Free browser-based tool that highlights AI writing markers - invisible characters, em dash habits, assistant phrasing and flat sentence rhythm - and explains each one.
This analyzer highlights stylistic markers that appear more often in text generated by large language models than in general human writing. It shows you every match, explains why each one is a signal, and — just as importantly — tells you what else legitimately produces it. Everything runs in your browser; no text is uploaded.
It is a highlighter, not a judge. It does not output a verdict, and it must never be used as evidence that a particular person did or did not write something.
Vendors including Anthropic and Google have described statistical token-level watermarking, where the model is nudged toward a pseudorandomly chosen subset of the vocabulary at each step. Detection requires the secret key used to derive that subset. Because the key is held by the vendor, no third-party tool can verify a statistical watermark, and any tool claiming to do so is measuring something else. What this analyzer measures is writing style.
Beyond word matching, the tool measures three properties of the text as a whole. The most useful is burstiness: the coefficient of variation of sentence length. Human writers mix very short sentences with long ones, while sampling from a probability distribution tends to smooth that out. It also measures paragraph-length uniformity and contraction rate.
Crucially, these structural signals corroborate rather than accuse. Formal technical writing legitimately has flat sentence rhythm, even paragraphs and no contractions — a human-written incident postmortem trips all three. So structural weight is scaled down sharply when no vocabulary or phrasing markers support it. This is exactly the failure mode that causes commercial detectors to falsely accuse non-native English speakers and technical authors.
Light editing, paraphrasing, translation, or mixing with your own writing removes most markers. A low score means there is nothing here to point at — not that a human wrote it. Below roughly 120 words the tool refuses to produce a score at all, because short passages cannot support a statistical claim without generating false positives.
Where a flagged word or phrase has a plainer alternative, the tool offers it: leverage becomes use, delve into becomes examine, and ritual filler like it’s worth noting that is simply deleted. These are ordinary copy-editing swaps that improve the writing on its own merits. They also demonstrate the point: a score that collapses after a handful of word substitutions is not something anyone should be treating as proof.