AI Watermark Remover

Why Claude text now carries a watermark

On 2 August 2026 Anthropic started embedding a machine-readable watermark into text produced by Claude models released from that date, including Claude Sonnet 4.6 and Claude Haiku 4.5. The technique is a version of Google DeepMind's SynthID-Text: it biases word selection in a pattern a detector can recognise, without changing how the text reads. Anthropic applied it worldwide rather than only in the EU, in response to the transparency duties in Article 50 of the EU AI Act.

How a rewrite removes it

A statistical watermark lives in the specific sequence of words the model chose. Change those words and the signal degrades. This is Anthropic's own account of what does and does not survive:

• Light editing usually survives. The mark "may persist through some editing", because swapping a word here and there leaves most of the original token sequence intact.

• Heavy paraphrasing does not. Anthropic states the mark "won't be detected if heavily edited, paraphrased, translated, or mixed into other writing".

A full model-to-model rewrite falls firmly in the second category. Every sentence is rebuilt by a model with no part in Anthropic's watermarking scheme, so the pattern a detector looks for is not there to find. Mechanically, this tool:

• Extracts the text from your .docx, .pdf, .txt or .md, or takes it straight from the paste box
• Counts the words and estimates the tokens before anything is sent, so an over-limit file is refused up front rather than partway through
• Splits the text into chunks on paragraph boundaries, so no paragraph is cut in half
• Rewrites several chunks at once through DeepSeek V4 Flash
• Reassembles them in their original order and hands the result back to copy or download

What a watermark can and cannot tell anyone

It is worth being precise about what a detection result would actually mean, because the reporting around this has been loose. Anthropic's own wording is that a detected mark "provides a signal that content was processed by Claude, but is not fully conclusive". Processed, not authored. If you wrote a report yourself and asked Claude to tighten the grammar, the output can carry the same mark as a document Claude wrote from a one-line prompt. A detector cannot distinguish those two cases, and neither can anyone reading its output.

That cuts both ways. It means a positive result is weaker evidence than it sounds, and it also means people who used AI in entirely ordinary ways, for proofreading, for translation, for tidying a draft, can be flagged by a system that was never designed to measure authorship. Anthropic has also said the mark is not present in every response: very short outputs do not carry enough tokens to hold a reliable signal.

Who uses a tool like this

The common cases are mundane:

• Agencies and publishers who run drafts through several models and need one consistent voice in the final copy rather than a patchwork of styles.
• Companies with a no-verbatim policy that forbids shipping third-party model output straight into their own documentation.
• Translators and editors handed text that passed through an assistant somewhere in a long chain, who want a clean starting point.
• Writers who used AI for proofreading only and do not want a grammar pass to read as though the whole piece were machine-written.

There is one case where you should stop and think. If you are submitting work to a university, a journal, a client or an employer that requires you to disclose AI assistance, rewriting the text changes what a detector sees but does not change what you are obliged to declare. The obligation attaches to how the work was produced, not to what a tool can measure afterwards. Rewriting to avoid a disclosure you owe is a decision about honesty, not a technical problem, and this page will not pretend otherwise.

What the rewrite preserves, and what to check

The model is instructed to rebuild every sentence while leaving the substance alone. In practice it holds up well on narrative and business prose.

Kept

• Meaning, facts, figures, names and direct quotations
• The order of your argument and your paragraph breaks
• The original language, register and approximate length
• Words inside .docx tables, which are read and rewritten too

Not kept

• Headings, tables, images and styling from the original .docx or .pdf, because text is extracted as text and returned as text
• Footnotes, endnotes and tracked changes
• Anything in a scanned or image-only PDF, which has no selectable text to read

Read the output before you use it anywhere that matters. Check numbers and units, direct quotations, and any technical or legal term of art in particular: a paraphrase that improves readability can quietly soften a precise claim, and that risk rises the more specialised your material is.

Which tool do you actually need

Four tools here overlap in ways that are easy to confuse, so this is what each one is for:

Limits, and why they are counted in tokens

Allowances here are measured in tokens rather than words, because tokens are the unit the underlying model actually consumes. A word is only a rough proxy: dense technical writing, code, tables of figures and languages other than English all break into more tokens per word than plain narrative, so two documents of identical word count can differ substantially in real cost. Counting in tokens means the limit is the same for everyone regardless of what kind of text they bring.

The page shows an approximate word figure next to every token budget, and the counter under the paste box converts as you type, so you can see where you stand before you submit. Uploaded documents are scanned and counted before any rewriting begins: if a file is over your allowance you are told immediately rather than after a long job fails partway through.

Accepted files: .docx, .pdf, .txt and .md, up to 12 MB. A password-protected or image-only PDF cannot be read, because there is no selectable text to extract.

Frequently asked questions

Related: AI Humanizer rewrites text to read more naturally, and AI Detector estimates whether a passage looks AI-written.