AI detector
Paste a text and read seven signals that are frequent in generated writing, each one measured and explained. The text stays in your browser. No signal is a verdict, and the page says why.
Paste a text
The signal board fills in as you paste. Nothing leaves this page.
What this detector measures
Language models produce text by choosing, word after word, the continuation that is most likely given what came before. Left to themselves they therefore write in a recognisable way: sentences of similar length, a vocabulary that stays in the middle of the road, transitions at the head of sentences, a formal register with no contractions, and a set of words and phrases that they reach for far more often than people do. The seven meters above measure those tendencies directly, on the text you paste, and show the measurement next to each one.
| Signal | What is measured | Flagged when |
|---|---|---|
| Even sentence lengths | Coefficient of variation of sentence length in words | below 0.35 over five sentences or more |
| Narrow vocabulary | Distinct words per 100-word window, averaged | below 60 of 100 |
| Repeated phrases | Share of three-word sequences that occur more than once | above 8 percent |
| Stock phrases | Hits from a list of words and expressions generated text overuses, per 1,000 words | above 6 per 1,000 |
| Uniform paragraphs | Coefficient of variation of paragraph length | below 0.30 over three paragraphs or more |
| Signposting openers | Share of sentences that start with a transition word | above 22 percent |
| No personal or informal marks | Contractions and first-person pronouns per 100 words | below 0.4 per 100, over 150 words or more |
The thresholds are our working values, chosen to flag a pattern when it is pronounced. They are not calibrated against a labelled corpus and we publish no accuracy figure.
What it does not measure, and why that matters
Commercial detectors add two things this page does not: a measure of how predictable each word is to a language model (perplexity), and a classifier trained on millions of human and generated texts. Both see more than seven meters do, and both have the same blind spot, documented in the studies cited below: they confuse "regular" with "generated". Formal, careful, second-language or template-driven writing is regular, so it gets flagged. Edited or paraphrased generated text is irregular, so it gets through. OpenAI withdrew its own classifier in July 2023 over its low accuracy; Vanderbilt University disabled Turnitin's detector a month later, after working out that Turnitin's own claimed false positive rate of 1 percent would have meant around 750 wrongly flagged papers a year on its campus.
This page shows its measurements instead of a probability precisely because a probability would be read as a verdict. A meter that says "sentence lengths vary by 0.28" is a fact about the text you can check by reading it. A badge that says "87 percent AI" is a claim about a person that nothing on this page can support.
How to use the signals well
Paste at least 300 words; under 100 the meters say nothing, and short texts have no rhythm to measure. Read the flagged signals, then read the text with them in mind: does every paragraph open with "Furthermore"? Does the vocabulary never leave the safe middle? Is there a single contraction, a single concrete detail, a single sentence that surprised you? Those are the observations a careful reader makes anyway; the meters only point at them faster. If you are a teacher or an editor, the signals are a prompt for a conversation about the text, not a finding to put in a file. If you are the author, they tell you where your prose has gone flat, which is useful whatever wrote it.
What happens to your text
Nothing leaves this page. The page downloads once; the analysis runs in your browser; the result is drawn on your screen and discarded when you close the tab. There is no account, no upload and no log of what was analysed. That is a design choice, and it is also why there is no "history" feature: we never had the text.
Questions people ask
Is this AI detector free?
Yes. Paste any length of text, as often as you want, with no account and no word limit. The analysis runs in your browser, which is also why it is free: there is no server doing the work.
Can an AI detector prove a text was written by ChatGPT?
No tool can. Detectors measure patterns that are frequent in generated text; a human can produce the same patterns, and a generated text can be edited out of them. OpenAI withdrew its own classifier in July 2023 for its low accuracy, and universities such as Vanderbilt disabled Turnitin's detector the same year. Treat any score, ours included, as a reason to look closer, never as evidence.
What does the regularity index mean?
It is a weighted average of seven measurements: how even the sentence lengths are, how narrow the vocabulary is, how many phrases repeat, how many stock expressions appear, how uniform the paragraphs are, how many sentences open with a transition word, and how few contractions and first-person marks there are. A high index means the text is regular in those ways. Formal human writing is often regular too.
Why does the detector say my own writing looks generated?
Because the signals are patterns, not authorship. Academic prose, legal drafting, technical documentation and text written in a second language are regular and formal by nature, and they score high on several meters. Studies have shown that commercial detectors flag non-native English writers far more often than native ones for exactly this reason.
Is my text sent to a server?
No. The page loads once, then everything happens on your device. We cannot see what you paste, and nothing is stored.
Does this work for languages other than English?
The stock-phrase and opener lists are English, so those two meters are only meaningful for English text. Sentence rhythm, vocabulary variety, repeated phrases and paragraph uniformity work on any language that separates words with spaces.
Sources
- TechCrunch, OpenAI scuttles AI-written text detector over "low rate of accuracy" (the classifier was withdrawn on 20 July 2023), read 25 October 2026.
- Vanderbilt University, Guidance on AI detection and why we are disabling Turnitin's AI detector (16 August 2023), read 25 October 2026.
- Liang, Yuksekgonul, Mao, Wu, Zou, GPT detectors are biased against non-native English writers (arXiv 2304.02819, 2023), read 25 October 2026.