AI Watermark Checker
Paste any piece of text and this page tells you exactly which invisible characters are hiding inside it, and exactly where — without changing a single one of them. Think of it as an invisible characters detector: read the report first, then decide what to do about what it finds. To actually take those characters out, the cleaner does that in one click; this page only looks.
Checking happens in your browser: nothing is uploaded, and nothing here changes your text — this page only detects. To remove what it finds, use the cleaner instead.
How the checker works
- Paste your text into the box above. There is no button to press first — the check runs the moment you stop typing, on your own machine.
- Read the count and the table. One line gives the total ("Found 4 hidden characters in 180 characters"), and the table underneath names every type it found — Zero Width Space, Word Joiner, Right-to-Left Mark — with how many of each.
- Look at the preview. Every character the checker flagged is shown in
place, written out as its code point (
U+200B,U+FEFFand so on), so you can see exactly where in your sentence it sits.
Is this a ChatGPT watermark detector?
People searching for a chatgpt watermark detector or a chatgpt watermark checker are usually hoping for something that can prove a piece of text came out of ChatGPT specifically. This page cannot do that, and no character-level checker honestly can — none of what it finds is unique to one assistant. What it can tell you, precisely, is which invisible and private-use characters this exact piece of text contains right now: the zero-width spaces, direction marks and stray formatting that survive a copy-paste out of ChatGPT, Claude, Gemini, Copilot, Grok, DeepSeek or Perplexity alike, or out of a PDF, a Word document or a web page that never went near an AI at all. That is a narrower claim than "watermark detector" usually implies, and it is the one this page can actually back up with a table you can read yourself, not a black-box score.
What an invisible characters detector actually finds
Every character it checks for falls into one of a few families: zero-width characters that glue or split words without showing anything on screen, Unicode tag characters that have no visible form at all, direction marks and isolates that can make text render in an order it was not written in, variation selectors and the byte-order mark, and the private-use block reserved for custom icon fonts. None of these are rare — a paste out of a chat assistant, a PDF or Word usually carries two or three of them without anyone intending it. The full list, with what each one actually does and where it typically comes from, lives on the invisible characters page; this one is built to show you what is in front of you right now, not to teach the whole catalogue.
Why check before you clean
Most of the time you will want to press Clean and move on — the cleaner does that job in one click and this checker is not needed for it. Checking first is worth doing when you want to see the problem before you fix it: a teacher or editor curious what a submission actually contains, a developer trying to work out why a pasted string breaks a regular expression, or anyone who just wants to know what is really in a block of text before deciding whether it matters. The checker changes nothing, so you can paste something sensitive, read the report, and close the tab without having produced any output to worry about.
Questions
What does an invisible characters detector actually find?
Zero-width spaces and joiners, Unicode tag characters, direction marks and isolates, variation selectors, the byte-order mark, soft hyphens and the private-use block — the characters that change what a computer reads without changing what your eyes see, and that are not doing a real spelling job (a joiner holding an emoji together, or one that is part of Persian spelling, is left alone). The table under the box names each type by name and counts how many of each are present.
Is this a ChatGPT watermark detector or a ChatGPT watermark checker?
It works as both, in the narrow sense that it actually delivers: it detects invisible characters in text, wherever that text came from. It cannot tell you that a piece of text specifically came from ChatGPT, because none of the characters it finds are unique to one assistant — Claude, Gemini and a plain PDF leave behind the same kinds of debris.
Does it work on text from Claude, Gemini or other assistants?
Yes. The checker matches characters, not sources, so text from ChatGPT, Claude, Gemini, Copilot, Grok, DeepSeek and Perplexity is scanned identically, and so is text copied out of Word, Google Docs, a PDF or a web page.
Does checking my text upload it anywhere?
No. The scan runs entirely in your browser — the text in the box never leaves this page, there is no account, and there is no limit on how much you check.
Why doesn't the checker clean my text automatically?
Because seeing what is there and removing it are two different jobs, and this page is built to do the first one well. Once you have read the report, the cleaner removes everything found in one click and confirms the result is actually clean afterwards — not just that it counted something once.
Can a watermark checker like this prove a text was written by AI?
No, and treat any tool that claims it can with suspicion. The characters this page finds are typographic debris, not proof of authorship — human-written text picked up in Word or a PDF carries the same kinds of characters. A separate, different thing entirely is a statistical watermark hidden in which words a model chose, which no character checker, including this one, can see at all. The section below explains that distinction in full.
What is a text watermark, exactly?
It usually means one of two different things. The first is invisible characters — zero-width spaces, tag characters and the like — that a chat interface's rendering leaves in text you copy out; this page detects exactly those. The second is a statistical watermark, encoded in which words a model chose to write in the first place, which leaves every character looking completely ordinary. This checker can only ever see the first kind; the section below explains why nothing that reads plain text can see the second.
Is this an ai metadata cleaner?
Only in a narrow sense. If "AI metadata" means the invisible Unicode payload some tools can encode directly into plain text — the tag-character block at U+E0000 to U+E007F, built to carry hidden data with no visible form at all — yes, this finds it and the cleaner removes it. If it means file metadata, such as EXIF fields on an image or content-credential tags some tools now write into a PDF or a photo, this is the wrong tool entirely: it only ever reads plain text you paste in, never a file's own metadata fields.
Related tools
Same detection logic behind every page here — only the framing changes.
- Clean text — remove what this checker finds, in one click.
- Invisible characters, explained — the full list of what gets flagged, and why.
- AI text remover — what "remove AI from text" actually means, and what does and doesn't do it.
- ChatGPT watermark remover for text copied out of ChatGPT.
- Claude watermark remover for text copied out of Claude.
- Gemini watermark remover for text copied out of Gemini.
What this tool cannot remove
There are two different things people call an "AI watermark".
- Invisible and typographic characters. This tool removes named characters that carry no visible glyph at all, as defined by the Unicode standard — not a number we made up. It also turns different space variants, including the narrow no-break space, into an ordinary space. Curly quotes and dashes go through a separate, optional step — typographic replacements, switched on by default; turn that option off and they stay exactly as typed. It also clears the three Unicode private-use ranges, including the tag range U+E0000–U+E007F used to hide data in plain text — those ranges hold no standard characters of their own, so we describe them rather than dress them up as a character count.
- A statistical watermark in the word choice itself. Google's SynthID-Text, live in Gemini since May 2024, is documented to change no characters at all, so no character cleaner — ours included — has anything to catch. Anthropic has marked Claude's output the same way since 2 August 2026, to comply with the EU AI Act, but its own explanation (support.claude.com) confirms only that the mark exists, not how it works — so we can't promise it is built the same way SynthID is. What we can say: a character cleaner like this one only ever touches Unicode characters, and no watermark that relied on those would survive an ordinary copy-paste, so there is nothing here for it to remove either way. As of 10 September 2026, OpenAI's own content-provenance page lists a text watermark for neither ChatGPT nor its API.
If a tool promises to strip "the AI watermark" without making this distinction, it is describing the first thing while letting you believe it does the second.
By