Can you tell for sure if ChatGPT wrote something?

No. Text alone can't prove who wrote it. Signs like leftover chat phrases or placeholder brackets make AI drafting likely, but style signs are shared by many human writers. Drafts, notes and a conversation with the writer are better evidence than any reading of the final text.

What is the most reliable sign that text came from ChatGPT?

Leftover chat phrases ("I hope this helps", "Let me know if you'd like changes") and name or date slots left unfilled in square brackets. They are traces of the chat window rather than style. They are also easy to delete, so their absence proves nothing.

Are AI detectors accurate enough to decide?

Their makers and researchers say not on their own. OpenAI withdrew its own classifier in 2023 for low accuracy, Turnitin says its report should not be the sole basis for action against a student, and a 2023 Stanford study found detectors flagged over half of a set of non-native English essays as AI.

Does one overused word mean ChatGPT wrote it?

No. One word means very little. Wikipedia's guide treats a dense cluster of these words as a stronger sign, and notes that the overused words change over time. Plenty of people use words from those lists naturally.

How do I stop my own writing from looking like ChatGPT?

Use the signs as an editing list. Cut scene setting openers, delete chat phrases, vary sentence length, swap vague words for specifics and break up lists of three. Most of all, add details only you would know.

How to Tell if ChatGPT Wrote It: The Signs Readers Notice

The 10 signs of ChatGPT writing

These signs come mostly from Wikipedia's field guide, Signs of AI writing, which its editors keep with real examples from AI text found on the encyclopedia. The guide itself says the list is descriptive, made of observations rather than rules, and that the patterns are only potential signs of a problem. The "how much it tells you" column below is our editorial read of that guide, not a measured rate.

Sign	What it looks like	How much it tells you
1. Leftover chat phrases	"Sure! Here's a draft", "I hope this helps", "Let me know if you'd like changes"	A lot. People rarely write these to a reader
2. Unfilled template slots	A name, company or date slot left in square brackets instead of filled in	A lot. Someone pasted without reading
3. Scene setting opener	"In the modern world, communication matters more than ever"	Some. Plenty of people write like this too
4. Even sentence lengths	Four or five sentences in a row of about the same length	Little on its own
5. Lists of three	"fast, reliable and affordable" in paragraph after paragraph	Little on its own; more when it repeats
6. A cluster of AI vocabulary	"enhance", "valuable", "fostering", "showcase", "underscore" close together	Some, when many appear together
7. "Serves as" instead of "is"	"The park serves as a vital hub for the area"	Little on its own
8. "Not just X, but Y"	"It's not just a product, it's a movement"	Little on its own
9. Big claims with no specifics	"reflecting broader trends", "a lasting legacy", no names or numbers	Some, especially in a short text
10. Heavy formatting in plain prose	Bold phrases mid sentence, title case headings, bullet lists with bold labels	Some, in a context where people rarely format

Notice what is missing: spelling mistakes. ChatGPT drafts are usually clean, so "no typos" is not a sign of anything.

Leftover chat phrases and placeholders are the strongest signs

The first two rows are different from the rest. They are not style choices, they are traces of a chat window. Wikipedia's guide calls the first group collaborative communication: phrases like "I hope this helps", "Of course!" and "Would you like..." that were addressed to the person chatting, not to the eventual reader. Placeholders in square brackets are the same kind of trace. Both are easy to fix, which is also why their absence tells you little.

Style signs are weak alone

Rows 3 to 10 are habits, and people share them. A student taught the five paragraph essay may write in threes out of habit. Marketing copy leans on "not just X, but Y" too. Treat a style sign as a reason to read more closely, never as an answer.

Which words and phrases does ChatGPT overuse?

Wikipedia's editors keep a list of AI vocabulary words, backed by published studies, that started appearing far more often in text after 2022, when chatbots became widely available. It includes delve, "tapestry", "testament", "pivotal", "underscore", "enhance", "fostering", "showcase", "valuable" and "vibrant". The guide notes that one or two of these words may be coincidence, while lots of them, lots of times, is among the strongest tells.

Two catches make word lists less useful than they look:

The list keeps moving. The same Wikipedia guide notes that delve was heavily overused by ChatGPT in 2023 and early 2024 and then dropped off sharply in 2025. A word list from two years ago describes older drafts.
Words travel. The guide also points to research suggesting people's own writing and speech are picking up chatbot vocabulary. A student who reads a lot of AI text may start writing "showcase" naturally.

So look for density, not single words. A paragraph with five of these words and no names, dates or numbers reads very differently from one that uses "valuable" once.

Why do AI detectors disagree with each other?

Most text detectors look at how predictable the wording is. Researchers call this perplexity: low perplexity means each next word was easy to guess. Chatbot text tends to be predictable. So does a lot of careful human writing.

That is why detectors trip over the same writers again and again. A Stanford study, Liang et al. (2023), "GPT detectors are biased against non-native English writers", ran seven widely used detectors on 91 TOEFL essays written by people and 88 US eighth grade essays. The detectors were near perfect on the US essays but misclassified over half of the TOEFL essays as AI generated, with an average false positive rate of 61.22%. The authors link this to the lower perplexity of writing with a more limited vocabulary.

OpenAI reached a similar conclusion about its own tool. Its AI classifier announcement from January 2023 reported that on a challenge set the classifier caught 26% of AI written text while labelling human text as AI 9% of the time, and a note added later says the classifier was withdrawn on July 20, 2023 due to its low rate of accuracy. OpenAI's current educator FAQ answers "Do AI detectors work?" with "In short, no, not in our experience", and adds that its own attempt labelled Shakespeare and the Declaration of Independence as AI generated.

Commercial detectors publish their own guidance too. Turnitin's help guide on AI writing detection in the classic report view says its model may not always be accurate and should not be used as the sole basis for adverse actions against a student. If you want the detail on what that report shows, read our guide to whether Turnitin detects ChatGPT.

How to use these signs fairly

The signs are useful for two jobs: checking your own drafts, and deciding when to ask someone a question. They are not useful for accusing anyone.

If you are reading someone else's work:

Look for several signs together, not one. A single list of three means nothing.
Weigh the writer. The Liang et al. study above is a reminder that people writing in a second language are the most likely to be wrongly flagged.
Ask about the process, not the verdict. Drafts, notes, version history and a short conversation about the argument tell you far more than the text does.
Remember the rules may allow it. Plenty of workplaces and some courses permit AI drafting with disclosure.

Even Wikipedia's guide warns its own editors not to rely too much on their judgment, citing a 2025 study that found people's ability to tell chatbot text from human text was no better than random chance. Humility is part of using the list well.

If you are checking your own draft, the table doubles as an editing list. Our guide on how to make ChatGPT sound more human turns each sign into a fix with a before and after. For a quick first pass, the ChatGPT to human text converter trims filler and varies sentence length, or you can try the free ChatGPT humanizer on a paragraph and compare. Then do the last read yourself.