AI watermark detector
0 words · 0 characters
Scan to list hidden characters. Your text stays in this tab.
Rewrite is not available yet.
What this detector finds
It finds characters you cannot see: zero-width spaces, byte-order marks, unusual spaces, and letters that only look Latin. It does not find statistical watermarks such as textGrain, Claude’s watermark, or SynthID-Text, because those live in word choice and need the provider’s secret key.
How to check a passage
- Paste.
- Press Scan.
- Read the list, then press Strip if you want those characters gone.
What a character scan cannot tell you
| This site — free | This site — paid rewrite | Typical zero-width cleaners | |
|---|---|---|---|
| Zero-width characters | Find and delete, in the browser | Same as free | Find and delete |
| Em dash | Optional punctuation cleanup — not called a watermark | Same as free | Often labeled as a watermark |
| textGrain / Claude statistical watermark | Cannot detect | Rewrite changes wording; no result promise | Often imply cleanup is enough |
| Privacy | Text stays in the browser | Rewrite is sent to the model for that request; we do not store it | Often unclear |
When a rewrite is the only thing that changes the wording
A paid rewrite changes the words themselves. It is a separate step from stripping hidden characters. It does not promise how any detector will score the result. See Pricing.
What an AI watermark detector actually checks
People search for an AI watermark detector when pasted text looks normal but fails a paste check, breaks a CMS field, or shows odd spacing in a spreadsheet. The usual cause is invisible Unicode: zero-width spaces (U+200B), zero-width non-joiners, byte-order marks left over from a file export, narrow no-break spaces, and look-alike letters that copy as Latin but are not. This AI watermark detector lists those code points in your browser so you can see exactly what was hiding in the string.
That is different from statistical watermarks. Systems such as textGrain, Claude’s published watermark research, and Google’s SynthID-Text change which words a model prefers. Detecting them needs the provider’s detector or key. This page does not claim to run those detectors, and it will not label ordinary prose as “AI written” based on style alone.
If you only need to clean a draft before publishing, scan first. If the scan finds nothing and a third-party AI checker still flags the text, the flag is about wording patterns, not hidden characters. In that case a rewrite changes the words; stripping invisible marks will not change a statistical score.
Common places hidden characters show up
- Copy-paste from ChatGPT, Claude, Gemini, or other chat UIs into Google Docs, Notion, or a CMS.
- Exports from Word, Pages, or PDF text layers that insert a byte-order mark at the start of a file.
- Marketing snippets pasted through Slack, email, or a design tool that normalizes spaces differently.
- CSV or JSON fields where a zero-width character breaks a unique key or a URL.
- Homoglyph swaps used in phishing demos or adversarial examples that look like ASCII.
In each case the visible letters look fine. An AI watermark detector that works at the character level is useful because it shows the code point, the count, and where each mark sits, instead of guessing from writing style.
How to use this AI watermark detector day to day
Paste the full passage into the box above. Keep formatting if you care about line breaks; the scan reads the raw string you paste. Press Scan. Review every listed character: zero-width marks, unusual spaces, and homoglyphs. If you want them gone, press Strip and copy the cleaned text back into your document.
For recurring workflows — editorial review, support macros, product copy — run the free scan before you publish. Keep a short checklist: scan, strip if needed, then spell-check. Do not treat a clean scan as proof that text was written by a human. It only proves those invisible characters are gone.
Teachers, editors, and security reviewers use the same flow when they need evidence. Export or screenshot the result list if you must show what was found. The tool stays local for Scan and Strip, so the draft does not leave the browser for those steps.
Use cases for an AI watermark detector
Publishing and SEO teams
CMS validators sometimes reject titles or meta descriptions that contain zero-width characters. Scanning the draft before upload avoids mysterious “invalid character” errors and keeps canonical titles readable in search results.
Developers and data pipelines
APIs that compare strings, hash cache keys, or build URLs fail when an invisible character sneaks in. Paste the suspect field here, strip the marks, and re-run the job with a clean value.
Education and research
When a paper or assignment is accused of carrying an “AI watermark,” separate character-level marks from statistical claims. This detector documents the former. For the latter, use the provider’s own tools when they exist, and state clearly what you cannot verify.
Trust and safety
Homoglyphs can make a domain or brand look legitimate. A character scan highlights non-ASCII look-alikes so reviewers can decide whether the string is safe to trust.
Limits you should expect
This AI watermark detector is honest about scope. It will not decrypt a provider’s secret watermark. It will not score “how AI” a paragraph sounds. It will not remove meaning by rewriting unless you choose a paid rewrite step, and even then there is no promise about how any other detector will react.
Em dashes and curly quotes are punctuation, not secret watermarks. Optional cleanup can normalize some of them, but this site does not market ordinary punctuation as proof of AI origin. If another cleaner labels an em dash as a watermark, treat that claim carefully.
AI watermark detector checklist before you publish
- Paste the final title, meta description, and body into the detector above.
- Scan and strip any zero-width or look-alike characters the list reports.
- Re-scan once to confirm the list is empty.
- Run your normal spell-check and link check.
- If a separate AI-writing score still looks high, treat that as a wording question — not a hidden-character problem — and decide whether a rewrite is worth it.
Following the same order every time keeps the AI watermark detector step boring and reliable. Most publishing failures from invisible Unicode disappear after one careful strip. Statistical watermarks and style-based detectors remain outside this tool’s job, and this page will keep saying so.
FAQ
What is an AI watermark detector?
An AI watermark detector looks for marks that survived a copy-paste. This site focuses on hidden Unicode characters: zero-width spaces, byte-order marks, unusual spaces, and look-alike letters. It lists each code point in your browser so you can strip them before publishing.
Can you detect ChatGPT’s textGrain?
No. textGrain is a pattern in word choice. OpenAI’s detector is not public, and this site does not claim to run it. Clearing invisible characters will not remove a statistical watermark.
Does the text leave my browser?
Not for Scan or Strip. A rewrite, when that feature is on, is sent to the model provider to produce new wording. We do not store the text.
Is this affiliated with OpenAI, Anthropic, or Google?
No. Unofficial. Brand names appear only so you can match the right help page for ChatGPT, Claude, or SynthID questions.
Will a clean scan prove the text was written by a human?
No. A clean scan only means those invisible characters are gone. Style-based AI detectors can still flag human writing or miss machine writing. Treat character cleanup and authorship claims as separate steps.
What should I do after the AI watermark detector finds marks?
Press Strip, copy the cleaned text, and paste it back into your document or CMS. Re-scan once to confirm the list is empty. If a third-party checker still flags the passage, that is about wording, not hidden characters.
Does stripping change the meaning of my draft?
Strip removes invisible or look-alike characters. It does not paraphrase. Visible words stay the same unless you run a separate rewrite.
Can I check large documents?
Paste the section you care about. Very long pastes can be slow in the browser; splitting by chapter or meta field is usually enough for publishing checks.