The Content Factory

The factory is already running

The simplest way to see what artificial intelligence does to popular culture is to stop forecasting and start counting. In April 2025, the streaming service Deezer estimated that more than twenty thousand fully AI-created tracks were arriving on its platform every day — roughly eighteen percent of new uploads, and nearly double the figure from January of the same year. Over the preceding twelve months, Spotify removed around seventy-five million tracks it classified as spam, a volume that became economical to generate only once the marginal cost of producing a song approached zero. An AI act calling itself the Velvet Sundown gathered more than a million streams before observers worked out that its folk-rock was assembled on the music generator Suno and that the band did not exist.

The pattern is not confined to music. By the middle of 2025, the monitoring organization NewsGuard had identified 1,271 bot-written sites operating across sixteen languages, each dressed as a news outlet and each producing auto-generated articles on politics, entertainment, and health. An analysis by the search firm Ahrefs of nine hundred thousand new web pages, conducted in April 2025, estimated that roughly seventy-four percent contained AI-generated content. At the smaller end of the same phenomenon sits a documented Pinterest operator posting around eighty AI-made pins a day — fake images fronting fake blogs, run entirely for advertising revenue.

There is even an early version of the most cinematic worry: a culture industry that tests ideas on an audience and manufactures more of whatever performs. In March 2025, the games publisher Activision ran advertisements for products that did not exist, including a fabricated “Guitar Hero Mobile,” and later described the campaign as market research. That is the pilot-testing loop with the content layer itself synthesized — generate a candidate, watch the response, decide what to build next, except the candidate was never built at all.

Why the protections collapse here

The survey’s opening chapter used Goethe’s Faust as a boundary stone. Three things, it argued, protect a work like Faust from automation: there is no objective function for greatness, the value lives in the statistical tails rather than the center, and audiences care about provenance — about who actually lived the work. Popular culture is interesting precisely because all three protections invert, and the first inversion is the one that matters most.

A machine-learning system improves fastest when it has a verifier: a cheap, fast, reliable signal telling it whether an output is better or worse. A chess engine has the game’s result; a coding model has a test that either passes or fails. Great literature has no such signal, which is why imitation — predicting the most probable next sentence — is roughly the ceiling a model can reach on it. Popular culture, by contrast, has the best verifier money can buy. Engagement is dense, immediate, and quantifiable: streams, completion rate, watch time, retention. Each of these is a number that arrives within hours and can be fed straight back into the next round of production.

Great literature has no cheap verifier; popular culture has the best verifier money can buy.

That difference is decisive because it is exactly the condition under which a particular training method, reinforcement learning, begins to outperform mere imitation. Reinforcement learning is the practice of letting a system try things and rewarding the attempts that score well, rather than teaching it to copy examples. It is how a program learns to exceed human play in a game: not by imitating recorded matches but by optimizing against the score directly. Where a clean score exists, the generate-measure-reinforce loop closes, and the system can be pushed past what any individual demonstrator showed it.

The second protection inverts just as cleanly. Faust lives in the tails of the distribution — it is valued for breaking inherited patterns in a way that turns out to be productive. Popular culture is largely valued for the opposite: for hitting the center of a distribution competently. Comfort, familiar genre beats, the next very-similar thing. That is precisely what a model built to predict the most probable continuation is good at, and it is the reason an automated culture factory is structurally plausible where an automated Faust is not.

Provenance, the third protection, scales with disposability. Almost nobody asks who composed the lo-fi instrumental they study to, or the library music laid under a video clip. Functional and ambient content carries essentially no authorship premium, which is why music and search-optimized information sites were the first to fall. So the structural answer to whether an AI-automated streaming service or topic site is possible is yes — for everything disposable, formulaic, and cheap to render.

Three ways the loop bends

The clean version of that thesis is too clean. The optimization loop has three failure modes, and each one bends the outcome away from “automation simply wins.”

The first is Goodhart’s law. Engagement is a proxy for value — a stand-in that correlates with it under normal conditions — and optimizing the proxy hard enough corrupts the thing it was standing in for. Pure engagement-maximization drifts toward the addictive rather than the good. One study found that twenty-one percent of recommendations served to new accounts on YouTube were AI slop, with a further thirty-three percent classed as “brainrot” — repetitive content engineered to trigger a small dopamine response and nothing more. Audiences habituate to this, and an early discount is already visible: engagement with AI-generated articles fell about forty percent over 2024, and human-generated content earned roughly 5.44 times more traffic. Rather than collapsing, demand for human freelance writers, designers, and editors surged through 2025. A service that optimizes a corrupted proxy does not converge on quality. It converges on a treadmill its audience slowly learns to taste.

The second failure mode is that the loop is structurally backward-looking. Reinforcement learning on engagement is interpolation over what has already worked. It can run a known vein dry with great efficiency — the ten-thousandth competent isekai, the next cozy mystery — but the things that create genuinely new audiences are not interpolations. The first cinematic universe, the true-crime-podcast boom, K-pop as a global form: these were distribution shifts, tail events that the existing data could not have predicted. The factory is therefore excellent at exploitation and nearly blind to exploration, which keeps it dependent on humans, or on luck, to seed each new genre. It is a strip-miner, not a prospector.

The third failure mode is commons-poisoning. A model improves by training on data; flood that data with synthetic output and future models trained on the polluted substrate degrade. Researchers call this model collapse — the slow corruption of the very material that made the trick work in the first place. Commentators have begun describing it as a corruption of the internet’s shared knowledge base. The strategy partly eats its own seed corn.

What falls, and what does not

Put the three failure modes together and the likely outcome is not that an automated service replaces the established ones. It is a bifurcation — a splitting into two layers. At the bottom, an effectively infinite ocean of cheap synthetic ambient content, where provenance is irrelevant and the proxy is all that matters. Above it, a premium tier in which verified-human, live, and scene-based work becomes both a luxury and an identity good — valuable because the synthetic layer is free. Provenance does not vanish from the picture. It concentrates upmarket and turns into a status signal: vinyl, live shows, “made by a real person” labels. Scarcity migrates from the artifact to its authentic origin.

The order in which categories give way follows a rough variable: disposability multiplied by low production cost multiplied by low provenance expectation. The higher that product, the sooner the category falls.

Even in the resistant categories, though, the development layer is already algorithmic. Streaming services have long A/B-tested thumbnails and let data shape which shows get made. The genuinely new step is not choosing among human-made options; it is generating the options.

The economics underneath

Two older lenses make the shape of this clearer. The first is Coasean, after the economist Ronald Coase, who asked why firms exist at all and answered that they exist to coordinate activity that would be too costly to organize through one-off market transactions. A studio or a record label is, in this view, a coordination machine for scarce creative labor. When generation approaches zero marginal cost, that coordination rationale erodes, and production disaggregates. But a new scarcity — trust, curation, verified authenticity, attention itself — becomes the thing firms reorganize around. The firm does not disappear; the scarcity it manages moves.

The second lens comes from lean manufacturing, where muda is the Japanese term for waste — any activity that consumes resources without adding value the customer would pay for. An automated content factory looks, at first, like the ultimate elimination of waste in production: output at almost no cost. But it optimizes the flow of a proxy metric rather than of value, which is the classic lean trap — maximizing the measurable instead of the valuable, achieving spectacular local efficiency while destroying value at the level of the whole system. The slop flood is muda dressed as throughput.

The endpoint is not a fully automated culture industry but a sharply stratified one.

The evidence so far supports the mechanism and the direction without ambiguity: where a dense engagement signal meets a disposable artifact, automation is already winning. What it does not support is the tidy endpoint of a culture industry handed over wholesale to machines. The shape the data actually points toward is a stratified one — a synthetic floor optimized to a corrupted proxy, and a human ceiling whose value rises precisely because the floor exists.