Guide · 12 min read

How to spot AI-generated writing: a complete guide

Last updated: May 2026

AI writing has a tell. Actually, it has about sixteen of them. Here's what to look for, why each one happens, and how to tell the difference between an AI draft and a slightly nerdy human writer who genuinely likes em dashes.

If you've spent any time around AI-generated text in the last year, you've probably noticed a feeling. Something's off. You can't always put your finger on it, but the writing reads too smooth. Too tidy. The sentences land in the same shape over and over.

That feeling has a name now, and a list. In 2024, Wikipedia's editors (some of the most experienced volunteer readers and writers on the internet) started seeing a wave of AI-generated articles being submitted to the encyclopaedia. They wrote down what they kept noticing. The result was a guide called Signs of AI Writing, and it's become one of the more useful documents on the subject.

This article walks through what they found. Some of these patterns are technical, some are stylistic, and some are the kind of thing you notice immediately once it's pointed out. None of them, on their own, prove a piece of writing is AI. But put together, they form a fairly reliable fingerprint.

QUICK CHECK

If you'd rather skip the reading and just have something analysed, paste it into Telltale. We check all sixteen of these patterns automatically and show you which ones fired.

Why AI writing has tells in the first place

Before we get to the patterns, a quick word on why they exist at all. Large language models like ChatGPT, Claude, and Gemini work by predicting the most probable next word, then the next, then the next. They're trained on enormous amounts of human text, and they get very good at producing fluent output. But they're optimising for likely, not distinctive.

This means when an AI faces a creative choice (which word, which sentence shape, which metaphor), it tends to choose the most statistically common option. The same option, in the same situation, every time. A human writer might write "delve into" once in a piece and then deliberately pick "explore" or "look at" or "get into" the next time, because they don't want to repeat themselves. The AI doesn't have that aversion. So certain words and shapes keep showing up.

The signs below are basically a list of the things AI keeps reaching for. They're not bad on their own (humans use all of them) but they show up at unusual rates and in unusual combinations when a machine has been involved.

The vocabulary tells

Words that turn up far too often

The most famous one is delve. It's a perfectly fine word. Academics have been delving into things for centuries. But ChatGPT in particular reaches for it so often that "delve" alone has become a meme among editors and researchers. One analysis of academic papers found a sudden spike in the word's use starting in late 2022, neatly tracking the public release of ChatGPT.

Other words in the same family:

You won't catch AI on a single instance of "delve." But when you see three or four of these words clustered together in a short piece, the probability shifts.

Corporate vocabulary

There's a second tier of vocabulary that isn't AI-specific but tends to cluster in AI output because the model picks up the marketing-speak from its training data. Words like comprehensive, holistic, robust, leverage, synergy, paradigm, groundbreaking, transformative. Each individually is fine. Together they create a particular texture: the texture of a LinkedIn post or a company press release.

The structural tells

The em dash problem

LLMs love an em dash. They drop them in everywhere, often for dramatic effect — like that — when a comma or simply rewording the sentence would do the same job. A real writer who uses em dashes tends to use them sparingly; AI uses them at three to five times the human rate.

If a piece of writing has more than two or three em dashes per paragraph and they're being used for emphasis rather than mid-sentence interjection, that's a strong signal.

Negative parallelism

It's not just X. It's Y.

This construction is a small rhetorical device that human writers occasionally use for emphasis. AI uses it constantly. "It's not just a tool, it's a movement." "This isn't merely a meeting, it's a turning point." The structure is meant to feel weighty and reframing. After the third or fourth time in a single piece, it starts to feel like a verbal tic.

A related variant: "No A, no B, just C." Same dramatic flourish, same overuse.

Present participle tailing clauses

This is one of the more reliable structural tells, and once you see it you can't unsee it. AI loves to end a sentence with a present-participle phrase that vaguely gestures at importance:

These clauses are filler. They don't add information; they add the feeling of having said something important. A human writer would usually either commit to the claim ("This was important because…") or skip it entirely. AI does neither.

The rule of three

Triplets are a classic rhetorical device, and one AI absolutely cannot resist. Adjectives in threes ("clean, simple, and elegant"), concepts in threes ("speed, precision, and care"), examples in threes ("from coding to writing to design"). It happens in human writing too, of course. But AI does it relentlessly. Often three or four times in a single short piece.

False ranges

From intimate gatherings to global movements, this idea has shaped how we live.

The "from X to Y" construction creates the impression of a thoughtful spectrum being covered, when in reality nothing concrete has been said. AI uses this constantly because it sounds substantive without requiring actual substance. A human writer would usually pick one specific example. AI prefers the vague sweep.

Compulsive summaries

"In conclusion." "To summarise." "Overall." "In short." AI was largely trained on essays and articles, and it ends sections like it's been told to. Even when the conclusion is unnecessary, even when it's just restating the previous sentence in different words, the model can't help wrapping up.

The transition phrase overload

Furthermore. Moreover. Additionally. It is worth noting. It should be noted. AI strings these together like a student trying to hit a word count. Human writers vary their transitions or skip them entirely. AI reaches for the same handful again and again.

The tonal tells

Promotional flattery

Everything in AI prose is breathtaking, captivating, majestic, stunning, awe-inspiring. Every city is vibrant. Every culture is rich. Every landscape is picturesque. The model defaults to superlatives because they're statistically common in the descriptive writing it was trained on (travel magazines, brochures, marketing copy). The result is text that sounds like a tourism advertisement.

Significance over-emphasis

AI keeps telling you why things matter rather than trusting the content to make its own case. You'll see phrases like:

A human writer usually shows significance through specific detail. AI states it directly because it doesn't have specific detail to draw on.

Vague attribution

"Many experts say." "Studies show." "Researchers have found." "It is widely accepted." AI loves these phrases because they sound authoritative without requiring it to actually cite anything (which it can't reliably do). A human writer who doesn't have a source typically either finds one or hedges more carefully. AI confidently asserts.

The formatting tells

Excessive bolding

AI loves to bold things. Key terms, important concepts, names, dates. It treats every paragraph like a textbook trying to highlight study material. A human writer uses bold sparingly because they trust the reader to find what matters. AI bolds compulsively.

Curly quotation marks

This one's a small technical tell, but it's specific. ChatGPT and some other models output curly quotation marks (" ") by default, even in contexts where the surrounding text uses straight quotes (" "). On its own this proves nothing (many writers use curly quotes too) but combined with other patterns, it's another small data point.

What this all means

Here's the important thing: any one of these patterns can appear in human writing. Real writers use em dashes. Real writers say "delve." Real writers begin paragraphs with "Furthermore." A guide like this could, if misused, lead you to accuse perfectly innocent human writing of being AI.

The way these patterns actually work is together, in combination. AI writing isn't characterised by any single tell. It's characterised by the density and co-occurrence of multiple tells, plus the absence of the things human writing usually has:

If you read a piece that hits half of the patterns above and has none of the human markers, it's very probably AI. If it hits one or two patterns but has rich personal detail and a clear voice, it's almost certainly human (just maybe an academic).

How to use this in practice

If you're a teacher worried about AI submissions, a journalist verifying a source, or just someone trying to read the internet more critically, the practical approach is:

  1. Don't rely on any single pattern. One em dash is nothing. One "delve" is nothing.
  2. Count the patterns together. If a short piece hits five or six of these, the odds shift heavily.
  3. Look for human markers. What specific, concrete, particular thing does this piece contain that an AI couldn't have generated without prompting?
  4. Don't accuse people based on this alone. These are signals, not evidence. AI detection tools, including ours, can be wrong.
  5. Pair pattern detection with judgement. The best AI detection isn't a tool. It's a tool plus a human reader who knows the writer's previous work.
TRY IT

Telltale runs through all sixteen of the patterns above automatically and shows you which ones fired in your text, with examples and explanations. Free to use, three checks a day, no signup required.

→ Analyse some text

A note on false positives

One last thing worth saying clearly. AI detection tools, including ours, produce false positives. Real people get flagged. This has become a real problem in education, where students have been accused of using ChatGPT for assignments they wrote themselves.

The patterns in this guide tend to over-fire on:

If you're a teacher, please don't accuse a student of using AI based solely on a detector's verdict. Talk to them. Ask about their writing process. Look at their previous work for comparison. The tools are useful as a flag, never as proof.

Further reading

Published May 2026 · telltale-ai.com
Back to tool · Privacy · Terms