ChatGPT AI Detector

Use this free ChatGPT detector to check whether a piece of English writing shows the patterns typical of ChatGPT and other AI chatbots. Paste 100 words or more into the checker above: there is no sign-up, and your text is not kept.

This guide explains what the result can and cannot tell you, the signs worth looking for yourself, how well the checker caught ChatGPT text in our own sealed test, and what to do next if you are a teacher, an editor or a writer whose work was flagged.

Can you tell if something was written by ChatGPT?

You can tell whether a text reads like AI writing, and a good ChatGPT checker can measure that far more consistently than a hunch. What nobody can do from the text alone is prove that ChatGPT wrote it. ChatGPT, other chatbots and some careful human writers can all produce the same smooth, even style.

So the honest answer to “is this ChatGPT?” has two parts: a measured signal from the detector, and the context only you have, such as the writer’s drafts, their usual style and how the text was produced. Used together, they give you a fair basis for a conversation. Used alone, a score is not proof.

How to check if a text was written by ChatGPT

1. Paste the full text. Use the whole essay, article or answer, not only the lines that made you suspicious. The minimum is 100 words; 150 words or more gives a more reliable result.
2. Run the check and read the zone first. Likely human-written, Uncertain or Likely AI-written tells you where the text sits; the 1–99 score and the stability range tell you how strongly.
3. Read the evidence line. Each result shows how often human writing of the same length crossed the AI line in our sealed test, so you know how much weight the result can carry.
4. Look at the writing yourself. Compare the text with the signs below and with other work by the same person.
5. Talk before you act. Ask about drafts, notes or sources. Save the result as a PDF if you need a record.

Two features help with harder cases. You can compare up to 3 passages side by side, for example a suspect essay next to two earlier pieces by the same student. And if a passage is short, adding more of the same text usually gives a clearer answer than checking a fragment twice.

Signs of ChatGPT writing to look for

Experienced teachers and editors often notice the same habits in unedited chatbot text. None of them proves anything on its own, because people write this way too, but several together are a reason to look closer:

• A broad opening and a tidy summary. The first lines set a general scene, and the last paragraph repeats what was already said without adding anything.
• Even rhythm. Sentences of similar length, paragraphs of similar shape, and transitions that are smooth but carry no real argument.
• Balance without a view. Both sides of every question are covered politely, and the text never commits to a position the writer would have to defend.
• Neat structures. Points that always come in threes, parallel phrasing and headings or bullet points where a person would simply have written prose.
• Few real specifics. No names, dates, figures or personal experience, or specifics that turn out to be wrong when you check them, such as sources that do not exist.
• A voice that does not match. The text sounds unlike the writer’s other work: a different level of vocabulary, different spelling conventions or a sudden absence of their usual mistakes.

These cues are easy to over-read, which is exactly why a measured check helps. The result card also lists a few plain observations about the text, such as how much concrete detail it contains; they are notes for your own review, not the reasons for the score. For a fuller list with examples, read signs of AI-generated text versus human writing.

How a ChatGPT detector works

Many early AI detectors relied on simple statistics, such as how predictable each word is. This checker uses engine V4, a fine-tuned transformer text classifier: a model that learned, from large sets of human and AI-written text, the differences between the two. It was trained once and frozen, and nothing is tuned on live traffic.

When you submit a text, the engine reads it in overlapping windows and averages them into one value, which is then shown as an AI-writing signal score from 1 to 99. It judges the passage as a whole; it does not mark individual sentences, it does not compare your text with other sources the way a plagiarism checker does, and it never names the model that wrote a text.

The score is not the share of the text written by AI, and it is not a probability. A result alone does not prove who wrote a text.

Because it measures AI-writing patterns in general, the same check works on text from Claude, Gemini, DeepSeek and other chatbots. See how the detector works for the full picture, and AI detection versus plagiarism checking if you need both.

How to read your result

Inside Likely AI-written, a score of 65–84 carries a “Moderate signal” chip and 85–99 a “Strong signal” chip. The AI line is 65 at every text length: each length band has its own frozen internal threshold, and the score is mapped so that 65 is that threshold.

The stability range shows how far the score moves when the source families behind the score map are resampled; the zone itself is decided by the frozen threshold. The evidence line under the result is the most useful part for a decision: it tells you how often human writing of your text’s length reached 65 or more in the sealed test. The guide to understanding results walks through every part of the card.

How V4 was trained on ChatGPT text, and how well it catches it

ChatGPT writing was part of V4’s training data, alongside eight other model families. We then tested the checker the way a sceptical reader would want: on new texts it had never seen, scored once, after every threshold was fixed. In that sealed test of 17 October 2026, all 525 ChatGPT texts of 150 words or more reached the AI line, and so did all 198 shorter texts of 100 to 149 words. The previous engine, V3, caught 65.71% of the same texts.

Catching AI text is only half of the job; the other half is leaving human writing alone. Of 2,863 human texts of 150 words or more, 4 were wrongly flagged, about 1 in 700. Short human texts are harder: at 100 to 149 words, about 1 in 100 crossed the line.

Two development tests add context. When ChatGPT was left out of training altogether, 99.22% of its texts were still caught, so the result does not depend on the engine having seen ChatGPT before. And of ten text attacks tried on all model families together, synonym substitution did the most damage and still left 97.40% caught.

What we tested

The ChatGPT texts were written between 10 and 13 October 2026 in the ChatGPT app, not through an API, by four models chosen by hand. They cover 35 task types, from essays, emails and cover letters to product reviews, news articles, short stories, forum answers and summaries, in 6 writing styles. Half used topics from a broad list and half were based on human seed texts or rewrite tasks; every other model family got the same tasks. Reasoning text was discarded, only the final answers were scored, and near-duplicates were removed first.

The evidence page has every family, length band and human source, and the methodology explains how the test was set up.

One real example

A development text, not a sealed one: ChatGPT wrote this discussion-board post of 208 words; it scored 96 with a stability range of 93–99 and the zone Likely AI-written. Its first 71 words:

The minutes describe the decision as unanimous, but the recorded procedure appears to have involved an invitation for objections rather than an explicit vote. I suggest that our analysis preserve this distinction instead of treating the two forms of decision as automatically equivalent. An absence of stated objection may indicate agreement, but it can also reflect uncertainty, deference, or a participant’s understanding that silence is sufficient for the meeting to proceed.

Edited, paraphrased and “humanised” ChatGPT text

Real texts are rarely pure ChatGPT output. People edit drafts, mix in their own paragraphs or run the text through a paraphrasing or “humaniser” tool. All of these change the AI-writing signal, and the more a text has been reworked, the less any single score can say about how it began.

We measured the effect of automated rewriting on development data: across ten attacks, synonym substitution caused the largest drop and still left 97.40% of AI texts caught. That is a useful sign of resilience, not a promise about every edited text. We do not publish ways to get around the detector. If you used ChatGPT for part of your work, saying so openly is worth more than any score.

When human writing is flagged

Any ChatGPT checker can mistake human writing for AI writing, and the risk is not the same for everyone. In our sealed test it was highest for short texts, for learners of English and for formal speeches:

• Short texts: about 1 in 100 human texts of 100–149 words reached the AI line, against about 1 in 1,000 at 150–299 words, about 1 in 300 at 300–599 words and none of 970 at 600 words or more.
• Learners of English: 22 of 1,641 learner texts (1.34%), about 3 times the rate for native writing at the same lengths, and 15 of 572 (2.62%) at 100–149 words. No Arabic-first-language learner writing was available to test.
• Formal speeches: 17 of 931 recent parliamentary, congressional and central-bank speeches (1.83%), the one kind of writing where V4 did worse than the previous engine.

If your own writing was flagged, gather what shows how you wrote it: drafts, version history, notes and sources. Offer a longer sample and explain your process. The guide if your text was flagged takes you through it step by step.

Who uses a ChatGPT checker

• Teachers and lecturers checking essays and assignments, ideally as the start of a conversation with the student. See checking essays and research papers.
• Editors and site owners reviewing freelance articles before they publish, alongside fact-checking and their own editing. The guide on using AI detectors for blog editing covers a sensible workflow.
• Managers and recruiters reading reports, cover letters or written tasks. See checking business reports.
• Writers and students checking their own work before they submit it, so a surprise flag can be explained in advance.

Limits of a ChatGPT AI detector

The checker cannot attribute a text to ChatGPT or to any other model. It works on English only; for other languages see language support. Short, heavily edited or mixed texts carry less information, and models released after the test are not covered by the results above. All of these figures describe our test set, not a promise about every text you submit.

Using a ChatGPT detector responsibly

Decide in advance what use of AI is allowed in the assignment or commission, so a result is measured against a clear rule. When a text is flagged, talk to the writer first and give them room to explain. Never fail a student, reject an article or make an accusation on a score alone, and take extra care with short texts, learners of English and formal writing.

ChatGPT AI detector FAQ

Related guides

The same checker covers every chatbot we tested. See all model guides, or go straight to Claude, Gemini, DeepSeek, Kimi, Qwen, GLM, MiniMax and Llama. Copilot can run models from more than one provider; its guide explains what that means for a result.

Tested and written by the AI Detector Checker team. Last tested: 17 October 2026.