Signs of AI writing: 12 tells, with examples and fixes

The most thorough public list is kept by Wikipedia's editors, who use it as "a field guide to help detect undisclosed AI-generated content." They explain where the tells come from: a model tends toward the most statistically likely wording, so specific, unusual facts get smoothed into generic praise, and "the highly specific 'inventor of the first train-coupling device' might become 'a revolutionary titan of industry'." Most of the twelve signs below are versions of that smoothing. The examples are ours, written to show each pattern, with a rewrite after each.

Word-level signs

1. AI vocabulary in clusters

Some words became far more common after chatbots arrived. A study of over 15 million PubMed abstracts found "an abrupt increase in the frequency of certain style words" and estimated that at least 13.5% of 2024 abstracts were processed with a language model. Wikipedia's list includes delve, pivotal, tapestry, underscore, testament and landscape (as an abstract noun). One such word may be coincidence; several together are a pattern. Example: "This guide delves into the pivotal role internal links play in today's SEO landscape." Rewrite: "Internal links tell Google which pages matter. Here is how to add them."

2. Stock openers and filler

Phrases that announce importance instead of showing it: "In today's fast-paced digital world", "It's important to note that". Example: "It's important to note that page speed plays a crucial role in rankings." Rewrite: "Our product pages took six seconds to load on mobile; after we compressed the images, they took two." Delete the frame, keep the claim, then make the claim specific.

3. Transitions at the start of every sentence

"Moreover", "Furthermore" and "Additionally" opening sentence after sentence; Wikipedia flags "Additionally" especially when it starts a sentence. Good sentences connect without scaffolding. Where a bridge is needed, make it say something: "That broke in practice" carries more than "Moreover".

4. Inflated significance

Ordinary facts dressed up as milestones, often with an "-ing" clause tacked on the end: "stands as a testament", "marking a pivotal moment", "highlighting its importance". Example: "The bakery opened in 1998, marking a pivotal moment in the neighborhood's culinary story." Rewrite: "The bakery opened in 1998 and still bakes 400 loaves before 7 a.m." Puffery is the opposite of SEO copywriting, which persuades with the detail the reader came for.

Structural signs

5. Everything in threes

Models "overuse the rule of three", from "adjective, adjective, adjective" to three short phrases in a row, often to make a thin point look complete. Example: "Our platform is fast, reliable, and scalable." Rewrite: keep the one adjective you can prove, and prove it.

6. "Not just X, but Y"

Negative parallelisms correct a misconception nobody had: "It's not just a plugin, it's a growth engine." Humans use them too, Wikipedia notes, but a text full of them reads as generated. Say plainly what the thing does.

7. Uniform rhythm and repeated openers

Every sentence the same medium length, several in a row starting with the same word. Read a paragraph aloud: if it never speeds up or stops short, it is flat. Cut one sentence to four words. Let the next run long.

8. Summaries that restate the section

Paragraphs and posts that end with "In summary" or "Overall" and repeat what was just said. Wikipedia files this under signs typical of older models, which often added a "Conclusion" section and ended paragraphs by restating their core idea. End on the last useful thing instead.

9. Chatbot formatting

Bulleted lists where each item starts with a bold label and a colon, headings in Title Case, bold scattered over key phrases, emoji as bullets. Use prose where prose reads easily, and sentence case in headings.

Factual and leftover signs

10. Vague attributions

"Experts argue", "Industry reports suggest", "Observers have cited": claims hung on an authority that is never named. Wikipedia also notes models present one or two sources' views as widely held. Name the source, or cut the claim.

11. Invented or broken sources

Links that never worked, studies that do not exist, identifiers that lead to unrelated papers. Google's own guidance explains why: "generative models don't retrieve facts, but predict a likely sequence of words," so their output "may contain inaccuracies (also known as hallucinations)," and it calls it "critical to manually factcheck and review all AI-generated content" before publishing. Open every link and check every number.

12. Leftovers from the chat

"I hope this helps!", "Certainly!", "Would you like a more detailed breakdown?", or hedges such as "While specific details are limited" that cover for a fact the model could not find. These are among the most reliable tells, because they are addressed to the person who prompted the model, not to the reader.

Signs that are not signs

Wikipedia's editors list indicators that mislead. Perfect grammar is one: plenty of human writers have it. Mixing casual and formal registers is another. Em dashes were a famous tell, but by Wikipedia's account the newest models increasingly avoid them, so their absence proves nothing either. And people are weaker judges than they think: a 2025 study the page cites found human ability to tell LLM text from human "no better than random chance", though expert annotators in a 2025 preprint caught 90% of the AI text.

What does point to a person, on the same page: plain "is" and "has" sentences, and plain verbs, such as "wrote" rather than "authored" and "used" rather than "utilized".

How to tell if something was written by AI

Look for clusters, not single signs: three or four from different groups in one text mean more than any one word. Compare the text with the writer's earlier work, since a sudden shift in vocabulary, rhythm or formatting tells more than any single sign. Then check what a model cannot fake. Do the links work and say what the text claims? Can the writer explain why a sentence is there? Was it published before ChatGPT's public launch on November 30, 2022, which rules a chatbot out? A detector score adds one more estimate, not a verdict.

If the text is your own AI draft, the fix is editing, not disguise. Most of these tells come from a draft with nothing specific to say, which is why a content brief that names the angle and the facts before drafting prevents more of them than any rewrite. Our guide to AI blogging strategy covers where AI drafting fits, and the post on a Yoast-optimized AI writer explains why a green readability light does not catch any of this.

How Structura handles this

The free AI humanizer is an analyzer, not a rewriter. Paste a text and it lists the tells it finds, quoting each one: stock phrases, formulaic transitions at the start of sentences, sentence openers repeated three times or more, and uniform sentence rhythm, in four languages. The analysis runs in your browser and the text is not stored. It does not rewrite, because automatic paraphrasing garbles meaning.

It covers signs 2, 3 and 7, plus a few phrases from 1 and 4 such as "delve into" and "plays a pivotal role". Its phrase list is deliberately short, to avoid flagging sentences a person would write; it does not spot triplets, negative parallelisms or vague attributions, and it cannot check facts or links, so signs 10 to 12 still need a human read.

Structura itself writes each post in an author persona you define, cites sources from a list you approve and fetches every outbound link before the post ships, which catches links that do not resolve. Whether a source says what the post claims is still worth a look during review. Approval mode holds each post as a draft until you have read it. How a reviewed draft becomes something worth ranking is the subject of SEO content writing.

FAQ