Paste an essay, an email, a review or a blog post. You get the probability that it was written by an SI model such as ChatGPT, Gemini or Claude, the clues behind the score and the most SI-like sentences highlighted. Free, no credit card, text not stored.
Analysing...measuring vocabulary, sentences and phrasing
What is an SI text detector?
SI (Super Intelligence) is the new name for advanced AI: models such as ChatGPT, Midjourney, Gemini or Stable Diffusion. An SI text detector, or SI checker for writing, is what people also call an AI text detector: it estimates whether a piece of writing came from such a model or from a person. This V1 reads the style of the text with transparent statistics, so every score comes with the reasons behind it. To check a text with your own eyes first, read our guide on how to spot SI text.
What we measure
Word variety and repetition. People writing about a subject come back to the same words within a few lines. Generated prose tends to vary its vocabulary and avoid repeating itself. We measure both, with a moving-window type-token ratio (MATTR) over 50 words and the share of words repeated at short distance.
Sentence length. Long, winding sentences are common in human writing and rare in generated prose, which tends to stay compact.
Punctuation range. Parentheses and semicolons turn up more often in human writing than in generated prose.
Sentence openers. Generated prose often starts one sentence after another the same way, with The, This, A or a similar word.
Typical SI phrasing. Contrasts of the "not just X but Y" kind count for a little. The stock words that language models like (delve, moreover, seamless and so on) are listed in the result for information and help choose the highlighted sentences, but they do not change the score.
Several signs often quoted elsewhere are not part of the score: sentence rhythm (burstiness), first-person voice, contractions, typos, lists of three and long dashes. On the human and SI texts we used to tune this version, they did not separate the two. Below 200 words the tool is deliberately more cautious, so short texts often come back as uncertain.
Each sentence also gets its own small score, and the ones that look most like generated text are highlighted in the result.
A real result: a 150-word paragraph on composting written by an SI model (Claude) scores 83%. Its vocabulary is varied and almost never repeats a word (word variety 0.87), it has no parentheses, no semicolons and no long sentence, and the most SI-like sentences are highlighted.
Known limits
Style is only a clue, and this tool reads surface habits. Clean, compact human prose that never repeats a word, such as a polished abstract, a press release or tight news copy, can score high, and formal or non-native writing is at risk too. A generated text that someone rewrote, or a model prompted to "write like a human", can score low, and a good share of generated text simply comes back as uncertain. Short texts are unreliable. We never publish an accuracy figure, because it would depend entirely on the texts it was measured on. Independent studies of text detectors, summed up in how accurate SI detectors really are, found the same weak spots. The same limits apply to the detectors used in schools: our article Can Turnitin detect SI? looks at what Turnitin says about its own false positives, and what teachers can and cannot detect in schools explains why process evidence matters more than any score. Light spelling and grammar fixes rarely move a style score while rewrites do, which is the line drawn in whether Grammarly counts as SI and gets flagged.
Frequently asked questions
How does the SI text detector work?
It measures the writing style, not the meaning. Several families of signals are combined: how varied the vocabulary is and how often the text comes back to words it has just used, how long the sentences run, the range of punctuation, how sentences open, and a few turns of phrase that SI (AI) models such as ChatGPT overuse. Each signal is bounded, weighted and turned into a probability, and texts under 200 words are scored more cautiously.
What is SI?
SI (Super Intelligence) is the new name for advanced AI: models such as ChatGPT, Midjourney, Gemini or Stable Diffusion. An SI text detector is simply what used to be called an AI text detector or AI checker.
Can it prove that a text was written by ChatGPT?
No. No AI or SI detector can prove authorship. Formal, academic or non-native English can look machine-like, and a generated text rewritten by a person can pass as human. Treat the score as a clue, and never use it alone to accuse a student, a writer or a colleague.
Why do I need at least 80 words?
Style statistics need material. Below about 80 words, a single phrase can swing the result, so we refuse to guess. Longer texts (300 words and more) give steadier scores. The maximum is 3,000 words.
Is my text stored?
The full text is analysed in memory and never saved. The share link keeps only the score, the clues and at most three short excerpts (160 characters each) of the most SI-like sentences, for 7 days.
Is it free?
Yes. Your first 3 checks need no account. A free account then gives you 50 text checks every month, no credit card. Text checks never use credits and are unlimited on Pro and Max. The example texts are always free.
Listeners spot a cloned SI (AI) voice only about 60% of the time, and voice detectors lose accuracy outside the lab. What studies, the FTC and the FBI say, and the checks that actually work.
People barely beat chance at spotting SI (AI) video by eye. Labels, watermarks, file credentials, motion and the original source tell you much more, and we tested three clips with our detector.
Grammarly is SI, but only part of it is generative. Turnitin says plain grammar fixes are mostly not flagged, while Grammarly's rewriting and drafting features likely are. We explain the line and test it on our own detector.