Signs of AI Writing: 21 Red Flags That Give ChatGPT Away

What are the most common signs of AI writing?

The common signs pile up. One of them is a style choice. Five of them in a short page is a pattern. None of them opens the chatbot and checks the history.

Structural red flags: formatting, headings, and bullet overload

Look at the skeleton before you look at the adjectives. Model drafts often arrive with a heading every few lines, a bullet under every heading, and a bold lead-in on every bullet. The structure is doing the thinking. The sentences under it repeat the heading. A student paper can be organized without looking like a changelog. If every section is the same length, and every bullet could move into another paper, the outline was generated.

Tone signals: hedging, flattery, and no personal voice

Helpful models hedge. It is important to note. While there are challenges, there are also opportunities. Both sides have merit. The hedge is polite and it refuses to choose. A writer with a stake picks a side and says what would change their mind.

Flattery shows up in emails and in drafts that were asked to sound encouraging. Great question. You are absolutely right to wonder. That voice has no memory of the class, the client, or the argument. It has a memory of being rated for warmth.

Personal voice is not slang. It is a detail only this writer would have: the Tuesday reading, the broken scale in the lab, the sentence a professor marked last time. If the draft could be signed by anyone, the tone is the tell. How to read an AI writing score is what to do when that feeling becomes a percentage.

How to spot ChatGPT vs Claude vs Gemini writing

ChatGPT defaults to confident structure. Short setup, three supports, a bow on the end. It reaches for delve, crucial, and the em dash. It will invent a neat example if you do not pin it to your notes.

Claude often writes longer and more careful. It names tradeoffs, apologizes for limits, and closes in a calm register. The hedges are the tell, not the warmth. A page that keeps saying the question is nuanced, then never uses a number from the assignment, is a Claude-shaped draft even if the words were edited after.

Gemini tends toward tidy explainers and list-shaped answers, close to a help page. Grok can be looser or more conversational, which does not make it human. The model name is a guess. The habits are the evidence. A watermark, when one exists, is a different test. What is the ChatGPT text watermark is that other test, and most people cannot run it.

Will I get flagged for AI if I use an em dash?

No. Not for the dash alone. Detectors are not punctuation counters with a guilty verdict attached. They look at how predictable the whole passage is. A novelist who lives on em dashes will not match a chatbot that uses the dash as a hinge and then repeats itself.

You can still get a high score on careful human prose. Formal assignments, methods sections, and very even English share a texture with model output. The dash is the detail people notice because they can see it. The score is usually coming from the sentences around it.

If you are worried about one mark, read the paragraph out loud. If you would say it, keep the dash. If the clause after the dash only restates the clause before it, cut one of them. That edit helps a reader. It is not a trick.

Why manual spotting is not enough

Eye tests drift. After you learn the word delve, you see it everywhere and miss a flat page that avoided the word. After you learn the em dash joke, you accuse journalists. A checklist of 21 signs is a way to slow down. It is not a detector.

Models are also steered. A prompt that says "sound like a tired junior" removes the stock words and leaves the even rhythm. A light human edit removes the worst sentences and leaves the rest. Manual spotting catches the unedited chat. It misses the careful one.

The other failure is the false alarm. Non-native English, template essays, and technical prose trip the same instincts. How do AI detectors work is the mechanics behind that miss. Are AI detectors accurate is why a confident eye, or a confident tool, still needs the draft in front of it.

How to verify a suspicious text in 3 steps

First, mark the habits yourself. Stock words, triples, generic closers, headings that repeat, and a voice with no local detail. Write down the sentences, not a mood.

Second, check the text with DetectGPT. Paste at least 50 words, stay under 5,000, or use a public URL if the words are already on a page. Read the yellow, orange, and red lines. How DetectGPT scores a draft is the map for those colors. The percentage is a share of the prose, not proof of who typed it.

Third, ask for the work behind the page if the stakes are real. Notes, an outline, an earlier version. A score and a checklist cannot see those. If you are the writer, keep them before anyone asks. OpenAI's own retired classifier is the reminder that a detection label was never a reliable verdict, even from the company that built the model.

FAQs

What are the biggest signs of AI writing?

Stock vocabulary, rule-of-three lists, generic conclusions, and a voice with no detail from the actual assignment. One of those is a style choice. Several together are a pattern.

Does ChatGPT overuse em dashes?

Often, yes. People do too. One em dash does not mean the page was generated. A dash that only repeats the first half of the sentence is the weaker tell.

Can I spot Claude or Gemini the same way?

The habits overlap, and the defaults differ. Claude hedges more. Gemini often looks like a help article. The model name is still a guess.

Will an em dash get me flagged?

Not by itself. Scores come from how predictable the passage is. Formal human writing can score high with no dashes at all.

Is a checklist the same as a detector?

No. A checklist slows you down. A detector marks lines. Neither one proves authorship. Use both, then look at drafts if the result matters.