Paste text — we estimate AI-written likelihood from style, repetition and clichés, with no model. A rough guide, not proof.
Free with no sign-up, and the check runs inside your browser, so what you enter never leaves your device. The result is a signal to help you decide, not proof for money or legal decisions.
Estimate whether a passage was written by a person or shows the tell-tale statistical fingerprint of generative AI such as ChatGPT or Gemini. Everything runs inside your browser — your text is never uploaded — so even sensitive writing is safe to check.
First, sentence-length rhythm (burstiness). People mix short and long sentences; AI text often flattens into an even, uniform length. Second, lexical and structural repetition: the same connectors ("furthermore", "in conclusion") or sentence shapes recurring on a schedule raises the odds of machine generation.
Third, clichés — bland, hollow phrases ("in today's fast-changing world", "plays a crucial role in") clustered together. Fourth, excessive formality and uniformity: writing with no typos, no colloquialisms and no emotional variance reads more like a model than a human.
The score runs 0–100% as a relative "likelihood it was AI-written". Roughly, above 65% is high, 35–65% is uncertain, below 35% is low. Each signal bar shows how much that factor contributed.
With very little text (two or three sentences), the statistics are unstable and a "low confidence" warning appears. Treat the score as a hint only. A full paragraph, ideally five or more sentences, gives a far steadier read.
Style-statistics detection has inherent limits. A carefully formal human report can be flagged as AI, while an AI draft lightly edited by a person, or output from the newest models, can slip through as human. Machine-translated text also skews the numbers.
So treat the result as a starting point for suspicion, not proof. Read for the things only a human can catch — factual errors, unsupported claims, missing context — and never use this as sole evidence in academic or hiring decisions.
This limit is not particular to this tool. OpenAI withdrew its own AI text classifier in 2023 on the grounds of low accuracy — and a model vendor judging its own output is in the most favourable position there is. Text, unlike images, leaves no physical trace like pixel statistics; all that remains are tendencies of style.
The more serious problem is that the errors are not evenly distributed. In an experiment published by Stanford researchers in 2023, writing by people whose first language is not English was misclassified as AI by several commercial detectors at a markedly higher rate. Simpler vocabulary and more regular sentence construction read as machine-like. Writing in a second language, formal official documents and beginner coursework all lose out for the same reason.
So this tool reports observations rather than verdicts. It shows which signals appeared and how strongly, and leaves the conclusion to a person. Lift the score out on its own as evidence and you carry that bias along with it.
An English original was run through machine translation. The sentences read smoothly, but particle choices were uniform and each paragraph reused the same connective structure. The score sits mid-range with only the repetition signal running long. Translation flattens style, so its statistics overlap with generated text. What this range calls for is not an AI verdict but a question about provenance — asking where the original is gets there faster.
A draft announcement was pasted in. Sentence lengths clustered tightly around the same count, and connectives like "moreover", "in conclusion" and "plays an important role in" recurred on a regular beat in every paragraph. The pattern score lands in the high band with long bars for burstiness and cliché density. The correct reading is not "this person lied" but "the first draft was probably model output, so the facts and figures need a human check".
A personal blog post was pasted in. One sentence ran four words, the next ran forty, and spoken-language asides and specific prices and dates were mixed throughout. The score drops into the low band. Concrete lived detail and irregular sentence rhythm are statistically human, and irregularity is exactly what this tool counts.
Academic abstracts, government reports and manuals — orderly, unemotional, typo-free — score high even when a person wrote every word. In our own gate set of 181 human-written texts, 12.7% were classed as "many patterns". That is roughly one in eight people getting an unfair result. This score can never justify an academic penalty or a hiring rejection; treat it as a place to start reading, not a verdict.
No. Analysis happens entirely in your browser and the text is never uploaded to a server.
There is no absolute threshold. A high score only means the writing is strongly "AI-like"; always combine it with human review.
Signals like sentence length and repetition are largely language-independent and generally work, but the cliché dictionary is tuned for Korean and English.
Please do not. As above, second-language and formal writing are structurally disadvantaged. Where plagiarism or misuse is genuinely at issue, the process — drafts, revision history, whether the author can explain their own reasoning — is what holds up, not a percentage.
It depends how far you went. Rewriting sentences and adding specifics and figures reads as human; changing only the wording leaves the original structure and scores high. Editing to lower the score is beside the point. What a reader needs is verified substance, not a low number.