as the products those contracts attach to.
• Collective licensing will be a retrofit rather than built fresh. That’s stopped being a prediction. Both schemes that launched this year were built on organizations that already collected and distributed money. Whether twentieth-century collective administration can handle multi-layered extraction is the open structural question — and it matters because there’s no realistic timeline to build anything new, and the one attempt to build for individuals is eighteen months old without a deal.
• The input/output divergence will widen before it closes. Catalog value appreciates as training demand grows. Output prices compress as synthetic supply floods in. The gap is structural and won’t close without intervention, and intervention isn’t imminent.
• The provenance premium won’t absorb the displaced. The organic analogy holds, and its number is the discipline: 6 percent of the food market after three decades of mandatory, audited certification. This month’s labeling scheme has none of those properties. And the premium may be more fragile still, because it exists only where disclosure exists, and disclosure is the thing rights holders have a measurable financial reason not to make.
• Watch display licensing. When Getty licensed images to OpenAI, it licensed them for display, not training — pictures surfacing inside ChatGPT with attribution and a link. Those deals have gone from a handful two years ago to dozens this year, while training licenses flatten. That’s a different compensable event from anything in the four architectures, all of which answer what is owed for the corpus. Display answers what is owed for the surfacing — and it’s the one event in this whole layer that can actually be counted, which makes it the only place a usage-based distribution formula could run.
• Cross-border rights management will break before it gets rebuilt. Either international harmonization emerges, slowly and painfully, or jurisdictional arbitrage becomes the operating norm — which is fast, and already underway.
• The attribution functions and exchange architectures will have to integrate. Seven attribution layers and four recompilation mechanisms are describing the same decomposed creative economy from two sides. Building a single framework that identifies contributions at each layer and compensates them is the structural work of the next decade. Nothing being built integrates at that level.

#The exchange layer

The first essay described attribution and accountability as the trust layer — the infrastructure that identifies who made what, how, and whether we can verify any of it. Value and exchange is the exchange layer, the infrastructure that moves compensation across those identified contributions. The two depend on each other. You can’t pay for what you can’t identify, but identification without compensation is just accounting.

What I’ve tried to map here is an exchange layer under construction — further along than the trust layer in some respects and more broken in others. The input-market deals are moving real money and setting real precedent. That’s progress. The pathway from those deals back to individual creators is broken, largely because the architectures in use were designed for aggregator-level rights management and not for the multi-layered reality of AI extraction. That isn’t progress. It’s a structural pattern we’ve seen before, rebuilt at higher speed and greater scale.

And there’s one more thing worth naming, because it runs underneath all of it. The mechanisms that have actually worked for individual creators in American law — the blanket mechanical license, SoundExchange’s statutory 45 percent, ASCAP’s rate court — all share a feature. In each case someone decided that leaving distribution to private negotiation would fail, and wrote the pass-through into law.

None of the four architectures now competing to govern AI compensation has that feature. Three of them can’t. The fourth would require Congress.

The recompilation question isn’t whether value can flow through the new mechanisms. It can, and it’s flowing right now. The question is to whom, and the current answer is: mostly to the aggregators closest to the licensing table, very little to the individual working artists the system was supposed to be about.

The third essay, Cultural Infrastructure and Ecology, takes up what happens to the broader cultural system when the trust layer and the exchange layer produce these patterns. Who enters the creative professions when the output market has been flooded? Which cultural forms get sustained and which get starved? What institutions shape the answers, which are being built, and which are being dismantled at the worst possible moment?

Attribution tells us who made what. Exchange tells us who gets paid for it. Infrastructure and ecology tell us what kind of culture the first two are producing — and whether it’s a culture any of us would have chosen to build.

View Sources