The Economic Models That Failed

The record label’s core business model has not changed fundamentally since the 1950s: find artists with potential, invest in recording and marketing, own the recordings, collect royalties. The internet disrupted the distribution piece. AI is disrupting the content piece. The record label survived the internet disruption because its value in marketing and artist development was not eliminated by digital distribution. The question for the AI era is whether any of its remaining functions are irreplaceable.

That is a specific question about a specific industry. The broader pattern — which creative business models have failed, which have held, and why — reveals something about the economics of AI disruption that the industry-by-industry coverage often misses.

The Commodity Middleman Problem

The economic models that have failed most completely share a structural characteristic: they intermediated between creators and buyers in markets where the product being sold was relatively undifferentiated and where AI can produce adequate substitutes at near-zero marginal cost.

Stock photography agencies fit this exactly. They aggregated undifferentiated images (businessperson at desk is not differentiated from another businessperson at desk), marked them up, and licensed them to buyers who needed image content. AI broke this by creating an alternative supply of undifferentiated images at near-zero cost. The stock agency’s value proposition — I have a lot of images you can use — evaporated when the buyer could generate a lot of images directly.

Production music libraries — the companies that licensed tracks to video producers, game developers, and advertisers — have the same structural problem. Background music is undifferentiated in the same way. The buyer needed music that felt appropriate to context; they did not need music from a specific composer. AI generates context-appropriate music adequately for most commercial purposes. The production library’s value proposition collapsed at approximately the same speed as stock photography.

Content marketing writing — the agencies and freelancers that produced blog posts, product descriptions, and SEO content for businesses — has experienced comparable collapse. Generic content at volume was never a creative-integrity market; it was a market for adequate text produced at scale. LLMs produce adequate text at scale and do so more cheaply than any agency model could match.

The Differentiated Premium Model That Held

The economic models that held are almost perfectly inverse: they sell differentiated, specific, high-trust relationships or outputs that AI cannot adequately substitute.

Fine art photography has held because the value proposition is documentary specificity and artist identity, not generic image adequacy. A photograph by Gregory Crewdson sells because it is a Gregory Crewdson photograph — the documented output of a specific artist with a specific vision and a specific cultural positioning. AI cannot produce Gregory Crewdson photographs. It can produce images that superficially resemble them; they are not the same thing, and the market for them is not the same market.

Concert performance has held for the reasons discussed elsewhere: the value is social and temporal, the scarcity is real, the authenticity is documentable. You can verify that you were in the room.

Custom commissioned art — a portrait of a specific person, a mural for a specific building, an illustration for a specific book — has held because the specificity requirement eliminates AI substitutability. AI cannot paint a portrait of your grandmother unless you have extensive reference imagery, and even then the result lacks the relationship that a portrait commission represents. The commission is partly about the relationship between artist and subject, not only about the physical output.

Bespoke editorial illustration — the kind of conceptually sophisticated illustration that runs in The New Yorker or is commissioned for the cover of a serious nonfiction book — has held because the value is conceptual and identifiable. An illustrator who can think visually about complex ideas, in a distinctive style that readers recognize and editors seek out, provides something that prompt-engineering into a generative model doesn’t produce reliably enough for that context. Whether this remains true as image generation improves is the open question.

The Models That Are Adapting (Slowly)

Between the failed and the resilient models, there is a population of economic structures that are under pressure but not collapsed — adapting, with varying speed and effectiveness.

Music publishing, which owns compositions (not recordings) and collects royalties when those compositions are performed or covered, is under moderate pressure rather than existential pressure. AI can generate new compositions, but it cannot claim copyright in them under current US law — and therefore cannot displace the royalty stream from existing human-written compositions. The risk is longer-term: as the supply of AI-generated compositions grows, the market for licensing existing human compositions may contract. But the existing catalog’s value is protected by copyright for another generation.

Commercial photography, for clients who want specific human subjects photographed in specific ways, is adapting by moving upmarket. The low end (product shots that can be generated against white backgrounds, simple lifestyle imagery) is largely gone. The mid and high end (brand campaign photography, luxury advertising, anything requiring real people in real environments with real lighting) is intact. The adaptation required abandoning the low end rather than trying to compete in it.

Graphic design has bifurcated. The design of visual systems — brand identities, user interface systems, typographic hierarchies — requires human judgment about relationships, consistency, and strategic communication that AI handles poorly. The execution of design work at the production level — generating variations on an established template, creating assets that fit within an existing design system — is largely AI-handled in studios that have adapted their workflows. The designers who remain employed are doing more strategic work and less production work.

What the Pattern Reveals

The common thread in the surviving models is not “high quality” — quality alone is not protective. It is specificity. Specific person, specific place, specific relationship, specific concept. These specifics cannot be generated because they require the existence of the specific thing that they are about.

Stock photography was not low quality; it was unspecific. A technically excellent photograph of a generic business meeting is still unspecific, which made it substitutable. A technically average photograph of a specific event that happened in a specific place with specific people in it is not substitutable regardless of quality.

The lesson is not reassuring for creative professionals who built their practices on the assumption that quality was protective. Quality helps at the margin. Specificity is structural. The economic models that failed were the ones that tried to succeed with quality alone in markets where specificity was not required by the buyer.

This distinction — specific versus generic, relationship versus output, documented versus generated — was not visible as a strategic consideration before AI, because the distinction did not matter economically. Generic adequate content was expensive enough to produce that it supported business models. Now it is not. The creative professional who wants to know whether their economic model is safe has a straightforward test: is what they produce specific in a way that AI cannot substitute? If not, the model is at risk. If yes, the model is resilient. The clarity of the test does not make it easy to act on.