How to Spot an AI Content Farm Site
NewsGuard counted 3,749 AI content farms by June 2026. You do not need a detector to recognise one, you need four checks and about two minutes.
You land on a news site you have never heard of, the article answers your question, and something about it feels thin. Knowing how to spot an AI content farm is now a basic reading skill, and the useful method is not a detector tool. It is four structural checks you can run on the page in front of you in about two minutes.
The scale is worth stating first. NewsGuard's tracking centre, last updated 23 June 2026, lists 3,749 sites it classifies as AI content farms. Its first count, in May 2023, was 49. These are not fringe pages: the whole design goal is to look like an ordinary local outlet or trade publication.
The four criteria that define an AI content farm
NewsGuard only classifies a site this way when all four of the following hold. Borrowing their criteria is useful because it keeps you from calling every badly written page a farm.
Criterion | What it means in practice |
|---|---|
Mostly AI-produced | Evidence that a large share of the content is machine-generated, not one assisted article. |
Little or no human review | Published unchecked. The giveaway is chatbot error text left in the body. |
Looks human-run | Layout, a generic plausible name, bylines and an about page that read like a real outlet. |
No disclosure | The site does not clearly say its content is AI-generated. |
Note what is absent: writing quality. A fluent article is not evidence of a human, and a clumsy one is not evidence of a machine. The criteria are about process and disclosure, which is why they survive better models.
The two-minute check to spot an AI content farm
1. Read the about page and the masthead
Open the about page. You are looking for named people with checkable histories. Farms tend to produce one of three things: no about page, a page of generic mission language with no names, or names with no footprint anywhere else. Search one byline. A real local reporter has a trail: other outlets, a social profile, a LinkedIn, a conference talk. A fabricated one has a headshot and nothing behind it.
Also look for an address, an ownership statement and a corrections policy. Legitimate small outlets almost always have at least one. Farms rarely bother, because the audience is search engines rather than a community that would complain.
2. Scan the publishing pattern, not the article
Go to the homepage or an archive page and look at timestamps and volume. Four hundred posts a month from a site with no named staff is arithmetic that does not work. So is a site covering municipal politics in nine unrelated cities. So is an archive that starts abruptly eight months ago with no earlier history.
This check is strong precisely because it is not about any one piece of text. A good model writes a good article. Nothing makes 400 articles a month plausible from three unnamed writers.
3. Hunt for unedited machine output
Use the site's own search, or a search engine restricted to the domain, for the phrases that leak when nobody reads the output before it publishes. As a language model. I cannot fulfill that request. I'm sorry, but. Certainly! Here is a. Regenerate response.
This is the single highest-value check and it takes fifteen seconds. A site that has published any of those strings has published without review, which is criterion two, directly evidenced.
4. Follow one factual claim to its source
Pick the most specific claim in the article: a number, a date, a quote, a named official. Then try to find it anywhere else. Farms recycle and mangle. You will often find the original story at a real outlet with a different number, or an attributed quote that exists nowhere, or a date shifted by a year. We wrote about this failure mode in the context of AI news aggregators in our pillar on AI risks, and the pattern is identical here.
If the claim has no origin and the site has no named staff, you have your answer. Stop reading.
5. Check how long the site has existed
A free bonus check if the first four leave you unsure. Look at the oldest article in the archive, then search for the domain's registration date in any public WHOIS lookup. A site that registered its domain nine months ago and has four thousand articles about a town's planning disputes is not a local newspaper that went digital.
Two related tells sit in the images. Farms either use the same handful of stock photographs across hundreds of unrelated articles, or they use generated images with the characteristic smooth, slightly wrong architecture in the background. Right-click a hero image and run a reverse image search. A real outlet covering a local council meeting has a photograph of that council meeting, or it has no photograph. It does not have a generic image of an anonymous modern building that also appears on nine other articles.
Neither of these is conclusive on its own. A new legitimate outlet exists, and plenty of honest small publishers use stock photography because they cannot afford a photographer. Treat them as weight on the scale once two or three of the core four have already pointed the same way.
Why a text detector is the wrong tool
The reflex is to paste a paragraph into an AI text detector. Resist it. These tools have a false-positive problem that is well documented, they do worse on short samples, and they are confidently wrong in both directions, which is the worst combination for a decision you actually care about. We went through the evidence in do AI text detectors work.
There is also a conceptual mismatch. Even a perfect detector would answer the wrong question. AI-written is not the same as untrustworthy, and plenty of reputable outlets now use AI assistance with disclosure and an editor. What makes a content farm a content farm is the absence of review and disclosure, and those are visible on the page without any tooling.
Who runs these and why
Two economics, both worth recognising.
The common one is advertising. Volume plus programmatic ads equals revenue, and the content only has to be good enough to rank and load. This is the bulk of the 3,749.
The rarer and more consequential one is influence. Anthropic's October 2026 election post describes foreign operators running a fake news site that used Claude to produce midterm coverage advancing Russian, Iranian and North Korean narratives, alongside other actors fabricating news purely for ad revenue. Both sets of accounts were banned. The detail that should stay with you is how those operations evade per-request safety checks, which we covered in why AI safety filters miss split-up tasks.
For the related problem of fake applications wearing a real product's name, the same instinct applies and the tells differ: see how to spot a fake AI app.
If you publish, this cuts both ways
Read those four criteria as a specification of what not to resemble. The two that are entirely in your control are human review and disclosure, and they are cheap. A named author with a real bio, an editor who reads the thing before it goes out, a note on how AI is used, a corrections policy, and a publishing volume your staff could plausibly produce.
That is not just reputational hygiene. Those are the features that distinguish a small publication from a farm to a reader in two minutes, and a reader who gets that far is worth more than the one who bounces.
Frequently asked questions
Is every AI-written article from a content farm?
No. Disclosure and human review are what the classification turns on. A site that says it uses AI assistance, has named editors and corrects mistakes is doing something different from a site that hides all three, even if both used the same model.
What is the fastest single check?
Search the domain for leftover chatbot phrases such as as an language model or I cannot fulfill that request. It takes seconds and a hit is direct evidence of publishing without review.
Can I report a site I think is a content farm?
You can report it to the ad networks serving it, which is where the revenue is, and to the relevant search engine as spam. NewsGuard also maintains its tracking centre, though it is a commercial service rather than a public complaints desk.
Do these sites rank well in search?
Some do, for a while. Google's documented position treats scaled content abuse as a spam policy violation, and its spam updates have repeatedly demoted sites publishing unreviewed machine content at volume. The pattern is that they rank until an update catches them.
Does a well-designed site mean it is legitimate?
No, and looking legitimate is explicitly one of the four criteria. Templates are free. Judge the masthead, the publishing volume and one traced claim instead.
How did this land?
About the author
Senior Editor, AI & Product
Cecilia leads the Swarmz editorial desk. She has spent a decade turning complex AI and product topics into writing people actually finish, and she owns the blog's quality bar.
