Why the internet is filling with AI slop

Because it became free to make, and nothing downstream was designed for that.

That is the whole mechanism. Everything else is detail.

The number

Graphite analysed English-language articles sampled from Common Crawl and tracked the AI-generated share month by month. It went from 2.2% in January 2020 to 51.7% in May 2025, passing human-written output for the first time around November 2024.

And then it stopped climbing. Since early 2025 the share has plateaued at roughly half. Coverage of the study, Axios among others, made a point of that plateau, and it is worth repeating: the honest headline is “about half,” not “the internet has been taken over.”

Half is quite enough to change how everything feels.

Why the economics break

Publishing has always been throttled by the cost of production. A page took someone an hour. That single fact did most of the quality filtering on the internet, invisibly, for thirty years, not because writers were virtuous, but because nobody bothers to write a hundred worthless pages by hand.

Remove the cost and the filter goes with it. Three things follow immediately:

Any positive expected return justifies infinite volume. If a page costs effectively nothing and has a one-in-ten-thousand chance of earning a few cents, the rational move is to make a million pages. This is not a loophole. It is arithmetic.

Ranking signals become gameable at scale. Search and feed algorithms rank by proxies, length, structure, keyword coverage, freshness. Those proxies worked because producing them used to require effort. They are now trivially satisfiable.

Speed beats quality in a feed. A recommendation system optimising for watch time does not have a quality input. It has an engagement input. Slop that engages outperforms good work that does not, and there is no term in the equation that notices.

The four engines

Ad arbitrage. Generate pages, rank for something, serve programmatic ads. Margins are thin per page and the model works entirely on volume.

Engagement farming. Accounts posting generated images tuned to provoke reaction, monetised through platform payouts or eventual account sale.

Marketplace flooding. Listing costs nothing on most platforms. Generated books, templates, prints and courses cost nothing to produce. A one-in-a-thousand sale rate is a viable business.

Ordinary corner-cutting. The largest category and the least discussed. Not spam, real businesses and real publishers, under real deadline pressure, shipping unread output. The Chicago Sun-Times reading list with ten invented books reached print this way: a legitimate supply chain where the checking step quietly stopped happening.

Why it will not simply be filtered away

Detection does not work reliably, the Stanford research on false positives, and OpenAI’s withdrawal of its own classifier, are the short version. So platforms cannot cleanly identify slop even where they want to.

More fundamentally, the incentives do not point that way. Platforms are paid for engagement, not for accuracy, and slop engages. There is no revenue line that improves when it is removed.

What is actually changing

The plateau is the interesting signal. It suggests something is pushing back, search updates, platform policy, ranking changes, or simply the ceiling on how much of the web can be spam before the arbitrage stops paying.

What follows is a split, already visible: verified, provenance-bearing, human-attributed content on one side, and an undifferentiated ocean on the other. Curation becomes the scarce good, because the abundant thing has no value and the scarce thing is knowing what is worth reading.

That is not a new equilibrium. It is the one publishing had before search engines convinced everyone that finding things was solved.