Skip to content
AI Watermark Remove

Check anything for AI watermarks

A checker that shows its working. Every finding is a named character or a named block of data. Each comes with a count and a place in the file. Never a made-up confidence score.

0 characters
More options
Checking is free: 5 texts and 2 images a day, no signup.

Checking your text

    Why this reports characters instead of a percentage

    Most tools in this space give you a number: "87% likely AI-generated". That number is generated by a classifier that was trained on some corpus and has no way to tell you why it produced 87 rather than 43. It cannot be reproduced or audited, and it is wrong often enough that several universities have stopped treating it as evidence.

    This tool reports different things, because different things are knowable.

    A zero-width space at codepoint offset 412 either is there or is not. So is a C2PA manifest in an APP11 segment. These are facts about the bytes in your file. You can verify them in a hex editor, and re-running the check after cleaning gives you a definite zero.

    For writing style, which genuinely is a matter of degree, there is a separate score. It is clearly labelled as a heuristic, and every signal that contributed to it is listed so you can judge whether you agree. It measures things a human grader would notice, and it never claims to be a watermark reading.

    The two confidence levels

    Findings are marked probable or informational, and the distinction matters.

    Probable covers zero-width characters, tag characters, free-floating variation selectors, private-use codepoints, bidi controls in Latin prose. None of these has a legitimate reason to be in ordinary text. Their presence means something inserted them.

    Informational covers space homoglyphs: non-breaking spaces, narrow no-break spaces, thin spaces. These appear constantly as ordinary formatting artefacts from web pages, PDFs, word processors. Treating them as evidence of AI generation would be overreaching, so we report them, remove them if you want, and do not pretend they prove anything.

    That separation is the difference between a report you can act on and a report that just makes everything look alarming.

    What a clean result actually means

    If the checker finds nothing, that is a real result, and a useful one: your text contains no invisible Unicode carriers and your image contains no provenance metadata. Both statements are verifiable.

    It does not mean the content was not AI-generated, and the tool does not claim it does. A statistical watermark like SynthID would be entirely invisible to this check, because there is nothing in the character stream to see. Neither would a pixel-domain image watermark.

    Knowing precisely which question you have answered is more useful than a number that gestures at all of them and answers none.

    Questions

    Is this an AI detector?
    No, and the distinction is deliberate. An AI detector guesses whether a human or a model wrote something. This reports what characters and metadata are actually in your file. The first is unreliable; the second is verifiable.
    What does the writing-style score mean?
    It counts surface patterns common in unedited assistant output: em-dash density, stock phrases, unusually even sentence lengths, curly quotes. Every contributing signal is listed. It is a "here is what a reader would notice" measure, not a probability that text is AI-generated.
    Can I check a whole document?
    Two thousand characters per free check. A pass goes up to two hundred thousand, which covers most dissertations in one go, and gives you the cleaned text.
    Does checking count against my quota even if nothing is found?
    Yes, a check is a check. The free allowance is three text checks and one image per day, resetting at midnight UTC.

    Related tools