--- title: "Why Is There So Much AI Slop? What the Evidence Can Show, and What It Can't" url: https://deadwritingtheory.com/blog/why-is-there-so-much-ai-slop/ summary: "The usual answer is that text costs almost nothing to generate and attention still pays, so volume rises. Studies can measure how much machine-written text exists and whether it reaches readers. They cannot measure why people publish it." published: 2026-11-05 updated: 2026-11-05 author: "Jack Stovell" publisher: "Adapt Progress Evolve Limited" language: en-GB --- # Why Is There So Much AI Slop? What the Evidence Can Show, and What It Can't *By Jack Stovell. Published 5 November 2026.* The usual answer is that text costs almost nothing to generate and attention still pays, so volume rises. Studies can measure how much machine-written text exists and whether it reaches readers. They cannot measure why people publish it. Short answer: the cost-and-attention story is plausible, and one measurement fits it. No study on this docket has measured why anyone publishes anything, though. The studies count the output. Everything about motive is inferred from that output. ## Three questions hiding in "why is there so much slop" Ask why there is so much AI slop and you usually get a tidy chain: it is cheap, it pays, so it floods the web, so it harms you. One smooth arc. The top search results on the question mostly give that chain, backed largely by anecdote. That one question is really three: 1. **How much is there?** This is a measurement question. You can answer it with a count. 2. **Why is it made?** This is about motive. No count of articles tells you what someone was thinking when they pressed publish. 3. **Does it matter?** This is about harm. It depends on who reads the text and what they would have read instead. Each question needs a different kind of evidence. Blend them, and a solid answer to the first lends its authority to guesses about the other two. ## "So much" first: what has actually been measured The studies on this docket measure two separate things. Prevalence is the share of text, articles or sites that look machine-written. Convergence is whether writing has grown more alike. This section covers prevalence. Convergence comes later. The DeGenTWeb preprint (He, Ardi, Govindan and Madhyastha, University of Southern California, 2026) classifies whole websites as LLM-dominant, meaning their content was generated with little human input. It asks how many such sites exist and whether they reach people through search. It finds a growing share, and finds those sites turning up in how-to search results. The [case for that study](/cases/the-websites-nobody-wrote/) sets out the figures. It is a preprint, not yet peer reviewed. The Graphite study, "AI Now Writes as Many Online Articles as Humans" (Paredes, Druck, Benson and Smith, 2026), estimates the share of new English articles that are primarily AI-generated. Its most useful finding is the shape of the curve: "since Q1 2025 the percentage of primarily AI-generated articles has plateaued at roughly 50%." The [case page](/cases/half-the-articles-and-a-plateau/) returns a split verdict. The size of the share counts for the prosecution. The stalled rise counts for the defence. Two caveats come with that number. Graphite is an SEO and growth agency, so this is a vendor study and carries a vendor badge. It also rests on commercial AI detectors, which can flag human writing as machine writing. Graphite reports low false-positive rates for its three detectors on articles from before ChatGPT. Detectors do have a record, though: in [one study](/cases/the-detector-that-convicts-the-innocent/), seven of them labelled more than half of a set of human essays by non-native English writers as AI. Read the share as an estimate with a margin of error. It is not a census. Then there is reach. A text existing and a text being read are separate claims. A [second Graphite study](/cases/half-the-articles-but-who-reads-them/) looked at what Google ranks and what ChatGPT and Perplexity cite. Most of the articles they surfaced were human-written. That study is also from a vendor and also uses a detector. Put it next to DeGenTWeb and the measured picture is this: a lot of machine text gets published, some of it reaches searchers, and how much depends on where you look and how. ## The usual explanation: cheap to make, still paid for attention Now the "why". The standard account runs like this. Generating text costs almost nothing. Attention still earns money through ads, affiliate links and recommendation feeds. When the cost of making something falls towards zero and the reward stays, people make more of it. The account hangs together and is probably partly right. A University of Florida news piece (4 March 2026) reports a version of it. The piece covers a study in the Journal of Marketing Research by Tianxin Zou, Zijun Shi and Yue Wu. Zou is quoted saying: "Because the quantity is so large, it congests the recommendation systems, so it gets harder to encounter the truly high-quality content." Notice what kind of evidence that is. According to the article, the authors used economic modelling. A model tells you what follows if its assumptions hold. It does not count congestion on any real platform. DeGenTWeb comes closest to the money question. It reports that 78.8% of LLM-dominant sites have a clear financial incentive, against 55.8% of other sites. That is a measurement, and it fits the standard story. Look at what it measures, though. It records a feature of the sites, observed after they were built. It does not record why anyone built them. ## Where the explanation is inference, not evidence A study can look at a site and record three things: the text is probably machine-written, the site shows up in search, and it makes money somehow. Each of those is an observation. Saying the owner published it because generation was nearly free and the revenue was positive takes a further step. Often it is a sensible step. It is still a step. Other motives produce the same output. Someone short of time fills a site quickly. Someone tests a tool. A small business wants pages to exist so its site does not look empty. A writer working in a second language gets help with the English. The text can look much the same in every case, and a count cannot tell these motives apart. Plenty of sites make money whoever writes them, which is why DeGenTWeb's comparison with other sites matters more than its headline share. So when an article says "people publish slop because X", check what its evidence measures. Nearly always it measures the text or the site. The writer of the article supplies the intent. ## Volume is not sameness: slop and machine voice are different claims A second mix-up: "there is loads of AI text" and "everything sounds the same now" get treated as one claim. They are two. Fitterer, Gangl and Ulbrich (Technische Universität Berlin, ACL 2025 Student Research Workshop) compared English news articles from 2018 with articles from 2024. LLM-style words rose. Lexical diversity did not fall. The [case page](/cases/the-news-that-didnt-flatten/) scores that as a point for the defence. It is not a knockout, because the study tests only vocabulary and only in edited news. It still shows the two can come apart: more machine-flavoured text does not by itself make writing converge. That gives two separate worries. One is quantity: how much of the web is machine-made, and does it crowd out the rest? The other is variety: do we all sound alike now? "Slop" mostly carries the first, plus a judgement that the text is worth little. "Machine voice" is a claim about style. Treat them as one claim and you can prove the easy one while slipping in the hard one unexamined. The [theory page](/theory/) explains how this site keeps them apart. ## How to test a "why" claim yourself Take any "why" claim and ask: what would I have to observe for this to be false? If the honest answer is "nothing", you are holding a story. It may be a good story. It is not yet a finding. Try it on the cost-and-attention account. Either of these would count against it: - machine text growing just as fast on sites with no ads, affiliate links or other way to earn money; - volume falling while generation stays cheap and attention still pays. You can picture either one turning up, which makes the account worth taking seriously. Nobody has run the test yet. Then put three checks to any explainer: 1. Does its evidence measure the text, or the motive? 2. Does it separate what is published from what is read? 3. Does it separate volume from sameness? Most confident explainers fail at least one. That is no reason to throw out the cheap-and-paid story. It is a reason to file it under explanation and keep the counts in a separate box. ## Sources - [He, Ardi, Govindan and Madhyastha, "DeGenTWeb: A First Look at LLM-dominant Websites", arXiv preprint 2605.00087 (2026)](https://arxiv.org/abs/2605.00087) - [Paredes, Druck, Benson and Smith, "AI Now Writes as Many Online Articles as Humans", Graphite research (2026)](https://graphite.io/five-percent/research/ai-now-writes-as-many-online-articles-as-humans-do) - [Eric Hamilton, "'AI slop' hurts consumers and creators. But high-quality AI could help both.", University of Florida News (4 March 2026)](https://news.ufl.edu/2026/03/ai-slop/) ## Questions people ask ### Why is there so much AI slop online? The usual explanation is that generating text costs almost nothing while attention still earns money, so volume rises. That is a common explanation. Nobody has measured it as a finding. Studies can show how much machine-written text exists and whether it reaches readers. The motive behind it is inferred. ### Is there actually more AI-written content than human-written content? One vendor study, from Graphite, estimates that primarily AI-generated articles make up about half of new English articles, and says the share has plateaued since Q1 2025. It relies on AI detectors, which can mislabel human writing, so treat it as an estimate. See the case page: /cases/half-the-articles-and-a-plateau/. ### Does AI slop reach readers through search, or does it sit unread? Some of it reaches readers. The DeGenTWeb preprint found LLM-dominant sites turning up in how-to search results. A separate Graphite study found that most of the articles Google ranks and ChatGPT and Perplexity cite are human-written. How much gets read depends on where you look. ### Is AI slop the same thing as every website sounding alike? No. Volume and sameness are separate claims. Fitterer, Gangl and Ulbrich found more LLM-style words in English news between 2018 and 2024, but no fall in lexical diversity. A rise in machine text does not on its own make writing converge. ### Will AI slop keep growing? Nobody can say from this evidence. One detector-based estimate of new English articles shows a plateau. DeGenTWeb found the share of LLM-dominant sites growing up to the first half of 2025. Neither is a forecast.