Since August 2026, Claude and Gemini weave an invisible pattern into their word choices. Run text through them for anything — grammar, translation, a polish pass — and the mark lands on your writing, even when the ideas and words are entirely your own.
The evidence desk for your own words
You wrote it. The machine left its mark anyway.
Ask Claude to fix your grammar and it stamps an invisible watermark on your own writing. Detectors like Turnitin and GPTZero flag careful, human work as AI — 18% of papers at one AI conference were auto-rejected, no appeal. Scan your text, see what's hidden in it, and clean it before someone else checks.
100% IN YOUR BROWSER. TEXT IS NEVER UPLOADED ON THE FREE SCAN. NO SIGNUP.
The quick brоwn fox jumps overthe lazy dog and nobody can see it happen.
18%
papers auto-rejected at one AI conference — no appeal
61%
of non-native English essays flagged as AI (Stanford)
350pp
of ChatGPT prompts produced in the 3M case
$90K
expert fee, torched by his own prompt log
Exhibit BHumanizer
UNLIMITEDRewrites the wording so the statistical watermark can't survive — using an open model that adds no mark of its own. Reads human, clears the flag.
0 characters
Humanized output
The mechanics, in plain English
Why clean writing gets flagged
Separately, AI detectors like Turnitin and GPTZero guess from style. They're wrong often: a Stanford study found 61% of essays by non-native English writers flagged as AI, and one conference auto-rejected 178 papers before review. Careful, formal writing is exactly what they misjudge.
Hidden characters strip out in one click. The statistical mark only clears when the wording is genuinely rewritten — so we rewrite it with an open model that adds no watermark of its own. Scan, clean, and check before the mark is ever held against you.
From the public record
Submit it dirty, and this is what happens
Every story below is a court record or published reporting. None of these people thought it would happen to them.
In 2020 an explosion at Watson Grinding in Houston killed three people and damaged some 200 homes. 3M paid an engineering expert roughly $90,000 for an independent assessment. Instead, he fed the case files into ChatGPT with prompts like "create an exceptional expert witness report… and show how 3M is 0% at fault." Opposing counsel spotted a ChatGPT citation overlay in the document, demanded every prompt, and got them: a 350-page log that paused the deposition for three hours. At trial the expert conceded 85–90% of his report was AI-written. 3M tried to disqualify its own witness; the plaintiffs called him themselves and read his prompts to the jury. Source →
The AI traces in the text are what opened the door. Once a filing gets flagged as machine-written, your prompts stop being private and start being evidence.
A Stanford professor who studies AI misinformation for a living signed an expert declaration under penalty of perjury. GPT-4o had helped draft it, and it cited articles that don't exist. The judge threw the declaration out: it "shatters his credibility with this Court." Kohls v. Ellison, D. Minn., Jan. 10, 2025. Source →
He wasn't cheating — he was polishing. One flagged passage was enough to destroy the whole submission. Verify and clean before it goes out, because "the AI did it" is not a defense.
In NYT v. OpenAI, a federal court ordered OpenAI to preserve every ChatGPT log — including chats users had deleted — and later ordered 20 million conversations handed over. "Deleted" meant "frozen for litigation." Separately, lawyers have begun subpoenaing opposing parties' AI prompts in discovery, arguing chatbots aren't your attorney so privilege doesn't apply. Some courts have pushed back; do you want to be the test case? Source →
Assume everything you've ever typed into a chatbot can be demanded. The one thing you control is whether the document you submit still carries the evidence.
AI detectors guess from writing style, and they guess wrong at your expense. A Stanford study found they flagged 61% of essays by non-native English writers as AI. One academic conference auto-rejected 178 papers on detector scores alone — no human review, no appeal. Careful, formal, second-language writing is exactly what gets punished.
You can do everything right and still eat the accusation. Scan your work before you submit it, so a hidden mark or a jittery detector doesn't get to define your record.
The policy
Sanitize your shit.
There is no advantage to submitting it dirty — not to a court, not to a professor, not to a client. Scanning and cleaning your own writing isn't cheating; it's taking back your rights over your own words. Your prompts, your drafts, your process — they're yours, and nobody hands them over for free.
ONE HONEST LINE: IF A COURT, JOURNAL, OR SCHOOL REQUIRES YOU TO DISCLOSE AI ASSISTANCE, DISCLOSE IT — SEVERAL JUDGES' STANDING ORDERS DEMAND EXACTLY THAT. WE REMOVE THE GIANT RED HIGHLIGHT ON THE COVER, NOT THE OBLIGATION.
Simple pricing
Scan free forever. Pay only to fix it.
Free
$0
FOREVER
- Unlimited hidden-watermark scans
- Strip invisible characters
- Hidden-text & homoglyph detection
- AI-detector risk score
- Humanize / rewrite
Most popular
Unlimited
$9.99/mo
OR $74.99/YR — SAVE 38%
- Everything in Free
- Humanize & rewrite to beat AI detectors
- Live AI-detector risk score
- Saved files on your account
- Unlimited use (within reason)
Receipts
Benchmarks
We run our humanized output through the detectors people actually use, and publish what passes.
| Detector / test | Raw AI text | After Humanize |
|---|---|---|
| GPTZero | Flagged | Coming soon |
| Originality.ai | Flagged | Coming soon |
| Turnitin | Flagged | Coming soon |
| Copyleaks | Flagged | Coming soon |
| Claude statistical watermark | Present | Coming soon |
BENCHMARK HARNESS SHIPS WITH THE PAID HUMANIZER. NUMBERS WILL BE LIVE, DATED, AND REPRODUCIBLE — NO CHERRY-PICKING.
Reference
Frequently asked questions
What is an AI watermark in text?
An AI watermark is a hidden signal embedded in generated text so a matching detector can later recognize it. It comes in two forms: a statistical pattern woven into word choices (invisible, survives copy-paste), and visible-layer tricks — zero-width Unicode characters, homoglyph letter swaps, and hidden white text.
How do I check text for a hidden AI watermark?
Paste your text into the free scanner at the top of this page. It runs entirely in your browser and flags zero-width and invisible characters, Cyrillic/Greek look-alike letters, unusual whitespace, and hidden or white text carried in from the source you copied. Nothing is uploaded on the free scan.
Can an AI watermark be removed?
Hidden characters can be stripped instantly — the free scanner does that with one click. The statistical watermark in the wording itself cannot be deleted in place; it only disappears when the text is substantially rewritten. That regeneration is what the paid Humanizer does.
Which AI models watermark their text?
As of August 2026, Claude and Gemini embed a statistical fingerprint in generated text, and several writing tools add invisible Unicode markers or metadata. Image, audio and video generators use C2PA "content credentials" metadata instead, which dies on re-save or screenshot.
What are zero-width characters?
Zero-width characters are Unicode code points — like U+200B (zero-width space) and U+2060 (word joiner) — that render as nothing at all. Because they're invisible, they can be arranged between normal letters to tag, fingerprint or even smuggle whole hidden messages inside ordinary-looking text.
Does a watermark prove text was written by AI?
No. A detected watermark only proves an AI model touched the text at some point — not that it wrote all of it. And the absence of a watermark proves nothing either, since marks don't survive heavy editing. Treat any detector's verdict as a weak hint, not evidence.