Claude watermark & provenance checker
Check text for the things that can actually be measured: hidden Unicode, formatting artefacts, look-alike characters and statistical patterns in the writing - with no claim to a verdict nobody can give.
What people use it for.
What is actually measured.
The text is read in your browser and reported on in four ways. Every character is mapped and classified, so invisible characters - zero-width marks, tag characters, direction controls - are listed by name and code point where they sit. Characters that imitate Latin letters, and words that mix Latin with Cyrillic or Greek, are flagged separately. Whitespace and formatting artefacts are counted: runs of spaces, trailing spaces, more than two blank lines, smart quotation marks. Finally the writing itself is measured - sentence length and its variation, repeated bigrams and trigrams, lexical diversity, Shannon entropy, punctuation density - and a small number of style signals are raised from those numbers, each one labelled with the evidence behind it and its own confidence. Values worth protecting, such as numbers, dates and citations, are extracted and listed so you can see what any later cleaning must not touch.
Where it stops: This cannot tell you whether a text was written by AI, and it does not try. There is no reliable detector for that, and the ones that claim to be wrong often enough to ruin people - so no score, no percentage and no verdict is produced here. The statistical signals describe the writing, not its author: uniform sentence lengths and formulaic transitions are common in careful human prose too, and are reported as observations with their evidence attached. Provider attribution is never inferred from style. Provider-specific notes are heuristic, and no proprietary watermark is claimed to be detected or removed unless an authoritative verifier is configured. Statistical watermarks such as SynthID Text live in word choice rather than characters, and no character-level tool can see them.
Common questions.
Can this tell me whether text was written by AI?
No, and any tool that says it can is overstating what is possible. There is no dependable way to decide that from the text alone. What this page gives you is what can be measured - hidden characters, artefacts, and statistics about the writing - with the evidence shown so you can judge it yourself.
What is a hidden Unicode watermark?
Invisible characters placed in text that survive copy and paste: zero-width marks between words, or Unicode tag characters, which are invisible copies of ordinary letters and can spell out an instruction a model reads and you never see. Those are characters, so they can be found exactly and removed.
Why does it not give me a percentage?
Because a percentage would be invented. The underlying signals are weak and overlap heavily with ordinary human writing, and presenting them as a score gets people accused of things they did not do. Each signal is shown with its evidence and its own confidence instead.
Does it detect Claude, ChatGPT or Gemini specifically?
No. Provider is your selection, not an inference, and notes for a provider are heuristic. Attribution is never guessed from writing style.
Can it remove what it finds?
This page reports. To remove invisible characters from text use the hidden character remover, and for a PDF use Sanitize PDF, which cleans the document and keeps it a PDF.
Is my text uploaded?
No. Every check runs in your browser and the text is never sent to a server.
Found something? Remove it.
This page reports what is in the text. To take the invisible characters out, use the remover, which shows each one before it goes and checks the result afterwards.