ChatGPT detector

The fingerprints of ChatGPT’s writing

ChatGPT is the most widely used writing assistant in the world, and its default voice is easy to recognise once you have read a lot of it: confident, upbeat, symmetrical and padded with signposting. Each model generation changed the details. GPT-3.5 opened with “Certainly!” and closed with “In conclusion,”; GPT-4 made “delve” and “tapestry” famous; GPT-4o leaned into em dashes, emoji bullets and “It’s not just X — it’s Y”; GPT-5 is plainer but keeps the structure. GPTTrace’s vocabulary list is era-tagged so that it rewards the newer tells more than the old clichés.

Artifacts that only come from copying

When people copy answers from ChatGPT’s search mode or deep research, they often bring along machinery that was never meant to be seen: internal citation tokens such as oaicite and contentReference, placeholders like turn0search3, source brackets like 【4†source】, and URLs tagged utm_source=chatgpt.com. These are close to proof that text passed through ChatGPT, and GPTTrace flags them prominently.

Structure over vocabulary

The single most stable signal across versions is structure. ChatGPT likes a short framing sentence, three parallel points, a bolded label at the start of each bullet, a contrast built as “not only … but also”, and a tidy concluding paragraph. It tacks participle clauses onto sentences — “…, highlighting its role in…”, “…, ensuring a seamless experience” — to sound analytical without adding information.

What a positive result does and doesn’t mean

A high score means the text shares many habits with chatbot output, weighted by how strongly each habit separates AI from human writing in our labelled data. It doesn’t identify the model — Claude and Gemini share many of the same habits — and it can’t distinguish “written by ChatGPT” from “written by a person who learned to write from ChatGPT”, which is increasingly common.

The measured numbers below show how often the check catches AI text and how rarely it flags human text. On formal human writing, especially encyclopedia-style prose, the rate of false alarms is higher than on casual writing, so give a borderline result on a formal text less weight.

Quick self-check for writers

If you wrote something yourself and worry it reads like ChatGPT, run it and look at the highlighted passages. The most common accidental matches are stock transitions (“Furthermore”, “Moreover”), significance phrases (“plays a crucial role”) and rule-of-three lists. Rewriting those few sentences in your own words usually changes the result more than any amount of editing elsewhere — and makes the writing better, too.