unmark

AI watermark checker

See what’s hiding in a piece of text: invisible characters that can fingerprint it, the punctuation and phrasing AI models lean on — and a straight answer about the statistical watermarks no public tool can see. Free, and it never leaves your browser.

Text to check0 chars

Updated .

What this checker finds

  • Hidden characters — zero-width spaces and joiners, byte-order marks, bidirectional controls, Unicode tag characters, private-use code points and look-alike spaces. Each is listed with its code point, how often it appears and where it first occurs. This part is definitive: if it’s listed, it’s in your text. Load-bearing marks, like the joiners inside emoji, are recognised and left alone.
  • Microsoft’s Windows AI watermark — the one AI watermark made of characters. Text from the Windows AI text APIs can carry the sequence U+200B U+034F U+2060 U+200B; the checker names it when it finds it.
  • AI punctuation — em dashes, curly quotes and the single-character ellipsis. Chat models use them far more than most people type them, so they read as a tell.
  • Formulaic phrases — openers and fillers such as “delve into” or “in today’s fast-paced world” that show up again and again in AI output.

What no public tool can detect

Most watermarks the big AI providers use are statistical: Google’s SynthID (Gemini), OpenAI’s textGrain (ChatGPT and Codex in the EU) and Anthropic’s Claude watermark all shape which words the model picks, leaving a pattern only a detector holding the provider’s key can measure. None of those detectors is public. OpenAI limits its detector to approved researchers and expert organisations; Anthropic’s detection API is in private preview for eligible organisations; Google’s public SynthID Detector checks images, video and audio, not text.

So treat any site that claims to “detect the ChatGPT watermark” in your text with suspicion — it has no way to. What you can do is remove it: a thorough rewrite breaks the pattern, and the facts for each provider are on the ChatGPT, Gemini and Claude pages, with sources.

If the checker flags something

  1. Hidden characters: copy the text without them using the button above, or send it to the tool — they are stripped automatically.
  2. A possible statistical watermark: if the text came from Gemini, Claude or ChatGPT in the EU, send it to the tool and run a rewrite. Aggressive and Aggressive+ change the most wording.
  3. AI punctuation and phrases: the tool can swap dashes, quotes and ellipses for plain ones, and its rewrite replaces formulaic phrasing.

FAQ

Can this tell whether text was written by AI?+
No — and no tool can do that reliably. The checker reports concrete things: hidden characters (which are definitely there or not), plus punctuation and phrasing that AI models lean on. Those are tells, not proof; plenty of people write with em dashes.
Can I check text for the ChatGPT, Gemini or Claude watermark?+
Not with any public tool. OpenAI’s textGrain detector is limited to approved researchers and expert organisations, Anthropic’s detection API is in private preview for eligible organisations, and Google’s public SynthID Detector checks images, video and audio but not text. A site claiming to detect those watermarks in your text cannot actually do it.
Which AI watermark can this checker detect?+
Microsoft’s. Since October 2026 the Windows AI text APIs can insert the invisible sequence U+200B U+034F U+2060 U+200B into generated text; the checker recognises that exact sequence and says how many copies it found. Stripping hidden characters removes it completely.
If the checker finds nothing, is my text watermark-free?+
Not necessarily. A clean result rules out hidden characters, but a statistical watermark lives in which words were chosen and is invisible to any character scan. If the text came from a watermarking model, a rewrite is the way to remove it.
Do hidden characters mean the text came from AI?+
Not on their own. Zero-width and other invisible characters also arrive through copy-paste from web pages, PDFs and editors, and some tracking schemes add them on purpose. Either way they can fingerprint the text, so they are worth stripping.
Is my text uploaded?+
No. The checker runs entirely in your browser, including opening .docx files — nothing is sent anywhere unless you choose to rewrite the text in the tool.

Related guides