Skip to the content
Dead Writing Theory

The Court of Public Prose

From the clerk's desk

AI Slop in the News: What the Headline Numbers Measure, and What They Don't

Most AI slop news stories lean on three kinds of evidence: a share-of-articles count, an anecdote about one bad item, and a reader's impression. Each answers a different question, and only the first is a measurement of the web.

By Jack Stovell · Published

Short answer: sort the evidence before you weigh it. A count can tell you how much machine text is being published, but it cannot tell you whether that text sounds alike. An example shows that something can happen. An impression shows what one person’s screen looked like. Most slop stories use all three in the same tone, so the job is to find out which one is carrying the weight.

What ‘AI slop news’ usually contains

The coverage has a familiar shape. There is a number near the top, then a vivid example: one rotten page, or one absurd image. Then comes a quote or a paragraph saying the internet feels worse than it used to.

Each part answers a different question. The number tells you how much. The example tells you whether it can happen. The impression tells you how it feels. Only the first is a measurement of the web, and even then it only measures the slice that was sampled.

Type 1: a share-of-articles count (prevalence)

This is the one that looks like science, and sometimes it is. Someone samples a pile of pages, labels each one as human or machine, and reports a share.

The most quoted version comes from Graphite, an SEO agency. The court’s record of its May 2026 study reads: “In Q1 2025, primarily AI-generated articles nearly equalled human-written ones: 49.6% against 50.4%.” The full entry, with its caveats, is at Half the articles, and a plateau. Retellings tend to drop two things. It is a vendor study, and by its own account the share has since stopped rising.

The court’s position is simple. A headline count of AI articles says nothing about whether the writing has converged. Prevalence asks how much and convergence asks how alike, and they are separate questions. A web that is half machine-written could still be varied. A web with very little machine text could still be uniform. Testing sameness needs a different instrument, for example vocabulary diversity compared across years. The news that didn’t flatten is that kind of test. On its narrow measure, it found no flattening in English news.

There is a second gap: published is not the same as read. Half the articles, but who reads them? looks at what search engines and chatbots actually show people. Again, that is a different question from how much gets published.

Type 2: the vivid single example (anecdote)

Then there is the anecdote: one bad item, held up to the light.

The standard example in science is a review that the journal Frontiers in Cell and Developmental Biology retracted in February 2024. In the words of its retraction notice, “concerns were raised regarding the nature of its AI-generated figures.” One of those figures, a rat, became a byword for machine-made nonsense in research. It is memorable, which is why it keeps being retold.

Memorable is the problem. A single glaring failure tells you that failure is possible. It cannot tell you how often it happens, because one case has no denominator. A thousand retellings of the same rat are still one rat.

That doesn’t make anecdotes worthless. They are usually how anyone notices a problem in the first place. The measurement has to come afterwards, though. The scientists got there first is what that looks like: an estimate of language model use across a large body of papers. Notice that it is also a prevalence study, and the case says so. It measures how much, not how alike.

Type 3: the platform or feed impression (opinion and sampling)

The third kind is the hardest to pin down, because it is built from someone’s own scrolling.

“My feed is full of it.” That may well be true. But a feed is a sample the platform chose for you, tuned to whatever you linger on, so you cannot read a rate off it. Click on one strange image and the feed may offer you twenty more. That tells you about the recommender, not about the web.

In short, this kind of evidence is opinion plus a biased sample. As an experience it is real. As a count it is useless. Broadcast segments and columns often lean on it, with a presenter or a viewer saying the web feels off. That can be true for them without measuring anything beyond their own screen.

Why a detector sits underneath most of these headlines

This part usually hides in the methods paragraph, if there is one. To report a share of AI articles, somebody has to decide which articles are AI. Nobody can read tens of thousands of pages by hand, so a detector usually makes that call.

That makes the headline number only as good as the detector behind it. Detectors produce false positives, meaning human writing labelled as machine writing, so part of any detector-based share is a share of mistakes. How large that part is depends on the tool and on the text. The detector that convicts the innocent shows how badly it can go for one group of human writers.

The better studies test their detector on pages written before ChatGPT existed and report the error rate. That gives you a floor to read the headline share against. A story that quotes the share and leaves out the error rate has dropped the most important caveat. A corpus estimate with a known error rate is also a very different thing from a detector’s verdict on one page. Keep them apart.

A five-question check you can run on any slop story

Ask these in order:

  1. Which type is it? A count, an example or an impression.
  2. If it is a count, what was sampled? Which pages, which language, which years, and from where.
  3. How was each item labelled as AI? Does the story give the detector’s error rate, or at least admit there is one?
  4. Does it slide from “how many” to “how alike”? A share of machine text is not evidence of sameness.
  5. Can you find the original study? Or are you reading a write-up of a write-up?

Most stories fall at the third and fourth questions. Plenty fall at the fifth.

What the court can and cannot say about the news cycle

The court is the set of cases on this site, and it has limits. It can tell you what each piece of evidence measured. A count is a count, a plateau is a plateau, and a detector’s errors belong in the story. The theory page sets out the framework those verdicts hang from.

It cannot rule on the news cycle itself, or on whether any outlet’s mood about the web is justified. And it will not say the web has converged because a share is high. That inference is the one the evidence does not support.

Sources

Questions people ask

What does 'AI slop' mean in news stories?

It is a loose label for low-effort, machine-made content that people find poor or irritating. There is no fixed definition, so check which kind of evidence a story uses (a count, an example or an impression) before deciding what it claims.

Is it true that AI content now outnumbers human writing online?

One widely quoted vendor study, by Graphite, says primarily AI-generated articles reached about half of new English articles in its sample and then stopped rising. It rests on commercial detectors and counts articles published, not what readers see. The details are in the case file Half the articles, and a plateau.

Does a high share of AI articles mean the web sounds the same?

No. Prevalence is not convergence. A share tells you how much machine text there is. Whether that text has become alike is a separate question, and you need a different measurement to answer it, such as lexical diversity compared across years.

How reliable are the detectors behind these headlines?

They make mistakes, and that includes flagging human writing as machine writing. Treat any detector-based share as an estimate with a false-positive caveat. Look for a study that measured its detector's error rate on text written before ChatGPT.

How can I tell whether an AI slop story is measuring anything?

Ask what was sampled, how each item was labelled as AI, whether the error rate is given, and whether you can find the original study. If you can't answer those, the story is probably an anecdote or an impression with a number attached.