Is This Article Written by AI? Six Signs to Check Before You Trust It

Key Points

• An Ahrefs study of 900,000 new web pages from April 2025 found 74.2% contained AI-generated content, though only 2.5% were written entirely by AI.
• NewsGuard has identified 3,749 AI content farm websites spanning 16 languages.
• Wikipedia’s own guide to AI writing warns readers not to rely solely on AI detection tools, which have non-trivial error rates.

In This Article

• How much of what you read online is written by AI?
• What are the six signs an article was written by AI?
• Should you trust an AI detector’s verdict?
• What will change for AI-written content over the next 12 to 24 months?
• What should you do when an article feels machine-made?
• What else do readers ask about spotting AI writing?

Quick Answer

You can often spot an AI-written article by its habits, such as inflated praise, vague sourcing and tidy lists of three, but no single sign proves anything, so look for several at once. A detector can add a second opinion. It can’t give you a final verdict.

The question matters because AI text is now everywhere. Most of it is a mix of human and machine work: in the Ahrefs study, 25.8% of new pages were purely human, 2.5% were pure AI, and 71.7% blended both. In my view, the useful question isn’t “Did a machine touch this?” It’s “Did a person check it before it reached me?”

At the far end are sites that don’t check at all. NewsGuard’s AI Tracking Center now counts 3,749 AI content farm sites. It classifies a site this way when a significant portion of its content is AI-produced, it’s published without meaningful human oversight, and the AI use isn’t disclosed to readers.

Below, I walk through the six signs editors look for, what detectors can and can’t tell you, and what to do next. The habit I’d keep is the three-sign rule: one sign is noise, but three together are a pattern worth taking seriously.

Top questions this article answers

• How can I tell if an article was written by AI?
• What are the most common signs of AI writing?
• Are AI detectors accurate?

How much of what you read online is written by AI?

A lot of it has some AI in it, but very little is pure machine output. What matters for you as a reader is whether a human checked the facts.

Ahrefs ran 900,000 newly created English pages, one per domain, through its own detector, and found 74.2% had AI somewhere in them. By a separate measure from the SEO firm Graphite, more than half of new English-language articles are now AI-generated or mainly AI-assisted.

A common misconception is that AI involvement automatically makes an article unreliable. The reality is that a person using AI to tidy a draft is very different from a site publishing thousands of unchecked articles. Graphite’s study found that while AI articles are published in huge numbers, top-ranking search results still tend to be human-written or heavily edited.

In practice, you’re not trying to catch every sentence a machine touched. The takeaway: look for signs that nobody was steering.

What are the six signs an article was written by AI?

The clearest signs are puffed-up praise, lists of three, “not X, but Y” corrections, vague sourcing, missing first-hand detail and commentary that tells you how to feel.

The best field guide comes from Wikipedia. Its volunteer editors in WikiProject AI Cleanup built a list of the signs of AI writing after reviewing thousands of flagged drafts. It’s a set of observations, not rules. Here are the six signs I’d check:

• 1. Everything sounds important. AI text often inflates significance with stock phrases like “rich cultural heritage” and “stands as a testament.” A person usually just states the fact.
• 2. Lists of three and repeated transitions. The guide flags rule-of-three phrasing and repetitive transitions. Three adjectives in a row, sentence after sentence, is a tell.
• 3. “It’s not just X, it’s Y.” AI loves to set up a contrast that corrects a misconception nobody actually had.
• 4. Vague or name-dropped sources. AI tends to prove importance by listing the kinds of outlets that covered a subject, and often credits its own shallow analysis to those sources. Look for claims you can actually trace.
• 5. No first-hand detail. Real writers include names, dates, prices and small observations only they could know. Generic articles could be about any place, product or person.
• 6. Editorial commentary. Phrases that tell you what to think, such as “it’s important to note” or “this highlights the value of,” stand in for evidence.

What about em dashes? Wikipedia’s editors note they appear in human writing too. It’s the combination with several other signs that makes human authorship less likely. That’s the three-sign rule in action: a single dash, a single list of three or a single “not just” proves nothing.

The takeaway: count the signs, don’t react to one.

Should you trust an AI detector’s verdict?

Treat it as a second opinion, not proof. Detectors make mistakes in both directions, and a “not sure” answer is often the most honest one.

Wikipedia’s AI Cleanup guide is blunt about the limits. It says detection tools produce both false positives and false negatives, and many sites now sell ways to “humanize” AI text to fool them. It also notes that the best tools score above 95% accuracy, and that a consistently high AI score can justify a closer look.

A common misconception is that a percentage tells you how much of an article a machine wrote. The reality is that most scores are confidence readings, not measurements of authorship.

If you do want a second opinion, choose a tool that admits uncertainty. You can paste text or a page address into a free AI content detector that gives two separate readings: how the writing reads, with machine-like phrases highlighted in place, and who likely wrote it. When no passage is clearly machine-written, it says “not called” instead of guessing. Full disclosure: my team built it, and like every detector, it isn’t 100% accurate. It also needs at least 150 words to give a reliable reading.

In practice, use the six signs first and a detector second. The takeaway: a detector can confirm your suspicion, but it shouldn’t create it.

What will change for AI-written content over the next 12 to 24 months?

Labels and penalties will spread faster than detection improves. I expect readers to get more help from platforms and laws than from any single tool.

What most people miss: the biggest protection for readers won’t be better detectors. It will be disclosure. If you want to follow these changes as they happen, my team keeps a running roundup of AI writing and detector news, with each story linked to the original reporting.

What should you do when an article feels machine-made?

Apply the three-sign rule, check one key fact against an original source, and only then decide how much to trust the page.

Start with the claim you care about most, such as a price, a date or a health or money fact, and look it up somewhere independent. If the article’s facts hold up, the writing style matters less. If they don’t, the style was a warning you now understand.

Then look at the site itself. An “About” page with real names, a way to contact the editors and dated corrections are all signs a person is responsible for what’s published. Missing all three, on a site publishing dozens of articles a day, is the pattern NewsGuard tracks.

In my view, the goal isn’t to avoid every article AI touched. It’s to trust the ones where a human clearly checked the work. Start with the next article that makes you pause: count the signs, check one fact, and decide from there.

Written by

Alex Shortov

CTO, AEO Content

Full-stack engineer and content infrastructure architect with 20 years of building enterprise systems.

Connect on LinkedIn

Frequently Asked Questions: What else do readers ask about spotting AI writing?

The questions below cover em dashes, detector mistakes, whether AI content is always bad, how much text a detector needs, and what an AI content farm is.

Are em dashes a sign of AI writing?

Not on their own. Wikipedia’s editors note that em dashes often appear in human writing too. They only become meaningful alongside several other signs, which is why the three-sign rule matters.

Can AI detectors be wrong?

Yes, in both directions. Wikipedia’s guide warns that detectors have non-trivial error rates. In a high-profile 2026 case, detectors flagged the first encyclical of Pope Leo XIV as machine-written, and New South Wales dropped take-home school work after detectors failed to give fair results.

Is AI-written content always bad?

No. Most new pages mix human and AI work, and only a small share are pure AI. The real risk is content published with no human checking the facts, which is what AI content farms do.

How much text does an AI detector need?

More than most people think. Short snippets swing on a single sentence, so a reading on a paragraph or two isn’t worth much. Some detectors ask for at least 150 words, and longer passages give steadier results.

What is an AI content farm?

It’s a site that mostly publishes AI-generated articles with little human oversight and doesn’t tell readers. NewsGuard tracks thousands of them, often funded by automated advertising.