3. Cultural Infrastructure & Ecology

In May 2025 the National Endowment for the Arts began terminating grants it had already awarded, under what it called “new priorities.” Those priorities included the usual culture-war directives and, tucked into the same list, the item “foster AI competency.” Among the terminated projects was one to SMU DataArts. It was for an AI project.

Nobody knows how many grants got canceled. The NEA has never published a count, and the only tally that exists was crowdsourced by the field itself. Two budgets have since proposed eliminating the NEA, the NEH and the Institute of Museum and Library Services outright but Congress hasn’t done it. However, the Corporation for Public Broadcasting didn’t survive. It was defunded in 2025, and dissolved by its own board in 2026, ending fifty-eight years of public media funding.

The dollar amounts were never the point. Americans gave $27 billion to arts, culture and humanities last year, and the NEA’s entire budget is less than one percent of that. What federal funding does that matters is confer legitimacy and distribute access. An NEA grant, like a serious review or a respected prize, is one of the ways a field tells itself what counts as good, and which work is worth sustaining when the market won’t. Cutting it removes money. But it also removes a signal, and there aren’t many left.

Meanwhile Stanford’s payroll analysis found a 16 percent relative employment decline for workers aged 22 to 25 in the most AI-exposed occupations, against an increase for older workers in the same jobs. I’ll come back to how contested that is, because the difficulty of pinning it down turns out to be perhaps the most important thing about it.

The institutions that mediate between creators and culture are collapsing at the same moment the entry ramp into creative professional life narrows. None of this is sudden. Over two decades we have been dismantling a public system for sorting cultural value while the platforms substituted their own, relocating value from culture itself to the attention that “content” could attract. The nonprofit model that has supported American arts culture since the 1960s was already failing, and after COVID the failure turned systemic. AI arrives on top of it, not the cause of the collapse but an accelerant.

The trust layer described in this series’ first essay is in the hospital. The exchange layer described in the second is being written by corporate actors largely without artists in the room. But neither functions without the cultural ecology they sit inside, the system we have paid least attention to and are disrupting the fastest.

This essay is about that ecology, and I want to argue something more specific than that it is under strain. Most of the AI-and-culture policy conversation may be aimed at the wrong problems. Not because the people in it are foolish, but because the instruments we have for seeing this transition were built for a different kind of harm, and they are showing us a picture that may be confidently, expensively wrong.

#What the ecology did

For most of the twentieth century, creative culture in democratic countries ran on an imperfectly connected system. Arts education seeded the pipeline. Guilds and unions anchored mid-career labor. Critics and journalists mediated between creators and audiences. Funders sustained what the market wouldn’t. Venues and publishers did the middleware work. Libraries, archives and public broadcasters supported the commons.

None of it was designed. It accumulated through political fights, and little of it was equitably distributed. But as a system it shared one assumption: culture was something to be cultivated, not merely produced. A creative field needs training grounds, evaluative judgment, public goods and patience. It cannot run on output alone.

Here is the part that makes these institutions so hard to defend now. They were never purely expressions of cultural values. They were solutions to resource problems — the most efficient way, in their era, to do expensive things nobody could do alone: train a generation, underwrite work the market wouldn’t, distribute it, keep a record. Reaching an audience was hard. Funding production was hard. Remembering was hard.

None of that is hard anymore. Everyone carries distribution in their pocket. And so the efficiency argument that built these institutions has evaporated underneath them, which means their defenders are reduced to arguing from sentiment at exactly the moment they need to argue from function.

That is the trap, and most of the sector is in it. The question is not whether to save the institutions. It is which of their functions turn out to have been essential, and that is a question largely overlooked because we are still arguing about the buildings.

Four functions determine whether a creative culture can sustain itself.

The pipeline. How new creators enter a field and develop mastery. It was always uneven, too dependent on class and geography, but it existed.

The middleware. The institutions between creators and audiences — critics, venues, platforms, publishers, galleries, editors, curators — doing active sorting and selection, deciding what gets seen, what gets taken seriously, and what gets connected to what.

The mediating institutions. The layer between creative work and civic life: journalism, criticism, universities, foundations, public funders. They tell a society what art it is making and what to remember, and they fund what the market won’t.

The commons. Public arts funding, libraries, archives, public broadcasting, the public domain. And now the training-data commons itself — the accumulated inheritance of everything ever made, converting into privately controlled assets at high velocity.

These aren’t separate. Fewer mid-career artists means less material for critics to sort, which erodes the mediating layer, which weakens the civic case for the commons. Every loop runs both ways. A culture is not the sum of its works. It is the conditions under which works get made, encountered, interpreted and passed on.

#The pipeline problem

The conventional framing treats AI displacement as a labor-market problem. Jobs lost, jobs gained, with the net effect to be determined. That framing isn’t wrong so much as beside the point. It’s a crude scorecard. And the argument it produces — will AI destroy creative work or democratize it? — is an empty binary.

Creative mastery on the way to becoming an artist is not produced, it accrues. A working illustrator doesn’t become a master by declining the median-competent work and waiting to produce masterpieces. She becomes a master by doing a decade of book covers, editorial spots, advertising and storyboards that keep her employed while her visual judgment deepens through thousands of hours of applied practice. That is the pipeline, not the art-school program, but the years of paid work during which mastery accumulates as a byproduct of earning a living.

AI is most capable of replacing that work. You can watch companies buy the capability: Netflix paid $587 million for Ben Affleck’s AI filmmaking company whose tools handle missing shots, background replacement and lighting correction — the tasks that have traditionally constituted the first decade of a post-production career.

So the visible damage so far is illustrators losing wages. The invisible damage is to the development of the next generation of masters, and here is the thing that should worry policymakers most: that harm is difficult to detect as it’s happening.

Consider what would have to happen for anyone to measure it. You would need to count the masters who didn’t happen. Not the jobs lost, or at least those we can eventually count, but the careers that never bloomed, the judgment that never deepened, the twenty-year accumulation of skill that didn’t occur because the ten-year foundation wasn’t there. There’s no statistical data for non-formation, and no unemployment claim filed by a person who never entered the field.

Which is why the struggle to document employment effects is stuck and contested, and why being stuck is itself the finding.

The macro case is contradictory and confusing. The Stanford study figure of employment is a relative decline, its clearest cases are software developers rather than creative occupations, and the authors have since narrowed the finding. The best standing national tracker still finds no occupational shift that clearly aligns with AI. The sector surveys point in a different direction, and the best freelance-market study cuts across the entry-level story entirely: seniority protected nobody, and top freelancers were hit hardest. The two best datasets disagree about who is most exposed — payroll data says the entry level, freelance data says the top.

Then there is New York. The state added a field to its layoff-notice system asking employers whether automation had contributed to the layoff. In the first year that more than 160 employers filed, Amazon and Goldman Sachs among them, not one attributed a layoff to AI.

But every caveat goes in the same direction. Nobody knows whether answering is mandatory. The AI category is undefined. The statute reaches only mass layoffs at large employers, so it can’t see the freelance and commission attrition that is creative-sector displacement. The government built an instrument to detect AI-driven job loss, made it voluntary and undefined, added it onto a statute written in the 1970s for plant closings, and got a year of zeroes. Not much help.

Put these three facts together and they look like something different. The official measurement is empty. The academic instruments were built to see national occupational shifts, not a small, freelance, code-scattered workforce. And the only parties with telemetry fine-grained enough to see what is actually happening to working artists are the algorithmically driven platforms themselves, who are under no obligation to publish it and have every reason not to.

In a vacuum, absence of evidence becomes evidence of absence, and the party that benefits from the finding owns the only data that could overturn it. That is not a gap in the literature, but a structural feature of how this labor transition will be judged, and it means the burden of proof is up to the people least able to count it.

The second half of the problem is that pipeline damage is path-dependent. The NEA can be dismantled in one budget cycle; the mentorship networks and funding relationships it supported cannot be rebuilt in one. Local newsrooms, once closed, do not reopen. Career illustrators who leave for gig work do not come back when the policy environment stabilizes. So we have another harm we cannot measure, on a timescale that exceeds any electoral cycle, and that becomes irreversible before it becomes visible.

Policy systems are very bad at this kind of problem. They’re built to respond to events.

The counterargument to all this is that in disruptive change, systems re-route, and many times this has resulted in a liberation rather than a loss. Photography didn’t end painting; it freed it. An orchestrator spends years learning the mechanics of notation before she can get the sound in her head onto the page. Move notation into a tool and she can spend those years on musical ideas instead. That isn’t a smaller art. Every real leap in toolset has relocated where skill lives, pushing the mechanics down into the instrument and handing more people with something to say the means to say it. The aesthetic history of art, in many ways, is a history of technology innovation — better horn valves lead to revolution in orchestral sound, the printing press revolutionizes who can be a writer, YouTube creates a generation of video creators.

We should take that seriously, and in some corners of some art forms it will prove true. But this case is different in a way the photography parallel obscures. The earlier tools relocated the mechanics while leaving the human inside a living lineage. Notation software still encodes real musical logic that a composer directs, and he is still extending a tradition he belongs to. AI is trained on the accumulated heritage and skill-lineage of human work, and when it recombines that material it isn’t building on the lineage so much as stripping it away, producing something that sounds like it belongs to a tradition and looks like it belongs but doesn’t.

It impersonates mastery without accruing the lineage that the next generation might build on. So the real question isn’t whether the pipeline re-routes, it’s what gets relocated into the tool when it does: the craft, which a culture can stand on, or only the appearance of it, which it probably can’t.

And the costs are asymmetric. An attribution mechanism built five years late is a missed opportunity. A pipeline collapsed in the same five years changes what kind of culture a society can produce for the next thirty.

#A middleware reframe

Here is a place where the Observatory data overturns conventional wisdom.

The AI slop flood is real. Deezer gets roughly 90,000 fully AI-generated tracks uploaded a day, more than half of everything uploaded. The natural inference is that audiences are being deceived, and that synthetic music is passing itself off as human and displacing it.

That inference turns out to be wrong, and two independent research groups recently reported it within six weeks, in two different media. In music on Spotify, AI is about 5 percent of the catalogue and barely 1 percent of the algorithmic recommendation graph. In self-published books on Amazon, titles with substantial AI text are a fifth of what’s released and an eighth of what sells. Deezer’s own figure confirms it: AI tracks are more than half of what arrives and between 1 and 3 percent of what gets heard.

AI content converts to consumption at a fraction of the human rate. Audiences are not being fooled. They are, if anything, quietly sorting. So the harm isn’t deception, but something more insidious. It’s dilution — a slowly growing revenue pool split across a catalogue growing five times faster than demand. Nobody has to be fooled for a musician’s income on a platform to fall. The pie just has to be split in ever-tinier slices.

And that should logically detonate a large part of the current policy agenda. Labeling schemes, watermarking, provenance standards, disclosure mandates — the bulk of what has been proposed and the entirety of what platforms have volunteered — are mechanisms designed for preventing deception. Suppose they all worked perfectly. Suppose every AI track on every service carried an unremovable, accurate, audited tag. The dilution would still be completely unchanged, because it was never really a problem of misidentification in the first place.

That’s not to dismiss disclosure. Provenance matters for other reasons, and the trust layer needs it. But a policy that reduces catalogue volume without changing the consumption ratio has not addressed this harm, and a policy that labels perfectly without changing either has not touched it. That is a test almost nothing currently being proposed passes.

There is a second finding in the same data, and it cuts the other way. AI music is four times underrepresented in the recommendation graphs relative to its catalogue share. Something is filtering: the algorithms. The middleware that sorts and directs listeners is doing part of its job, but opaquely and unaccountably. We have outsourced a core cultural function to systems whose operators could reverse that filtering tomorrow, for commercial reasons, and tell no one.

Which is the deeper middleware problem, and it isn’t volume at all. A curated space forces a contextualizing relationship with the work; its sequencing and contexts are deliberate acts of interpretation. Algorithmic middleware does the opposite. The threat isn’t that it can’t authenticate what’s offered. It’s that its architecture is hostile to the contextualizing work middleware exists to do, and meanwhile the organic discovery layer itself is disappearing. As more people rely on AI answers in their searches, they stop clicking through to complete the search themselves. Google’s organic traffic to publishers is down roughly a third in a year, a disaster for a traditional web business model where clicks and views are the basis for monetizing content. And the latest web survey reports that 53 percent of all web traffic now is not human, but bots. The internet is becoming machines talking to machines.

Supply is exploding, while demand is routing around the institutions that used to organize it. These are the same dynamic.

#The Substitution Problem

The first essay described one institutional failure mode: disruption. The Copyright Office released its most consequential AI report, and the Register of Copyrights was removed the following day. She sued, was reinstated, and remains Register today. The institution held. What didn’t hold is the work: the report is still, fifteen months later, labeled a pre-publication version on the Office’s own website.

The second essay described a different failure: capture. The labels moved from suing AI companies to licensing catalogs to them, and the musicians’ union has now sued the labels — not the AI companies — alleging they took the money and licensed the catalog without compensating the session players. Policy is being written by private contract, at a speed no public institution could match, and now contested by the performers it was written over.

The ecology layer has a third failure mode, and it is the one that might matter most. Substitution.

The nonprofit structure that has underpinned American culture since the 1960s is failing at the system level. Here is what the failure looks like in one institution. Last season, the Minnesota Orchestra earned record high revenue, had an all-time-high annual fund, and eighty-two percent paid capacity — a return, its board chair said, to pre-pandemic attendance. The deficit grew anyway. The fifth in six years.

Then there’s Broadway, which can no longer survive selling out. Cats: The Jellicle Ball won three Tony Awards, played to better than 93 percent of capacity, and closed four months after opening because a million dollars a week in ticket sales couldn’t service its capitalization. Death Becomes Her grossed $102.5 million over twenty months at 92 percent capacity and has never announced break even. Roughly seven percent of the musicals that have opened since the pandemic have paid back their investors. A full house is no longer a business model. With numbers like these, the whole business isn’t even really a business model anymore.

Audiences came back. The business model didn’t. The structural failures — rising costs, philanthropy that can’t keep up, a growing gap between what can be charged for tickets versus what it costs to produce — these aren’t AI-caused. But they produce a crisis at precisely the time the question of how cultural value gets created and rewarded has been reopened — by technology companies. That is the substitution. Not philanthropy replacing public funding, but a whole logic of cultural value, slowly and civically determined, replaced by one that is fast, private and platform-determined.

The means by which that happens is not licensing deals or whatever AI guardrails are put in place. It is the platforms’ internal architectures and rules. Spotify, for example, says that a track must now clear a thousand streams in twelve months to enter the royalty pool at all, plus an undisclosed minimum number of unique listeners. By simple decree, a projected billion dollars over five years is redirected.

We should read that as a governance fact rather than a business one. An unpublished, unappealable rule determining which recordings get paid, set by a company, applying to everyone. No rate court. No consent decree. No notice-and-comment. No standing for anyone to contest it.

We have a century of legal machinery for exactly this problem — consent decrees, rate courts, statutory licenses, compulsory arbitration — all of it built on the premise that when a private actor’s decisions govern an entire market, the decision needs a forum of adjudication. That machinery now governs a shrinking share of how musicians actually get paid, while the decisions that govern most of it are made in product meetings.

The substitution isn’t that private actors are making cultural policy, because they always have. It’s that policy has migrated from mechanisms that can be appealed to architectures that cannot. A bad statute can be amended. A bad consent decree can be revisited. A recommendation weighting cannot even be seen.

There is one counter-case in everything this project has tracked, and it is not American. In July 2026 the Munich Regional Court held that the AI music generator platform Suno had infringed copyright in six German compositions, granting GEMA’s (GEMA is the German equivalent of BMI or ASCAP in the US) claims for an injunction, disclosure and damages in substantial part. Suno is weighing an appeal.

The copyright holding belongs to the second essay. What belongs here is that the German composers had somewhere to bring it. GEMA is a collecting society, and under its deed its members assign it their reproduction rights — GEMA holds them als Treuhänderin, as trustee. So a German composer whose song is memorized in a model has no claim to bring, because the institution (GEMA) has it. Collective administration is simultaneously why the claim was enforceable and why no individual could have enforced it.

Out of everything in this discussion, it is the only instance of a rule about AI and music being made somewhere a creator organization had the standing to carry it.

#The functional lens

Both previous essays argue for organizing this by function rather than by issue or jurisdiction. At the ecology layer that produces an additional conclusion.

Start with convergence: The illustrator losing early-career work, the newsroom laying off its arts critics, the arts council rebranding itself as an economic-development office, the regional theater going dark. Conventional tracking treats these as distinct stories in distinct fields. But they are the same ecological failure at different functions, failing in a coordinated way not because anyone is coordinating it, but because they share underlying conditions.

Then look across borders, and the pattern becomes ever starker. Ireland pays two thousand artists a weekly income. European law requires member states to guarantee stable funding for public service media. The UK’s AI adoption plan commits government to work “in lockstep with trade unions and guilds to ensure that workers are given a formal seat.” Germany kept a statutory collecting society with the standing and the balance sheet to take an American AI company to court and used it. Meanwhile, the United States has a second consecutive proposal to eliminate the NEA, the NEH and IMLS, and a Corporation for Public Broadcasting that no longer exists.

Almost none of that divergence is explained by AI policy. Not one of those differences between jurisdictions is the product of a good or bad decision about artificial intelligence. They are the product of decisions taken decades ago about which cultural institutions to build and keep. Germany didn’t out-think anyone on AI, it had a collecting society with resources and teeth. Ireland didn’t design a clever AI intervention; it had a working arts agency capable of running a national experiment and evaluating it.

So perhaps the AI policy conversation has its fundamental terms backwards. We talk as though jurisdictions will diverge because some write better AI rules. The evidence says they have already diverged, on the basis of institutional inheritance, and that AI rules will mostly be routed through whatever machinery a country still has. A country’s AI policy is downstream of its institutional infrastructure built to support culture. You cannot legislate a remedy if you have no body capable of administering it.

That reframing makes several current fights look smaller than they are being played, and it makes one domestic failure look much larger: have American national arts service organizations adopted official AI policies or frameworks? Arts Council England, by contrast, bars AI from assessment decisions, publishes a toolkit so arts organizations can write their own policies, runs a leadership academy, and has adopted digital arts as its tenth supported artform.

The gap isn’t ideas. It’s that one of them is an institution with the administrative standing and capacity to hold a position, and the others are membership bodies that would have to ask permission to have one.

Arts philanthropy is not absent on AI: a half-billion-dollar collaborative of ten major foundations now names humanities and culture among its priorities. What hasn’t moved is the guidelines and rules.

Which brings the tiering problem into focus one last time. Screen actors have contract language. The Writers Guild has provisions in its 2026 agreement: what the contract requires is written notice if companies license writers’ work for AI training, after which the Guild may request a discussion. No consent, no payment, no vested interest.

Everyone else — illustrators, freelance writers, translators, indie musicians, session players, photographers — has no guild, no training fund, no institutional pathway back into the field. For them the ecology is the infrastructure, and the ecology is being dismantled.

#What is actually at stake

Capital expenditure on AI infrastructure this year is running north of half a trillion dollars. That investment rests almost entirely on the accumulated creative record of humanity, which makes the question a large one: who should own the value of that inheritance, and who should realize it?

Right now it is being settled by default, in favor of whoever scraped the data first.

The answer can’t be to defend the system we have. It is hard to defend a model that left most working artists outside its protections, concentrated reward at the top, and priced culture by clicks. The more interesting question is whether this renegotiation is a chance to build something better than the traffic-based value machine the platforms handed us. If AI’s economics rest on the common creative record, is there a public premium, a share of value flowing back to the commons it came from, or to the creators whose work taught the machines?

The Observatory is tracking the first crude attempts. Te Hiku Media’s Kaitiakitanga Licence, over the te reo Māori speech corpus (an Indigenous data sovereignty legal instrument rooted in Māori guardianship — kaitiakitanga — that ensures te reo Māori speech data and the AI speech models that are derived from it are used for the benefit of Māori people), is the only instrument we’ve found that writes the return into the license itself: royalties on commercial profits shared back with the communities the data came from. It was drafted by a community broadcaster in the Far North of New Zealand, doing something that is unbuilt in the wider discussion about value derived from AI training.

But the obstacle to figuring this out on a wider scale is substantial. Owed back to whom, and collected by what? A public return presupposes a public body to receive, steward and distribute it, and we have documented the dismantling of exactly those bodies that might do it. That is the paradox at the center of this moment: the case for a public premium from AI arrives just as the mechanisms and institutions that might be capable of doing it are being scrapped. And the players still standing are mostly private, so routing a premium through the platforms stops it from being a public return at all.

Which is why the American debate over AI policy may be arguing about the wrong half of the problem.

We have spent three years of policy discussion on ownership — whose work, what license, opt in or opt out, whether training is fair use. Munich is a useful corrective, because the German composers won on an ordinary property right, the same one an American songwriter already has. What they also had was somewhere for that right to live: an institution owning it, funding the case, choosing which songs to bring, and able to turn the repertoire into a licensable dataset.

Ownership settles who is wronged but it doesn’t settle who can act. A right nobody can afford to exercise, over an opus nobody can price alone, is not much of a right, and that is the position most working creators are actually in, whatever the statute says.

So the question moves from should creators own this to what would administer it if they did. That is an institutional question. It is boring, and it is unglamorous, and it will likely ultimately determine everything.

Which loops back to where we started. We built these institutions because concentrated institutions were once the most efficient way to fund, make, distribute and remember culture. They are failing partly because that is no longer true. So the question was never how to save them. It is which of their functions are still essential — building new masters, underwriting non-commercial work, preserving the record, telling a society what it’s making, holding rights on behalf of people too dispersed to hold them individually — and what should perform those functions now.

So far that is being decided by default, mostly by accident, by companies optimizing for something other than culture or the public good.

#The ecology layer

Disruption, capture, substitution: three failure modes, with one underlying problem. The rules for the AI creative economy are being written without the civic process you might hope would shape them, and without the people whose livelihoods depend on the outcome.

This is the last of three essays mapping how AI is restructuring the systems that govern creative work. Attribution and accountability is the trust layer — what identifies who made what, and whether we can verify any of it. Value and exchange is the exchange layer, which moves compensation across identified contributions. Cultural infrastructure and ecology is the ecology layer: the conditions under which trust and exchange either sustain a living creative culture or don’t.

The three depend on each other, and the dependency runs in both directions. You cannot pay for what you cannot identify. You cannot sustain a creative field without mechanisms that move value back to the people producing it. And you cannot produce a next generation of work worth attributing or exchanging if the conditions that form creators have gone.

The uncomfortable point is that attribution and exchange can both be solved correctly and the ecology can still fail. A perfect disclosure regime and a fair licensing system are necessary and insufficient. If the pipeline collapses, the mediating institutions are dismantled, and the cultural record converts into private capacity whose value never returns to the public, then twenty years from now there will be an elegant framework for identifying and paying for creative work, and much less creative work to identify and pay for.

That is why this project organizes the question by function. The three layers are the same system seen from three angles, and they rise or fall together. A policy conversation that treats them as separate debates — one for copyright lawyers, one for labor economists, one for arts-policy people — will build a framework that fails at the seams, and it will fail in the direction of whoever is best organized to exploit the gaps.

So the question the next decade has to answer is not whether AI belongs in creative life. It does, and it will, whether anyone licenses it or not. The questions worth fighting over are whether the conditions for creative life survive the transition, who realizes the value of the inheritance the whole system runs on, and whether we still have institutions capable of asking either question on the public’s behalf.

That is a decision, and it is being made right now: in who funds what, who dismantles what, who builds what, and who is in the room when the value of everything we have ever made gets priced.

The trust layer is what we build before we decide whom to believe. The exchange layer is what we build before we decide who gets paid. The ecology layer is what we build before we decide what kind of culture we can still produce — and whether anyone is left who could notice if we couldn’t.

View Sources