How to tell if something was written by ChatGPT

There’s no single tell. ChatGPT, Claude, and Gemini have writing habits, people have writing habits, and the two overlap a lot. What works is looking for several signs at once, then checking the one thing a model can’t produce after the fact: the history of how the document was written.

Patterns that show up in AI drafts

These turn up often enough in unedited model output that editors and teachers have learned to spot them:

• Sentences of similar length, one after another, with few very short or very long ones.
• “Moreover”, “Furthermore”, and “Additionally” opening sentence after sentence.
• Contrast formulas like “It’s not just X, it’s Y” and “not only X but also Y”.
• Lists of exactly three adjectives or phrases, even where the writer had two things to say, or five.
• Paragraphs whose last sentence restates the paragraph, and a closing section that restates the whole piece.
• Claims with nothing attached: “many experts believe”, “this plays a crucial role”, with no names, numbers, or examples.
• Citations, quotes, or statistics that look real and can’t be found. Models sometimes invent these, and a reference that doesn’t exist counts for more than any style pattern.

Why none of these is proof

Every item on that list also appears in writing by people. School essays reward even rhythm and tidy summaries. Writers working in a second language often rely on a smaller set of transitions. Business writing is full of claims with nothing attached. And someone who edits an AI draft carefully can remove most of these patterns.

This is also why detectors get some groups wrong more often than others. In a 2023 Stanford study, seven detectors were run on 91 TOEFL essays written by non-native English speakers. On average they labeled 61.3% of these human-written essays as AI-generated (Liang et al., Patterns).

Check the document’s history

The strongest evidence is how the document came to exist. In Google Docs, File > Version history > See version history lists saved versions with timestamps and the name of whoever made each change. Word keeps a similar list under File > Info > Version History for files stored in OneDrive or SharePoint.

Writing that grew over several sessions, with sentences added, cut, and moved around, looks different from a document where nearly all of the text appeared in one step. That second pattern is worth a question. It isn’t proof either, since some people draft in another app and paste in the finished text.

If you’re the writer, keep that history. If you’re the reader, ask for it before you decide anything.

Ask about the content

Most people can talk easily about something they wrote: why they picked an example, where a quote came from, what they cut. Ask about one specific passage and listen for familiarity with the material.

Using a detector

A detector does a faster, broader version of the pattern check above, trained on far more text than any person has read. It’s still a pattern check. Scan the whole document instead of a paragraph, since short passages give any model little to work with, and look at which sentences it highlights. Then read those sentences yourself.

Our AI text detector calls a text “Likely AI-generated” only at 90% or higher and reports the wide middle range as “Inconclusive”. Whatever tool you use, find out where its cutoffs are before you act on a score. What an AI detector score actually means goes through this in more detail.