App Features
AI ProofreaderText HumanizerParaphrasing ToolAI SummarizerAI TranslatorFree ToolsFAQsUser GuideBlogPricingContactSign up - It's freeSign inBy signing up, you agree to our Terms and Conditions and Privacy Policy.
Anthropic began embedding an invisible watermark in Claude's text output on August 11, 2026, on every Claude model released from August 2 onward. It is a statistical mark woven into the model's token choices, not a hidden character, so it survives copy and paste and the tools that strip invisible characters do nothing to it. Rewriting the text is what removes it. Paste up to 500 words and this tool re-expresses every sentence in fresh wording, while your facts, numbers, hedges and citations stay exactly as you wrote them.
Paste or type your text, then run the tool. Results appear below in seconds.
0 / 500 words
Trusted by researchers publishing in
Between 40 and 500 words. The lower bound matters: a watermark measured across a token distribution needs a reasonable stretch of text, and so does a rewrite that has to keep your meaning intact.
A different model regenerates the passage word by word, varying sentence openings and clause order. Because the tokens are new, a signature keyed to Claude's sampling cannot carry through.
Compare the rewrite against your original. Facts, numbers, hedges, terminology and citations are held fixed, but you are the one who confirms the meaning survived intact.
Paste your own paragraph in for a copy-edit and it comes back watermarked. The mark cannot tell an edit from a draft.
Removing hidden characters cannot touch a sampling-bias watermark. Regenerating the tokens is the only mechanism that works.
May stays may and suggests stays suggests. Numbers, dates, names and hedges come out exactly as they went in.
Quotations, in-text citations and reference entries are copied character for character, never reworded or renumbered.
On August 11, 2026, Anthropic announced that it would embed an invisible watermark in the text its models generate, set out in its own documentation at How Claude marks AI-generated content. Models released on or after August 2, 2026 support the marking at launch, and Anthropic says it is working to add it to older models during a transition period. The policy reaches the Claude API, claude.ai, Claude Code, Claude Cowork and Claude served through AWS, Google Cloud and Microsoft Foundry. It applies worldwide, with no opt-out for users on any plan. The driver is the EU AI Act's Article 50(2) Code of Practice on transparency for AI-generated content, which obliges providers to mark machine-generated output; Anthropic chose to apply the mark globally rather than to European traffic alone.
Not as a hidden character, and not as file metadata. As it writes, the model repeatedly chooses between words that are statistically near-equivalent at that point in the sentence. The watermark biases those choices according to a secret key, so across a long enough passage the resulting token distribution carries a measurable signature. Nothing is inserted into the text, which is why the output reads normally, why Anthropic can say the mark does not affect meaning or quality, and why it survives copy and paste, a change of file format and conversion to plain text.
The distinction is mechanical. Anything that leaves your actual words in place leaves the signature in place: copying, reformatting, changing the file type, stripping invisible characters. Anything that reconstructs the token sequence destroys it: a substantial rewrite, heavy paraphrasing, translation into another language. Anthropic has acknowledged as much, describing the system as a first step that editing can defeat. Very short passages are unreliable in a different way, since a statistical measurement needs enough text to accumulate a signal before it can say anything.
This is where most of the free tools currently marketed as AI watermark removers go wrong. They scan for zero-width Unicode characters and invisible spaces and delete them, which was a reasonable response to a different problem and does nothing at all here. Running that kind of cleaner over Claude output returns text with exactly the signature it started with.
The complaint that dominated the response to the announcement did not come from people hiding AI drafting. It came from writers who use the model as an editor. Feed a finished paragraph of your own into Claude for a copy-edit and what comes back is watermarked, because the mark records that the model produced those tokens, not that it produced the ideas. Lawyers, academics and researchers made the same objection within hours: their human writing, lightly edited, would now carry a signal that an employer, publisher, journal or institution could read as proof of AI authorship. Re-expressing the passage in fresh wording is the direct answer to that attribution problem.
Only through Anthropic. Detection is planned via the company's own API, and general-purpose AI detectors cannot read this mark. Even then the result is probabilistic: a positive says the text may have been processed by Claude, without establishing authorship, the proportion that was machine-generated, or whether the model drafted or merely edited. A negative says almost nothing, since unmarked text may still be AI-generated. Treating either result as proof is a misreading of what the measurement can support.
The tool on this page rebuilds a passage. The ProofreaderPro AI humanizer goes further, and for a watermarked document it is the more complete answer. A humanization pass does not edit your text in place or patch the parts that look machine-written. It reads the document and writes it again from the ground up, moving AI-style prose into human style writing: sentence lengths stop being uniform, clause patterns stop repeating, the register loosens where a person would loosen it, and the vocabulary that gives machine drafting away is replaced with wording a human writer would actually reach for.
Scrubbing the watermark is a by-product of that, and a total one. Anthropic's mark is carried by the specific sequence of tokens Claude chose. When the humanizer rewrites a document, not one of those tokens survives into the output: every sentence is generated afresh by a different engine, with different sentence boundaries and a different distribution of word choices. There is no partial case to worry about, no residue in the paragraphs the humanizer touched lightly, because it does not touch anything lightly. The statistical signature has nothing left to be measured against.
The practical difference from this page's tool is scale and depth. This tool takes 500 words at a time and re-expresses them faithfully, which is right for one paragraph you need cleared. To humanize AI text across a full manuscript, thesis chapter, report or article, and to have the result read as though a person wrote it rather than merely read as different, the AI humanizer handles the whole document in one pass while protecting your citations, figures and technical terms exactly as this tool does. Anyone dealing with a watermark across a complete draft should start there.
ProofreaderPro.ai is an AI academic editing suite that proofreads, paraphrases, humanizes and cites, built for researchers, students and professionals writing for publication. This tool is the narrow case: a rewrite pass whose only job is to re-express a passage faithfully. For a full manuscript, the ProofreaderPro editor works across a complete document with tracked changes you approve line by line, and it does not watermark your text. If your goal is to move machine-sounding prose into a natural human register rather than simply to re-express it, the AI text humanizer is the tool built for that, and a full humanization pass removes a token-level watermark as a side effect of how it works. The rest of the free tools cover the checks around it.
A Claude-edited passage and its rewrite. Read them side by side: the claims, the numbers and the hedges are identical, and not one sentence survives in its original wording.
The findings of this study suggest that remote work arrangements have a measurable effect on employee productivity, though the magnitude of that effect varies considerably across sectors. Participants in knowledge-intensive roles reported higher output when working from home, while those in collaborative or client-facing positions reported the opposite.
This study's results point to a measurable productivity effect from remote work arrangements, although how large that effect is differs considerably from one sector to the next. Among participants in knowledge-intensive roles, output was higher at home; participants in collaborative or client-facing positions reported the reverse.
Nothing here was deleted or added. The paragraph makes the same two claims, at the same strength, about the same two groups, in 48 words instead of 50. That is the whole trick: the meaning is portable, the token sequence is not.
As Claude writes, it constantly hits ties between words that would all fit. A secret key breaks each tie. No character is added to the text; the pattern of picks is the watermark. That is why a detector holding the key can read it, and why only new words can erase it.
"The results suggest a measurable effect across sectors." At each tie, the key nudged the pick:
"The results point to a clear shift between industries." Every token regenerated by a different engine:
Chance agreement never falls to zero, which is why detection is statistical and needs a long passage. It is also why Anthropic's own explainer concedes the limit: "a complete rewrite where every word is replaced" removes the mark.
All three regenerate the token sequence. That is the entire mechanism, and it is the one this tool uses.
Claude's watermark is not a novel invention. It is the deployment of a research programme that has run in public since 2022, and its known limits were published along the way.
In a lecture at UT Austin, Scott Aaronson describes his OpenAI project: bias the model's choices among near-equivalent words with a secret cryptographic function, so the text itself proves where it came from.
Read the lectureKirchenbauer and colleagues at Maryland publish A Watermark for Large Language Models, the peer-reviewed recipe: promote a keyed subset of tokens during generation, detect with a statistical test, no model access required. ICML 2023.
arXiv:2301.10226Krishna and colleagues build DIPPER, a paraphraser that collapses AI-text detectors, watermarking included; DetectGPT falls from 70.3% to 4.6% detection. The weakness Anthropic would later concede was documented three years in advance. NeurIPS 2023.
arXiv:2303.13408Google DeepMind publishes SynthID-Text in Nature, the first sampling watermark deployed at scale in a live product. This is the technique Anthropic names as the basis of Claude's mark.
Nature (2024)Every Claude model released from August 2, 2026 marks its text output, worldwide, on every plan, with no opt-out, implementing Article 50(2) of the EU AI Act. Three days later the company publishes its technical explainer.
Anthropic's explainerHow the watermark works, what defeats it and why is not our opinion. It is documented in Anthropic's own publications, in peer-reviewed research, and in the regulation that prompted the rollout. Read the sources directly.
The company's own technical explainer: the mark alters the source of randomness in word selection, detection is probabilistic, it cannot separate drafting from editing, and a complete rewrite removes it.
anthropic.comThe policy document: which models mark, which platforms are covered (API, claude.ai, Claude Code, cloud providers), worldwide scope, and no opt-out on any plan.
support.claude.comSynthID-Text, Google DeepMind's production watermark and the published technique Anthropic's implementation is based on. The definitive description of how a sampling watermark works at scale.
doi.orgThe foundational scheme: partition the vocabulary with a key, promote the keyed tokens while generating, and detect the bias with a statistical test on the text alone.
arxiv.orgThe robustness result behind this page: regenerating the wording collapses detection across methods, watermarking included, because the signal lives in the token sequence and nowhere else.
arxiv.orgThe legal driver: providers of generative systems must ensure outputs are 'marked in a machine-readable format and detectable as artificially generated'. Anthropic chose to comply globally rather than for EU traffic alone.
artificialintelligenceact.euAll external links open in a new tab and were verified live before publication. Citation details follow each venue's own record.
The ProofreaderPro editor proofreads, tightens and formats your complete paper with tracked changes you approve line by line, and it does not watermark your text. Trusted by researchers worldwide, free to try.
Edit my full paperProofread, humanize, paraphrase and translate your academic writing in one place. Free to start, no credit card required.
Get Started Free