AI models are starting to watermark their text

Remove the AI watermark.
Rewrite every word.

Claude, Gemini and ChatGPT mark their text by biasing which word the model picks. GenPolish strips the invisible characters and rewrites the word choices the mark is made of, then shows you how much of your draft changed.

Try it free, no account See how it works Which models mark their text

5,000 words / month on FreeNo cardDrafts never used for training

Optional analytics

Help us measure the journey from a visit to rewriting and payment — never your drafts, results or email.

Privacy and analytics preferences

Your draft3 successful rewrites without an account

Demonstration of the free rule-based tool, not an AI result from your draft.

Human score96%

Original38% human

In today's fast-paced world, it is important to note that leveraging AI for content creation can unlock a myriad of opportunities. Whether you are a seasoned marketer or just getting started, it is crucial to understand that the landscape is constantly evolving. By harnessing the power of these tools, teams can significantly enhance their productivity and drive meaningful results.

Polished96% human

AI writing tools change what a small content team can ship. Two concrete wins: first drafts arrive in minutes instead of hours, and editing replaces staring at a blank page. The catch is that unedited output reads like output — hedged, padded and interchangeable.

AI watermark database

What the watermark actually is

2 of these 5 assistants mark the text they generate, and ChatGPT does so in part. None of the marks is made of characters, so no scanner can read it — it is a bias in which word the model picks next, applied thousands of times across a passage.

That decides what can be done about it. Cleaning characters does not touch it. Replacing the words does, because the words are what it is made of, and that is what a rewrite replaces. Each vendor says so in its own documentation.

ClaudeSynthID-Text, keyed token selectionMarked
GeminiSynthID-Text, tournament samplingMarked
ChatGPTtextGrain, keyed word choice; EU and API opt-inMarked in part
CopilotC2PA on images, nothing for textNot marked
GrokVisible mark on media onlyNot marked

Read the evidence, provider by provider

How it works

Three steps, no black box

Several of the same checks also run as standalone browser tools, without an account and without sending your text anywhere.

Paste

Drop in text from Claude, ChatGPT or Gemini. The invisible characters come out before anything else happens.

Rewrite

Pick a voice. The rewrite replaces word choices, which is what a statistical watermark is made of.

Review

Every change carries its category and its reason. Reject the ones you disagree with and you still keep the clean text.

What we catch

The tells that make text read like output

Every polish sorts its edits into 9 published categories, each with a rationale you can argue with. There is no proprietary “humanising” score you cannot inspect.

Stock openersHedgesFiller phrasesVague claimsClichésPaddingRepeated shapesListicle reflexOver-signposting

What each one costs to fix, with examples from real drafts: the tells that make AI text obvious.

Stock opener

In today's fast-paced world,(removed)Adds no information about the subject. Cutting it costs nothing.

Filler phrase

it is important to note that(removed)Hedging preamble. The sentence is stronger without it.

Vague claim

unlock a myriad of opportunitieschange what a small content team can shipReplaces an abstract benefit with a specific one.

Hedge

it is crucial to understand that(removed)Instructional hedge aimed at the reader rather than the subject.

9change categories, each explained

5of 8 invisible characters cleaned from every rewrite

90days before a saved rewrite expires

0drafts used for model training

Those 8 characters are not the watermark — the watermark is a bias in word choice, and the watermark database records which models apply one and what the evidence is. 5 of the 8 come out of every rewrite; the 3 joiners stay, because Arabic, Indic scripts and emoji need them, and the Unicode cleaner removes those on request.