Make your writing sound like you wrote it.
Paste your draft and WatermarkRemoverPro opens a workspace with it rewritten, a comparison of every change, and exactly what a detector would still measure in the result. All of it runs in this browser tab.
Paste, drop a file, or upload .txt, .md, .rtf or .csv. Everything is read and processed on your device.
Your text is processed in this browser tab and never uploaded. Here is how that works.
- Runs on your device: open the network tab and watch
- English, Spanish, French, German, Portuguese
- 20,000 free rewriting tokens every 7 days
New, 5 October: OpenAI will watermark ChatGPT and Codex text in the EU with textGrain. Its own tests show 25% of words replaced drops detection to about 17%, and only approved researchers can run the detector. Read more or see textGrain explained
Just a few of the brands that trust WatermarkRemoverPro
Ledgeur
Bill100
Brandmity
Chartely
Contextely
Contract10
Denial7
Duebay
Feedlark
Grannio
JobPlumb
Ledgerary
Lugbird
Patent77
Patientary
PermitBird
PortRobin
RenewBird
Rigbird
RotaBay
Scoutern
Sourceory
TableHelm
Underhaus
Vouchity
Hi Crafty
GradeHack
WillThisHappen
Frifti
Benefily
Classify7
Conformery
DeckHelm
Event70
Job13
Ledgerage
Ledgerler
Mealary
Meterary
ModelCharter
RoofHelm
Screen100
Slopeify
Spend7
WageCoachThree steps, and none of them involve a server.
The engine downloads to your browser once. Your document is read, measured and rewritten there, so there is nothing for us to store, log or hand to anyone.
Put your text in
Paste it, drop the file on the box, or use the upload button. Plain text, Markdown, RTF and CSV are read straight off your disk.
One button
Language, engine and how much to change all have sensible defaults. They are there under Advanced settings if you want them, and out of the way if you do not.
A workspace, not a wall of text
The button opens your workspace: the rewritten draft, how much AI evidence is left, a comparison you can act on paragraph by paragraph, and the full analysis when you want it. Everything you run is kept there, on this device.
A real measurement, not a percentage out of thin air.
Every figure carries the count it was computed from and a band around it. Where something could not be measured, the result says so instead of printing a zero and letting you assume it meant something.
The chance line
Where the number sits when nothing is there. A result on this line is a null result, not a low score.
The hatched interval
How far the measurement could move on a slightly different sample. Hatched, so it can never be mistaken for the value itself.
The marker
What was actually measured. Amber only ever means a mark was found under a key that was tested.
A mark was detected under one key.
WatermarkRemoverPro open reference scheme
not a vendor key
Green-list rate
68.0%
- Expected by chance
- 50.0%
- z
- 4.37
- p (one-sided)
- 0.000006
interval 60.1% to 75.0%
Measured over 147 distinct word pairs. Repeated pairs are counted once. The thin vertical line is the rate this statistic takes when no mark is present.
The chair thanked the working group for drafting the supporting brief, which those present found clear and remarkably short.
watermark z 2.36 · p 0.0092 · style distance not computed (passage too short to measure)
A question was registered about the budget line for the following stage, and the bursar committed to present exact totals to the following meeting rather than approximate them from recollection in the room.
watermark z 2.69 · p 0.0035 · style distance not computed (passage too short to measure)
What it does, and what it refuses to do.
It shows its working
Every figure is a computed statistic carrying the count it came from and a band around it. Where something could not be measured, the result says so.
It refuses to overclaim
A green-list mark is keyed, and no model vendor publishes its key. Results always name the keys that were tested, so a null result reads as “no mark found under these keys” and never quietly becomes “you are cleared”.
Your text never leaves
Rewriting is on-device on every tier and every surface, with no hosted mode at all. It cannot guarantee defeating a vendor’s undisclosed watermark, and it never says it can.
A style measurement, kept in its place.
How far a document sits from contemporary reference prose in the same language is a real, useful number. It is also the single most misread number in this field, so it is reported separately and never folded into the verdict.
Distance from reference reflects register, subject and translation. Technical writing, fiction, translated text and non-native prose all sit far from an encyclopaedic reference for entirely ordinary reasons, and none of that is evidence about how a document was produced.
0.64 SD
from contemporary reference prose
- Share of words used exactly once0.79 (+1.14 SD)
- Function words per 1,000 words493.83 (+1.13 SD)
- Commas per 1,000 words37.04 (-1.01 SD)
Do not take our word for it.
A detector you cannot test is a detector you have to trust. On the verify page you can mark a passage under our published reference key in your own browser, run the check, and watch the statistic move. Then run the same passage against a different key and watch it sit at chance. That is the difference between arithmetic and theatre, and it takes about fifteen seconds.
What that demonstration shows
- Marked text scored under the key it was marked with, far from chance.
- The very same text scored under an unrelated key, sitting at chance.
- Both runs happening in your browser, on text generated in front of you.
WatermarkRemoverPro is for writing you wrote yourself.
Screening someone else’s work, a student’s essay or a contractor’s copy, is a different job with different ethics and a different tool. Learnaway is built for checking someone else’s work for AI use. This one is not, and running a self-check tool over work someone handed you is the fastest way to misread its output.
Questions people actually ask
- Is a detected mark proof I used AI?
- No. A detected mark is not proof of authorship. Marks can appear in text that was quoted, translated, edited, or written with assistance, and a statistical test reports a probability rather than a fact about a person. WatermarkRemoverPro reports the strength of the signal and the passages carrying it, and states this limit on every result and every exported report.
- Can this tool guarantee my writing won't be flagged?
- No, and any tool that claims a 100% or guaranteed result is overselling a probabilistic process. WatermarkRemoverPro can reduce detectable AI-style evidence: both statistical watermark signal, where structurally possible, and human-perceptible AI tells like em dashes and stock phrasing. It cannot guarantee defeating a model vendor's undisclosed watermark, because nobody outside that vendor holds the key it was applied with, and no honest tool can promise otherwise. The rewrite runs entirely on your device, on every tier, and shows you the before/after evidence so you can judge the result yourself rather than take a guarantee on faith.
- Does WatermarkRemoverPro detect ChatGPT's or Claude's watermark specifically?
- Not unless this deployment holds that vendor's detection key, and no vendor currently publishes one; the result lists exactly which keys were tested, under "keysTested", so you can see that for yourself rather than take a claim on faith. WatermarkRemoverPro also reports a separate, key-free "AI-style likelihood" score: a heuristic that flags surface habits common in Claude, ChatGPT and other current model output (dash-clause connectors, stock phrasing, elevated vocabulary, uniform sentence length), deliberately biased to catch more real AI writing rather than to minimise false positives. It is not a statistical test and it is not a substitute for a real watermark result; see /guide/claude-ai-watermark for what a provenance mark actually is and why the key is the whole story. OpenAI began adding its own watermark (textGrain) to ChatGPT and Codex text in the EU in October 2026, and is limiting its detector to approved researchers and expert organisations, so the same limit applies: no public key, no third-party detection. /guide/chatgpt-watermark has the detail.
- What's the difference between the Standard and Pro rewrite engines?
- Standard is deterministic substitution against the core AI-tell library: instant, no download, and included on every plan. Pro is a real language model that runs in your browser, generating more candidate rewrites per passage and using the extended AI-tell library; it comes with a Pro subscription rather than being metered on the free plan. What the free plan meters is how much rewriting you do: 20,000 tokens every 7 days, where a token is one word as this product counts them. Checking is never metered on either plan. Neither engine ever sends your text anywhere.
- Does my document ever leave my device?
- Not on the free check, and never on the rewrite feature, on any tier or surface. Checking runs in your browser: the detection engine is downloaded to your device and the text is measured there; you can confirm it by opening your browser network tab and watching while the check runs. The API, the MCP server and the PDF evidence report necessarily run on our servers for CHECKING, because a programmatic caller has no browser, and those paths are documented separately and are opt-in. Rewriting is different and stricter: it runs entirely on-device or in-process everywhere, in the browser, the API package, and the MCP server alike, with no hosted mode at all, on any tier.
- Which languages are supported at launch?
- English, Spanish, French, German and Portuguese. Each has its own reference baseline measured from real contemporary prose in that language; a language without a measured baseline is reported as unsupported rather than analysed against a substitute, because comparing Portuguese text to a Spanish reference produces deviations that look like findings and are artefacts.
- An absent mark: does that prove I wrote it myself?
- No, and this matters as much as the first answer. Marks survive editing poorly, are not applied by every system, and cannot be detected at all without the key used to apply them. No model vendor publishes its detection key. WatermarkRemoverPro tests the keys it holds and names them, so "no mark detected" always means "under these keys" and never "this document is clean".
- What can I actually hand to someone who has accused me?
- The Pro evidence report: a dated PDF carrying the signal strength with its confidence band, the per-passage breakdown after false-discovery-rate correction, the full stated limits, the list of keys tested, and a SHA-256 hash of the exact document analysed. The hash is what ties the report to one specific file, so the report cannot be waved at a different draft.
An on-screen result answers your question. An appeal needs a document.
The Pro evidence report is a dated PDF carrying the signal strength and its band, the per-passage breakdown after correction, the keys tested, the full stated limits, and a SHA-256 hash tying it to the exact file you checked. The hash is what stops the report being waved at a different draft.