Signs of AI writing: how to spot it in a draft

The short answer
The most common signs of AI writing are nine: stock vocabulary, em dashes, lists of three, "not just X but Y" lines, heavy formatting, cheerful assistant openers, hedging with no opinion, vague details and a conclusion that repeats the intro. While one sign proves nothing, a cluster of 4 or 5 signs is a strong hint.
What are the 9 signs of AI writing?
The 9 signs of AI writing are stock vocabulary, em dashes, lists of three, "not just X but Y" lines, heavy formatting, cheerful openers, hedging, vague details and a conclusion that repeats the intro. Each one's a tendency, not a rule, and each has a fix you'll make in a minute.
Checklist: 9 signs
- Vocabulary tells. Stock vocabulary: favourite verbs and adjectives keep coming back. Fix: keep one, change the rest.
- Em dashes in nearly every paragraph. Fix: use a comma, a colon or a full stop.
- Structure tells. Lists of three, like "fast, simple and reliable". Fix: give two items, or four.
- Contrast lines such as "this isn't a tool, it's a system". Fix: state the claim plainly.
- Heavy formatting: bold type, bullets and headings on a short note. Fix: turn it back into sentences.
- Tone tells. Cheerful openers ("Sure, here's a draft") or a closing offer of more help. Fix: start with the point.
- Hedging, as in "it depends on many factors". Fix: say what you think and why.
- Content tells. Vague details: unnamed sources and examples that fit any topic. Fix: add a name, a number or a date.
- A conclusion that repeats the intro. Fix: cut it, or end on the next step.
Which words and phrases give AI writing away?
The words and phrases that give AI writing away are stock verbs and adjectives repeated across drafts, assistant openers and closing summaries that restate the intro. Pangram sells an AI detector, and its word list is long, though the company calls it non-exhaustive. Wikipedia's field guide for editors flags overused "AI vocabulary" as well, and notes that models skip plain "is" and "are" phrasing.
Vendors retrain models, so word lists age fast, and last year's list won't match this year's model. Don't memorise a list. Count repeats in your own copy instead, and swap each stock opener for the sentence that carries your point. A plain swap works best: "use" for a grand verb, a number for a grand adjective.
Which structure tells point to AI-written text?
The structure tells that point to AI-written text are lists of three, "not just X but Y" frames, bullet-heavy layout and bold labels on every paragraph. Uniform paragraph lengths and tidy headings on a short piece add to the picture.
- Triads. People use them too, but a page that groups everything in threes looks machine-made.
- Negative parallelism. Wikipedia names the "not just X but Y" pattern. One use in an article is style; a use in every section isn't.
- Layout. Broken Markdown ranks among the most common recent signs in Wikipedia's guide, for example stray asterisks and pound signs, and a bold label opening every paragraph reads as a template.
What tone and content tells show up in AI text?
The tone and content tells in AI text are hedging with no stated opinion, details that fit any topic and no first-hand fact a reader could check. A typical assistant opens with enthusiasm, stays positive about each idea and closes with an offer of more help, and Pangram adds a formal tone and a reluctance to criticise a viewpoint.
Content tells matter most, since they're the ones that cause real errors. Watch for "studies show" with no study, or "an expert says" with no name. Pangram notes that AI skips concrete details and proper nouns, and it defaults to generic names. Check every claim that's got a number, a date or a name, and cut it if you can't find the source in a 2-minute search. Wikipedia's guide gives citation problems a category of their own: broken links, invalid identifiers and book citations without page numbers.
What does vendor data say about AI writing signs?
Vendor data on AI writing signs comes from Pangram's own corpus, which found markdown and AI phrases about 12 times more often in AI text than in human text. Pangram describes the corpus as millions of human and AI documents and gives no date for the analysis; we read the page in November 2026.
| Signal | Human text | AI text |
|---|---|---|
| Markdown | 8 | 90 |
| AI phrases | 3 | 30 |
| Em dashes | 2 | 17 |
| Bullet lists | 3 | 28 |
| Triads | 5 | 19 |
| "Not just X but Y" | 1 | 3 |
Read these as one vendor's own analysis, not an independent study: nobody else has published matching rates that we've found. The em dash's frequency is high in AI text, yet its proof value on one page is low. The page isn't consistent with itself on em dashes: its table gives humans 2 per 10,000 words, and its em dash section says 5. Pangram sells a detector, so its advice to use one is a sales pitch as well as guidance.
Why does no single sign of AI writing prove anything?
No single sign of AI writing proves anything since people use every one of these habits, so only a cluster of 4 or 5 signs in one text is worth a closer look. Wikipedia's WikiProject AI Cleanup calls its page "descriptive, not prescriptive", says the signs are only potential indicators and notes that many patterns appear in human writing too.
Detectors don't settle it either, and they aren't neutral judges. According to Wikipedia, tools like GPTZero and Pangram do better than chance yet still carry non-trivial error rates, with false positives on human text, and edits can distort their results. Human judgment isn't reliable either, it adds. Treat a cluster as a reason to ask for notes, drafts or sources, then judge the work on accuracy.
How do you remove AI tells from your draft in 10 minutes?
To remove AI tells from your draft in 10 minutes, search for the stock phrases, break up lists of three, cut the summary ending, add one real detail and read it aloud. Give each step about 2 minutes; it's a timed pass, so don't polish.
- Search for stock openers and repeated adjectives, and swap all but one.
- Break any list of three into two items, or a sentence.
- Cut a final paragraph that only restates the intro.
- Add one detail only you know: an outcome, a mistake or a tool you used.
- Read it aloud and mark any line you wouldn't say.
The em dash gets one pass of its own, covered in the dedicated guide. The full editing method, with a before and after paragraph, is in how to humanize AI text.
Signs of AI writing before and after
The same request, asked two ways. Read the second one slowly. Each line does a job.
Sure! Time management is more than a skill, it's a mindset. Plan, prioritise and review, and you'll see results. Studies show this works. Let me know if you'd like more tips!
I keep a paper list of three jobs a day. On Mondays I cut one of them. When I tried an app for a month, I spent more time sorting tasks than doing them, so I went back to paper.
Why it works
The rewrite removes four tells at once: it drops the opener and the closing offer, replaces "studies show" with a specific habit and adds a detail only the writer has.
Signs of AI writing examples
Find the AI tells in a draft
Asking for quoted lines and rewrites gives you edits instead of a verdict.
Read the draft below and list every sign that it was written by an AI assistant. Group them as word, structure, tone and content signs. For each one, quote the line, say why it reads as machine-made in one sentence and suggest a rewrite. Don't claim the text is or isn't AI-written, since these signs are tendencies and not proof.
Draft:
{{draft}}How to use
- Copy it with the button above.
- Fill the blanks in {{double braces}} with your own details.
- Paste it into ChatGPT, Claude or Gemini and send it.
- Then follow up: ask for one change at a time.
Before you run “Find the AI tells in a draft” for the first time, the lesson on how to use AI prompts step by step walks through copying, filling and sending it.
Style rules that reduce AI tells
Each rule removes one tell from the list, and the sample anchors the voice.
Write in plain, direct sentences. Open with the answer, not with a greeting or a recap of my request. Vary sentence length, and use short sentences for key points. Use no more than one list of three per page. Name specific people, tools, numbers and dates, and if you don't know one, say so instead of guessing. Skip closing offers of more help. Match the voice of this sample: {{sample}}How do you prompt an AI model for fewer writing tells?
To prompt an AI model for fewer writing tells, give it a banned phrase list, a sample paragraph of your own and a rule against summary endings. To apply these style rules to a real draft prompt, start from the human-sounding ChatGPT prompts, a write like a human prompt pack that puts the banned phrases and the voice sample in place for you. You won't need to retype the style rules each time.
A standing rule in ChatGPT's custom instructions beats repeating it in each chat, and the custom instructions generator builds one. This page covers diagnosis; the prompt pack's job is generation, and that's where you'll get a natural first draft. More guides sit in the prompt guides.
Signs of AI writing FAQ
Still stuck? Build it in Prompt Studio. It'll score each part as you go.
How can you tell if ChatGPT wrote a text?
You can tell if ChatGPT wrote a text only by a cluster of signs, such as stock phrases, lists of three and a repeated conclusion, and never by one sign alone. Four or five together justify a question, not an accusation.
Are AI detectors accurate at spotting AI writing?
No, AI detectors aren't reliable at spotting AI writing, and they flag human text as AI often enough that a score alone proves nothing. One edit can swing a result, so treat a score as a single hint.
Is the em dash a sign of AI writing on its own?
No, the em dash isn't a sign of AI writing on its own; it counts only when it appears in most paragraphs alongside other tells. Wikipedia's guide files em dash overuse under historical indicators, since it shifts as models change.
What words does AI writing overuse?
The words AI writing tends to overuse are stock verbs, grand adjectives and transition words that open sentences, and the same few repeat across unrelated drafts. The set shifts with each model release, so count the repeats in your own draft.
Can you make AI writing undetectable?
No, you can't make AI writing undetectable: nobody can promise that, and the better goal is an accurate draft in your own voice. Detectors make mistakes in both directions, so edit for facts, detail and voice instead.