Claude's Text Watermark Explained: How It Works and How HumanizeMyPaper Removes It

In August 2026 Anthropic started doing something no major AI lab had shipped at scale before: every new Claude model now leaves an invisible, machine-readable watermark in the text it writes. If you use Claude to draft essays, blog posts, or emails, that mark travels with the text when you copy and paste it. This guide explains what the Claude watermark actually is, what it is not, who can detect it, what survives editing, and how the HumanizeMyPaper AI humanizer removes it. Where a claim comes from Anthropic or from the underlying research, we say so.

What changed in August 2026

The trigger was Article 50 of the EU AI Act, whose transparency obligations took effect on 2 August 2026. Providers of generative AI must mark their output in a machine-readable way so it can be identified as artificially generated. In July 2026 Anthropic signed the EU Code of Practice on Transparency of AI-Generated Content alongside roughly 190 other signatories, including Google and OpenAI, and on 11 August it confirmed how it would comply. A more detailed technical post followed on 15 August.

The practical facts, as Anthropic has stated them:

• It applies everywhere, not just in Europe. Watermarking is enabled wherever Claude is available, across the Claude app, the Claude API, Claude Code, Claude Cowork, and Claude Tag.
• It applies to new models first. Models launched on or after 2 August 2026 carry the mark from release. As of early September Anthropic lists Claude Fable 5.1 and Claude Mythos 5.1 as supporting it, with older models being retrofitted during the transition period, which runs to 2 December 2026 for systems already on the market.
• There is no opt-out. No setting, plan tier, or API parameter disables it, for chat users or for developers building on the API. Products that pass Claude output to their own users carry the mark whether or not they disclose it.
• Anthropic says quality is unaffected. A watermarked response is meant to be indistinguishable from an unwatermarked one to a human reader.

How the Claude watermark actually works

A language model writes one token at a time. At each step it has a list of candidate words with probabilities, and for most steps several candidates would do equally well. "The results suggest" and "The results indicate" are both fine. Normally the model picks among them with a random draw.

The watermark changes where that randomness comes from. Instead of a true coin flip, Claude uses a pseudo-random source driven by a secret key, which nudges the choice among near-equal candidates in a consistent, hidden direction. Anthropic describes it as using the many low-stakes choices in a passage to leave a pattern that is undetectable to the reader but detectable to anyone holding the key. Technically it is a version of SynthID-Text, the scheme Google DeepMind published in Nature in 2024 and has used in Gemini since, with the core idea going back to a 2022 proposal by Scott Aaronson.

Detection is statistical. A detector with the key scores each token against the pattern and adds up the evidence. One or two sentences do not contain enough choices to be sure, which is why Anthropic has indicated the mark is meaningful on passages of roughly 200 tokens, about 150 words, or more. A 1,500-word essay contains far more than enough.

What the watermark is not

Most people who hear "invisible watermark" picture hidden characters: zero-width spaces, unusual apostrophes, doubled spaces. Claude's watermark is none of those. There is no extra character, no metadata, and nothing to find by pasting the text into a hex viewer. Anthropic has been explicit that no Unicode characters are embedded and that the mark lives in the words themselves.

The confusion comes from an older story. In April 2025 the education startup Rumi reported that OpenAI's o3 and o4-mini models were leaving narrow no-break spaces (U+202F) in longer responses. OpenAI said the characters were a training quirk rather than a watermark, and it has separately confirmed that it built a statistical text watermark and chose not to release it. So the "delete the hidden characters" advice was about a ChatGPT formatting artefact. Applied to Claude, it does nothing. Any "Claude watermark remover" that only strips invisible Unicode is cleaning something that was never there.

Files are marked differently

Alongside the text mark, images Claude produces (PNG, JPG, SVG and similar) get C2PA Content Credentials, a cryptographically signed note in the file's metadata saying the file was made or processed with Claude. That is the same open standard camera makers and photo editors use. It is separate from the text watermark, it does not apply to essays, and it is fragile: re-saving, converting, resizing, or screenshotting the image discards the manifest.

Who can detect it, and who cannot

This is the part most coverage skips. Detecting the watermark requires Anthropic's key, so it is not something GPTZero, Originality, or Turnitin can do on their own. Anthropic is offering a watermark detection API, currently in private preview for the groups the EU law names: regulators, law enforcement, media and fact-checkers, independent researchers, educational organisations, and EU civil society groups, plus enterprises that must verify marking for their own compliance. Anthropic says access will expand over time.

Two consequences follow:

• Educational organisations are on the list. A university does not have detection today, but the category is explicitly eligible. Assume that will change.
• The watermark and AI detectors are two different problems. Turnitin and GPTZero do not look for Anthropic's pattern. They score perplexity, burstiness, and other statistical fingerprints of machine writing, which we explain in how AI detectors work. Text can carry no watermark and still be flagged, and in theory the reverse too. A humanizer has to solve both.

Anthropic is being cautious about opening detection widely for a reason the security community calls the "evasion oracle" problem: a freely available detector lets anyone paraphrase, check, and repeat until the mark is gone. One analysis put the cost of a paraphrase pass over a 1,000-word article at about four cents at API rates.

What survives editing, and what does not

Because the mark is a running statistical tally over tokens, its robustness is a matter of degree rather than a switch. Drawing on Anthropic's statements, the SynthID-Text paper, and independent testing reported since launch:

• Copy and paste: survives completely. The mark is in the words, so it goes wherever the words go.
• Light edits (fixing typos, swapping a few words, tightening a sentence): mostly survives. Every untouched run of tokens still carries evidence.
• Cutting sections: survives as long as what remains is over the detection threshold.
• Mixing with your own writing: dilutes the signal. If your own sentences outnumber Claude's, the average evidence per token falls, and the detector may not reach confidence.
• Translation and summarising with another model: weakens or removes it, because the output tokens were chosen by a different process.
• A full rewrite by a different model: removes it. Anthropic itself said the mark "persists through some editing" but that pasting the text into another chatbot and rewriting it could destroy the signal. The Nature paper reports degraded detection under paraphrasing, and later evaluations found sampling-based marks losing detectability after a single meaning-preserving paraphrase pass.

The pattern is simple once you see the mechanism: anything that keeps Claude's exact token sequence keeps the watermark, and anything that replaces the token sequence removes it.

How HumanizeMyPaper removes the Claude watermark

HumanizeMyPaper was built to turn AI drafts into text that reads as naturally written, and it happens to do the one thing that defeats a sampling watermark: it does not edit Claude's tokens, it replaces them.

1. Full regeneration, not word swapping

When you paste a Claude draft, the humanizer rewrites it sentence by sentence with a different model and different randomness. Sentence boundaries move, clause order changes, long sentences split and short ones merge, and vocabulary shifts with the new structure. The output shares the meaning of the input but not its token sequence, so the pattern Anthropic's detector looks for is simply not present in the result. This is also why a thin paraphraser is the wrong tool: a synonym swap leaves long runs of Claude's original tokens intact, and every intact run still carries watermark evidence. We compare the two approaches in AI humanizer vs paraphrasing tool.

You can choose how far the rewrite goes. The Base model is a lighter, faster pass. Standard and Advanced restructure more aggressively and are the right choice for text that will face a detector as well as a watermark check. See the pricing page for what each plan includes.

2. Invisible-character cleanup

The humanizer also normalises exotic whitespace and zero-width characters on the way in and out: non-breaking spaces, narrow no-break spaces, thin and hair spaces, zero-width spaces, and soft hyphens. To be clear, this is not what removes Claude's watermark, because Claude's watermark is not made of characters. It handles the other problem: artefacts that come from chat interfaces and from copying between apps, the kind that made the 2025 ChatGPT story. Your result is plain, clean text.

3. Verification on the detector side

We cannot run Anthropic's detector, because it is in a gated preview, and we would rather say that plainly than pretend otherwise. What we can say is how the mark works and why regenerated text does not carry it. For the statistical-detector side, which is what your reader is far more likely to run, HumanizeMyPaper includes a Turnitin-style AI report you can run on the humanized output before you use it. If a passage still reads as machine-written, humanize it again or edit it in your own voice.

The workflow in practice

1. Paste the Claude draft into the humanizer, or upload the .docx to the Document Editor if you want formatting preserved.
2. Pick a model. Use Standard or Advanced for anything that will be checked.
3. Run it, then read the output. It should say what you meant, in different sentences.
4. Run the AI report on the result.
5. Edit in your own voice: add an example, a specific detail, an opinion. This is the step that makes the text yours, and it is also the step no detector or watermark can argue with.

Should you remove it? The honest answer

The watermark records one fact: these tokens were sampled by Claude. It does not record whether the ideas were yours, whether you were allowed to use AI, or how much you edited afterwards. That cuts both ways.

If you used Claude to polish, restructure, or translate your own thinking and your context permits AI assistance, removing the mark by rewriting is the same as running the text through any other editing tool. Non-native English writers, who already face high false-positive rates from AI detectors, have the strongest case here.

If you are passing off generated work as your own where that is prohibited, a humanizer does not change the rule you are breaking. Turnitin's 2026 update added a category for AI text modified by bypasser tools, and under EU law, organisations publishing AI-generated text to inform the public on matters of public interest have their own disclosure duty that a rewrite does not remove. Our guides to humanizing AI text for Turnitin and whether HumanizeMyPaper is safe cover the academic side in detail.

Summary

• Claude watermarks text since August 2026, on every surface, with no opt-out. New models first, older ones being added.
• The mark is statistical, a hidden bias in word choice based on SynthID-Text. There are no hidden characters to delete.
• Detection needs Anthropic's key and is in private preview for regulators, media, researchers, educational organisations, and compliance-bound enterprises.
• It survives copy-paste and light edits, and it does not survive a full rewrite by a different model.
• HumanizeMyPaper regenerates the text, which replaces the marked token sequence, cleans stray invisible characters, and lets you check the result against a Turnitin-style AI report.

If you have a Claude draft open right now, paste it into the free AI humanizer and compare the two versions side by side. The difference between an edit and a rewrite is the whole story.

Frequently asked questions

Does Claude add a watermark to the text it writes?

Yes. Since August 2026 Anthropic embeds a machine-readable watermark in text generated by new Claude models, on every surface including the Claude app, the API, and Claude Code, worldwide. Anthropic lists Claude Fable 5.1 and Claude Mythos 5.1 as watermarked and is adding older models. There is no opt-out for users or developers.

Can I remove the Claude watermark by deleting hidden characters?

No. Claude's watermark is not made of zero-width spaces, narrow no-break spaces, or metadata. It is a statistical pattern in the words Claude chose while generating. Tools that only strip invisible Unicode characters leave the actual watermark untouched.

Does paraphrasing remove Claude's watermark?

A light edit does not. A full rewrite does, because the watermark lives in the exact sequence of tokens Claude sampled, and text regenerated by a different model with different randomness no longer contains that sequence. Anthropic itself has said rewriting Claude's output with another model can destroy the signal, and the SynthID research the mark is based on reports degradation under paraphrasing.

Can Turnitin or GPTZero detect the Claude watermark?

Not directly. Detecting the watermark requires Anthropic's key, and the detection API is in a private preview for regulators, media, researchers, educational organisations, and enterprises with compliance duties. Turnitin and GPTZero use statistical detectors instead, which is a separate problem a humanizer also has to solve.

How does HumanizeMyPaper remove the Claude watermark?

By regenerating the text rather than swapping words. Every sentence is restructured and rewritten by a different model, so the token sequence that carried Claude's mark is replaced. It also normalises stray invisible characters, and you can run a Turnitin-style AI report on the result to check the statistical-detector side.