AI text watermark remover

Which AI text watermarks exist and which one your text carries

Not all AI text watermarks are the same. Before you remove anything, you need to know whether your text carries a statistical watermark, hidden characters, or only stylistic signs that a separate AI detector may dislike.

Claude now uses a statistical text watermark. On August 11, 2026, Anthropic announced that Claude output carries a key-based signal created by biasing word choices. There are no hidden characters to delete. The system applies worldwide to Claude models launched on or after August 2, 2026, including the app, Claude Code, Cowork and API, with older models being added progressively.

Gemini uses SynthID-Text, a statistical watermark deployed since 2024. It is described in a Nature paper by Dathathri et al., published in October 2024, and Google has a SynthID Detector portal.

ChatGPT is different. OpenAI has a text watermark internally, but it is not deployed as of August 2026. What many people call a ChatGPT watermark is usually invisible characters, non-breaking spaces, zero-width marks, odd hyphens and formatting debris.

If you need the wider map, see which AI vendors watermark text. It explains the difference between vendor watermarks and style-based AI detectors such as GPTZero, Turnitin or Originality. Unmarker measures watermark signals, not whether a text reads like it was written by a model.

Remove the Claude watermark

The Claude watermark is statistical. It survives copy-paste and may remain after light editing because it is not stored in formatting or metadata. A normal cleaner cannot remove it by stripping characters, because the signal is in word choice patterns.

Unmarker rewrites Claude text with a model that carries no watermark. It then checks every sentence for meaning, including entities, numbers and quotes, and shows a before and after watermark score with a diff. The aim is a controlled rewrite, not a vague humanising pass.

Very short passages are harder to assess because statistical signals need enough text. Heavy paraphrase or translation can weaken a watermark, but they can also change meaning. Unmarker is designed to keep Markdown and HTML formatting while showing you exactly what changed.

A detected Claude mark means Claude touched the text. It does not prove Claude wrote every sentence. Human text that was proofread, translated or edited with Claude can carry enough signal to be flagged, and false positives are possible.

Unmarker is an independent product, not affiliated with or endorsed by Anthropic.

Remove SynthID-Text (Gemini)

SynthID-Text is also statistical. It does not rely on invisible characters, so the removal job is closer to Claude than to a formatting cleanup. You need a rewrite that changes the watermark pattern while preserving facts and structure.

Unmarker will treat SynthID-Text as a separate job because Gemini watermarking has its own detection context. The product is planned as one engine with focused tools, rather than one generic button that hides which signal it is handling.

When a SynthID check is available in the workflow, the useful output is not only the rewritten text. You should also see the detected score before editing, the score after rewriting, and a diff that lets you reject changes that alter meaning.

Clean invisible characters and "ChatGPT watermarks"

As of August 2026, ChatGPT text watermarking is not deployed. In practice, requests for ChatGPT watermark removal usually mean cleaning copied text from editors, browsers, PDFs or messaging tools.

That cleanup can remove zero-width characters, non-breaking spaces, unusual hyphens and broken spacing. It can also normalise pasted Markdown or HTML where formatting debris creates odd behaviour in Word, Google Docs, CMS editors or email tools.

This is a different problem from Claude or SynthID-Text. Cleaning invisible characters does not remove a statistical watermark. Likewise, rewriting a statistical watermark is not the same as fixing a document full of hidden Unicode characters.

A clear tool should tell you what it found. If the issue is invisible characters, the answer is a character and formatting cleanup. If the issue is a vendor watermark, the answer is measured rewriting with meaning checks.

What a text watermark remover cannot do

A text watermark remover cannot promise outcomes from style-based AI detectors. Tools such as GPTZero, Turnitin and Originality answer a different question: does this text read like a model wrote it. Their results can include false positives, especially on edited, formulaic or non-native writing.

A watermark detector asks a narrower question: does this text carry a vendor key-based signal. That distinction matters. Removing or reducing a watermark does not mean a style detector will score the text differently.

A remover also cannot make unlawful or dishonest use acceptable. If a school, publisher, employer or platform requires disclosure of AI use, you should follow that rule. Unmarker is not advice on hiding AI use where disclosure is required.

It also cannot preserve meaning without review. Automated checks can compare entities, numbers and quotes, but you remain responsible for reading the final text. Any rewrite may introduce wording changes that matter in legal, academic, medical or technical contexts.

One engine, one tool per job

Unmarker is planned as one engine with specific tools for different text watermark problems. The same interface can measure, rewrite, clean and compare, but each tool should be honest about what it is handling.

• Claude watermark remover: rewrites Claude text, checks sentence meaning and shows a before and after watermark score.
• Claude watermark detector: measures whether a text appears to carry the Claude statistical signal.
• ChatGPT watermark remover: cleans invisible characters and formatting debris often mislabelled as a ChatGPT watermark.
• SynthID text detector and remover: handles Gemini SynthID-Text as a separate statistical watermark workflow.
• AI text cleaner: normalises pasted AI text, spacing, characters and formatting without pretending to solve every watermark case.
• Claude humanizer: rewrites Claude output for clearer human review while preserving facts and structure.
• AI watermark detector: checks for supported vendor watermark signals rather than style-based AI detector scores.

This separation avoids confusing three different jobs: detecting a vendor watermark, reducing that watermark through controlled rewriting, and cleaning hidden characters. It also makes the limits visible before you edit a text.

Planned pricing

Unmarker is pre-launch and the waitlist is open for early access. Planned pricing is simple, but may change before launch.

• Free: 3 texts per day, up to 1,500 words.
• Pro: $9 per month.
• Founders: $5 per month locked for the first 500 founders.
• API: $0.50 per 10k words.

A free detector is planned. The paid product is for rewriting, checking, diffs, formatting preservation and repeated use.

Unmarker is built by STRATINET SARL in Cannes, France, founded by Alexandre Fuchs. For access or questions, you can join the waitlist or contact contact@alexandrefuchs.fr.

Free tools that run in your browser

Seven small tools for the visible side of the problem, free, without an account, and without sending your text anywhere. They do not touch a statistical watermark; they fix what you can see or count.

• Invisible character remover: zero-width characters, odd spaces, soft hyphens, bidi marks, lookalike letters, by category.
• Em dash remover: comma, period, colon, semicolon, hyphen or parentheses, with a diff.
• AI text cleaner: hidden characters, Markdown, quotes, dashes, bullets and spacing in one pass.
• AI words checker: counts AI writing tells by category, highlighted; not a detector.
• Unicode inspector: every character with its code point, name, block and flags.
• Smart quotes to straight quotes, and back.
• Markdown remover: Markdown to plain text.

Who builds these tools

Unmarker is built by STRATINET, a French company based near Cannes, founded by Alexandre Fuchs, with no outside investor and no partnership with any AI vendor. Who we are, why we started after the August 2026 announcement and how we work (measured scores, no detector promises) are on the About Unmarker page; the legal identity is on the legal notice.

Questions people ask