How the AI Detector Works
The AI detector scans a text for the signs of AI writing, weighs them against the signs of human writing, and turns the total into an AI score and one of three results. Every sign it uses is listed below, with how often it appears in human and AI text.
The three steps
- Count. The checker searches the text for 12 signs of AI writing and 8 signs of a person typing, and measures how much the sentence lengths vary.
- Weigh. Each sign has a weight, set from 3,379 test texts. Signs that separate AI from human text more sharply count for more. A sign of AI can only raise the score and a typing slip can only lower it.
- Decide. A high score gives "Probably AI", a low score gives "Probably not AI", and everything between gives "Can't tell". The cut-offs are set so that about 1 in 30 human texts is called "Probably AI".
All of this runs in your browser. The text is not uploaded.
Signs of AI writing
The figures show the share of test texts that contain the sign at least once. "Typed human" is essays, emails and forum answers. "Published human" is books, journals, Wikipedia and news, where editors shape the punctuation. "Unedited chatbot" is 160 emails, essays, letters and posts we generated for this test.
| Sign | Typed human | Published human | All AI | Unedited chatbot |
|---|---|---|---|---|
| Em dash (—)There is no key for it. Chatbots use it constantly; people typing by hand almost never do. | 0.5% | 14% | 35% | 65% |
| Markdown bold (**text**)Chat interfaces format answers in Markdown. The asterisks come along when the text is copied. | 0% | 0% | 10% | 53% |
| Markdown heading (# Title)A line starting with # is a Markdown heading, the way chatbots structure longer answers. | 0% | 0.1% | 2% | 5% |
| Bullet listChatbots break answers into bullet points far more readily than people writing a message or an essay. | 2% | 0.4% | 2% | 8% |
| Bracketed placeholder ([Your Name])A blank the chatbot left for the user to fill in. | 0% | 0.1% | 3% | 30% |
| Word AI overusesOrdinary words, but current models use them several times more often than people do. | 9% | 19% | 59% | 78% |
| Stock phraseA ready-made phrase that models reach for. | 3% | 4% | 24% | 32% |
| "Rather than" contrastModels like to define a thing by what it is not. | 3% | 4% | 27% | 10% |
| "Not just X" framingThe same contrast habit: "not just X, but Y". | 10% | 4% | 18% | 12% |
| "It's not X, it's Y" framingA denial followed at once by the correction. | 0.9% | 0.2% | 2% | 0% |
| Set-up phrase before a reveal"Here's the thing:" or "The result?" before the point is made. | 2% | 2% | 5% | 4% |
| List of threeModels group things in threes ("clear, concise, and effective") much more than people do. | 34% | 66% | 81% | 91% |
Signs of a person typing
| Sign | Typed human | Published human | All AI | Unedited chatbot |
|---|---|---|---|---|
| Double hyphen typed as a dash (--)What a person types when they want a dash and have no key for one. | 5% | 5% | 1% | 0% |
| Hyphen typed as a dash ( - )The other keyboard stand-in for a dash. | 10% | 3% | 0.6% | 0% |
| Three typed dots (...)Typed full stops instead of the ellipsis character. | 7% | 2% | 1% | 0% |
| Lowercase "i"Models capitalise it every time. | 12% | 1% | 0.2% | 0% |
| Sentence starting in lowercaseModels capitalise every sentence. | 25% | 18% | 5% | 2% |
| Contraction without its apostrophe (dont)Models do not drop apostrophes. | 15% | 0.3% | 0.2% | 0% |
| Common misspellingModels rarely misspell. | 8% | 0.1% | 0% | 0% |
| Chat slanglol, tbh, gonna and the like. | 3% | 0.2% | 0.2% | 0% |
Measured but not counted
These signs are in the table for the record. They do not affect the score.
| Sign | Typed human | Published human | All AI | Unedited chatbot |
|---|---|---|---|---|
| Chatbot phraseA line addressed to the person who asked, left in by mistake. | 0% | 0% | 0% | 0% |
| Sentence opened with a connective"Moreover," "Furthermore," "Ultimately," at the start of a sentence. | 13% | 14% | 28% | 21% |
| Curly quotation mark or apostrophe | 0.9% | 37% | 47% | 79% |
| Single-character ellipsis (…) | 0.2% | 1% | 0% | 0% |
| Spaced en dash ( – ) | 0.4% | 6% | 0.5% | 2% |
| Two spaces after a sentenceA typing habit; models do not produce it. | 23% | 0% | 5% | 0% |
| No space after punctuationA slip of the fingers. | 9% | 2% | 2% | 0% |
- Curly quotes, the ellipsis character and the en dash are common in AI text, but Word, Google Docs and phone keyboards also insert them into text people type. Our test texts are mostly plain text, so they understate that.
- The other signs in this table were either too rare in the test texts to weigh reliably, or added nothing once the rest were counted. One of the test datasets matches typing quirks in its AI text to the rate at which people produce them, which is a fair reminder that such quirks are easy to fake.
Sentence rhythm
People mix short and long sentences more than models do. The checker measures how much sentence length varies and counts even rhythm as a sign of AI and varied rhythm as a sign of a person.
What this approach cannot see
The checker reads surface habits. It does not understand the text, and it cannot tell a person who writes like a chatbot from a chatbot. Someone who removes the dashes and the Markdown, or asks the model to write plainly, will often get past it. The accuracy page shows how often.
Check a text for AI, free.
Open the AI detectorFigures from our test run of 2026-10-09.