Black and white photo of a person in a hoodie by the river reading a phone, with the informed.now logo and the words news without noise

AI News Quality

What is AI slop, and how much of your news is it?

Content farms now publish machine-written filler faster than anyone can read it, dressed to look like news. Learn where AI slop reaches readers and why choosing your sources beats trying to detect it.

· 4 min read

Summary: AI slop is cheap machine-written content published at volume to collect ad money rather than to inform. It fills a large share of the new articles on the open web but a far smaller share of what you read, because search engines mostly leave it out. The slop that reaches you comes through feeds, shares, and unfamiliar sites, so choosing your sources protects you better than trying to spot machine text.

Publishing an article used to cost somebody an afternoon. Now it costs almost nothing, and when writing costs almost nothing, the only scarce thing left on a page is a name that answers for it.

What counts as AI slop?

AI slop is text generated by a language model, published in bulk, and dressed as journalism so that ad systems pay for every visit. The page earns the same money whether its claims are accurate or invented, so nobody at the site checks them.

As of June 23, 2026, NewsGuard had identified 3,749 AI content farm sites in 16 languages. The tracker's definition, as it stood in June 2026, has 4 parts: A substantial share of the site's content is produced by AI, the content is published without significant human oversight, the site presents the content as if journalists wrote it, and the site discloses none of this.

The signs NewsGuard listed in 2026 are easy to picture: A generic site name, dozens of new articles a day, a chatbot's error message left sitting in the copy, false claims, and revenue from programmatic ads, which are ad slots sold by automated auction with no person choosing the page.

A newsroom that uses AI to draft or transcribe while an editor checks the result is a different subject, covered in our post on AI-generated news. Slop comes from sites with no reporter, no editor, and no disclosure.

How much of the web is AI slop?

By count of new articles, roughly half, according to a 2025 Graphite study. Graphite ran an AI detector over 43,000 English-language article URLs published between January 2020 and May 2025 and pulled from a public archive of the web. AI-generated articles reached 39% of new articles 12 months after the launch of the chatbot that put language models in public hands, passed human-written articles in November 2024, and then held roughly level. Graphite's own follow-up averages 3 detectors instead of 1 and puts the share 3.3 points lower on average.

The half from 2025 is a rough ceiling for slop: Graphite counted every article its detector judged machine-written, a wider group than NewsGuard's content farms, and a machine-written article with a careful editor behind it would land in the same count.

AI slop rarely reaches readers through search

Graphite reported in 2025, citing a separate study of its own, that AI-written articles mostly don't appear in search results or chatbot answers. A content farm can stock the web's shelves all day while search sends shoppers down other aisles.

The raw count of pages overstates a reader's exposure: Slop takes a smaller share of what a person reads than of what gets published.

Neither study measured feeds or shared links, so nobody has counted the slop in those routes. We'd bet the exposure concentrates in those unmeasured routes: In a feed that ranks posts by clicks, in a link a friend forwards, and on the unfamiliar site either of them leads to. Each route delivers a page without a ranking system vouching for the page.

Chatbots remain a small route to news. In the 2025 Reuters Institute Digital News Report, 7% of respondents used AI chatbots for news each week, and 15% of respondents under 25 did, 8 points more than the overall figure. Readers worry all the same. In that 2025 report, 58% of respondents said they worry about telling what's real from what's fake online, and the US figure was 73%, 15 points above the overall figure.

Isn't AI slop just clickbait with a new name?

Partly, yes. People wrote filler for ad money long before language models existed, and clickbait ran on the same business: A cheap page, a loud headline, and an ad beside both.

The label is fuzzy too. A detector decides what counts as machine-written, and detectors err. Graphite reported in 2025 that its detector marked 4.2% of human-written articles as AI and missed 0.6% of AI-written ones. An article that a person drafts with a model and then rewrites sits between the categories, and no detector can say where. Nobody can settle that middle case. "AI slop" has no clean edge, and every count of it, the 2025 figure of roughly half included, inherits the blur.

What changed is the price. A human content mill paid a writer for each article, and the wage bill capped how much the mill could publish. A language model removes the cap, which is how a site with almost no human oversight posts the dozens of articles a day that NewsGuard described in 2026. The motive is old and the volume is new.

Choosing sources works better than detecting slop

Graphite's detector, built for the job, wrongly flagged 4.2% of human-written articles in 2025. We'd bet a reader judging paragraphs by eye does worse than the detector, because a language model writes fluent sentences by default and fluent sentences used to be the sign that someone took care.

The question a reader can answer is who published the page and what the publisher loses if the page is wrong. These 4 habits answer that question, and every habit works with any site, paper, or feed.

  1. Check who runs an unfamiliar site. Find the about page and look for a named owner, named editors, and a way to reach them. A generic site name with no people behind it matches the content farms NewsGuard described in 2026.

  2. Read laterally. Search the site's name in a separate tab and see what other sources say about the site. A site describes itself in its own favor, so lateral reading gathers the evidence from everyone else.

  3. Prefer named outlets with bylines and corrections. A byline names a person you can look up, and a real reporter has a history that an invented name lacks. A corrections page shows the outlet repairing its errors in public, a cost a content farm has no reason to pay.

  4. Slow down on shares in feeds. A shared link shows a friend's name where the publisher's name should be. Read the domain before you read the article or forward the link, and when the domain is unfamiliar, go back to Step 1.

All 4 habits ignore how the prose was made. They work the same on junk a person wrote, so the fuzzy label stops mattering to the reader.

We built informed.now as one way to choose a source once: Its 6 summaries arrive by text message, and one transparent filter picks all 6. It's one text every weekday morning.

A track record still costs years

A language model can generate the article, the byline, and the site around both for almost nothing. A track record still costs years of being right in public and correcting the misses where readers can see them. The name on the page deserves more of your attention than the prose under the name.

Have questions about informed.now? Get answers.

Keep reading

Black and white photo of 2 people sitting close together on a couch looking at a phone, with the informed.now logo and the words news without noise

Can you trust AI news summaries?

Black and white photo of a smiling person with curly hair in a T-shirt holding a phone with both hands, with the informed.now logo and the words news without noise

How accurate is ChatGPT on today's news?

Black and white photo of a smiling person with short curly hair and round glasses holding a phone, with the informed.now logo and the words news without noise

How do you tell when the news was written by AI?

Get factual news by text

Free forever with no bias, clickbait, or ads