IS THIS AI GENERATED

No, and it is built specifically not to be one. It is a stylistic tell counter. It measures 48 observable properties of a piece of text, such as em dash density, sentence length variance and the presence of set phrases, and reports how many of them match patterns common in language model output. It never outputs a verdict on authorship, because no tool can determine authorship reliably.

Not accurate enough for the decisions people make with them. OpenAI withdrew its own AI Text Classifier in July 2023, citing its low rate of accuracy. The deeper problem is that the errors are not evenly distributed: a 2023 Stanford study found that GPT detectors classified more than half of a sample of essays written by non-native English speakers as AI generated, while essays by native speakers were classified correctly almost every time. That is the reason this tool reports evidence instead of a verdict.

Because of what they actually measure. Most detectors score text on how predictable its word choices are, a property related to perplexity. Someone writing in their second language tends to use a smaller and more regular vocabulary, which produces exactly the low-perplexity signature that detectors treat as machine generated. The bias is structural, not a bug that a better threshold would fix, and it is why using a detector score as evidence against a student is unsafe.

It is the single most recognisable tell, but on its own it proves nothing. The reason it works as a signal is mechanical: on a standard keyboard the em dash needs a key combination most people do not know, so typical hand-typed prose contains between zero and about 0.5 of them per thousand words. Language models produce them freely. This tool flags em dash density above 1.5 per thousand words and treats above 4 as a strong signal. Plenty of professional writers and anyone typing on a Mac with smart substitution enabled will exceed that honestly.

The recurring single words include delve, tapestry, testament, underscore, pivotal, crucial, robust, seamless, leverage, foster, showcase, meticulous, intricate, myriad, realm, landscape, cornerstone and catalyst. The recurring phrases include "it is important to note", "it is worth noting", "in today's fast-paced world", "navigate the complexities", "a testament to", "plays a crucial role" and "in conclusion". The strongest structural pattern is the negation contrast frame, which takes the shape "it is not just X, it is Y". This tool counts how many distinct ones appear rather than raw hits, so one writer with a favourite word does not trip it.

Burstiness is the variation in sentence length across a piece of writing, measured as the coefficient of variation. People write a long sentence, then a short one, then a fragment, so human prose typically lands between 0.45 and 0.75. Language model output settles into a narrow band and typically lands between 0.20 and 0.40. This tool flags anything below 0.38 and needs at least 15 sentences before it will measure it at all. It is the strongest statistical signal in the set, and it is also the easiest one to defeat by editing.

No. Nothing can prove that, and this tool is designed to make that limitation visible rather than hide it behind a percentage. A high tell density means the text contains patterns that are common in generated output. Careful, formal, heavily edited human writing produces the same patterns. If you are deciding something that affects a person, the output of any detector, including this one, is not sufficient grounds, and the false positives fall disproportionately on people who already get the worst of it.

No. There is no server component. The entire analysis runs in a Web Worker inside your browser tab, on your own device. Nothing is transmitted, nothing is logged, and there is no account. You can disconnect from the network after the page loads and the tool still works. This is the main practical difference from hosted detectors, which require you to send them the document you are suspicious about.

It is free, there is no account, no email and no usage limit. Because the analysis runs on your own machine, there is no cost per check to recover.

At least 50 words to run at all. Between 50 and 149 words it runs only the exact-match signals, phrases and punctuation, and suppresses the numeric score, because statistical measures on a short passage are noise. From 150 words up it produces a score, and confidence in the statistical signals keeps rising until about 400 words. Longer is better. The upper limit is 50,000 characters.

No. Every threshold in it is calibrated on English prose, and applying them to another language produces meaningless numbers. The tool runs language detection first and shows a banner telling you the analysis does not apply if the text is not English.

Easily, and that is worth saying plainly. Replacing the em dashes, varying the sentence lengths and cutting the stock phrases will drop the score substantially in a few minutes of editing. Any tool of this kind measures surface style, and surface style is the easiest thing about a text to change. Treat a low score as weak evidence and a high score as a prompt to read more carefully, not as a result.

Usually because it is formal, edited, or both. Clean mechanics, consistent spelling, uniform paragraph lengths and a steady reading level are all properties of careful writing, and they are also properties of generated writing. The tool reports them as tells because that is what they are, and it shows a caveat when the text reads as formal. It is not evidence that you did anything.

It is a tell density from 0 to 100, not a probability. It is the weighted sum of every signal that fired, with signals pointing toward human authorship subtracting from the total, run through a saturating curve so that three strong tells already read as many. A 64 does not mean a 64 percent chance the text was generated. It means the text is dense with the patterns listed underneath it, and those patterns are what you should be reading.