ClaudeClean

CLAUDE WATERMARK HUB

Claude watermark explained: what it marks and who can read it

Since August 2026, Anthropic marks the text Claude writes. Here's what that mark is, who can check for it, and where the rest of this guide set fits.

Scan my text free Updated

Short answer

Claude now hides a statistical watermark in its word choices, plus a signed C2PA credential in files it generates. Anthropic says nothing is added to the text and there are no hidden characters. Only Anthropic's secret key can read the text mark, through a detection API that is in private preview. Public tools, classroom detectors included, can't check it.

THREE LAYERS, ONE NAME

What is the Claude watermark, and what does it mark?

People say 'the Claude watermark' about three different things. Only two of them come from Anthropic.

TEXT

A statistical watermark in word choice

Claude's sampling is steered by a secret key and the words before it, so some word choices become slightly more likely. Anthropic calls it a version of Google DeepMind's SynthID-Text, per its watermark explainer.

FILES

A C2PA credential in generated files

Files Claude creates, such as .png, .jpg or .svg, get a signed content credential in their metadata. Anthropic's help center says this applies where file generation is offered, in the Claude apps and the API.

NOT THE WATERMARK

Hidden Unicode characters

Zero width spaces, odd no-break spaces and lookalike letters are a separate, visible-to-code layer. They can enter text from many places, and they are the part you can scan and clean yourself. See the difference.

FOUR WAYS TO LOOK

Who can see which layer of the same output

Pick a lens. It sweeps across the output and shows only what that observer can read.

Illustration

Code points light up

Zero width and no-break spaces are a separate layer. They can enter text from many places, and you can scan and clean them yourself.

Sample sentence, keyed words and code points are illustrative. Captions restate Anthropic's explainer and help center as cited on this page.

The text watermark is the one most people mean. It changes no characters, so a copy and paste keeps it, and Anthropic's help center says the mark will travel with the text when it's copied and pasted elsewhere.

It also says nothing about who you are. Anthropic writes that the watermark carries no identifying information and can't be traced to a specific person, organization or chat.

HOW WE GOT HERE

When did Anthropic start watermarking Claude text?

The short answer is August 2026, pushed by an EU rule. Here is the full sequence, including what other labs did next.

  1. JULY 2026

    Anthropic signs the EU Code of Practice

    Anthropic joined the Code of Practice on Transparency of AI-Generated Content, which its explainer puts at around 190 signatories in total.

  2. AUG 2, 2026

    EU marking duty starts to apply

    Article 50 of the AI Act requires AI providers serving the EU to mark generated content. Claude models launched in the EU from this date support marking at launch.

  3. AUG 11, 2026

    Anthropic announces the text watermark

    Coverage such as the Northeast Times and ABC News reported invisible watermarks in all Claude text, framed around concerns about student use.

  4. AUG 14, 2026

    The technical explainer goes up

    Anthropic's explainer laid out the method, the limits, and why it marks text globally instead of only in the EU.

  5. AUG 15, 2026

    Some users push back

    Gizmodo reported users on X saying they had cancelled. Anthropic told Business Insider it hadn't seen an increase in cancellations.

  6. SEP 1, 2026

    Detection API enters private preview

    Anthropic updated its explainer with a detection API for regulators, law enforcement, media, fact-checkers, researchers, educational organizations and some enterprises.

  7. OCT 5, 2026

    OpenAI adds its own EU text watermark

    OpenAI's EU text provenance post describes textGrain, a statistical signal for eligible ChatGPT and Codex text in the EU, per BleepingComputer.

  8. OCT 7, 2026

    Google opens SynthID Detector to everyone

    Google's announcement covers images, video and audio from several providers. Text is not on the list.

  9. OCT 17, 2026

    Bedrock rollout finishes for newer models

    Anthropic's help center flags some models whose text marking on Amazon Bedrock completes by October 17.

  10. DEC 2, 2026

    EU grace period ends for older systems

    Per Fedil, systems already on the market before 2 August 2026 have until 2 December 2026 to integrate machine-readable marking.

THE LAYER YOU CAN SEE

Check what's sitting in your text right now.

The statistical mark needs Anthropic's key. Hidden Unicode doesn't. Paste your text and ClaudeClean shows every zero width space, odd space and lookalike letter, in your browser.

Map my text's hidden layer Pro is $5/month
  • 300 rewrite credits a month
  • 20,000 characters per request
  • .txt, .docx and .pdf files

BY THE NUMBERS

How big is the AI text watermark rollout?

Four sourced figures that frame how seriously labs and regulators take text marking right now.

16

Claude models listed with text marking

Anthropic help center

~190

Signatories of the EU Code of Practice on AI-generated content

Anthropic explainer

20M

Gemini responses in the SynthID-Text live test, with no quality change found

Nature, 2024

17%

OpenAI's detection rate after 25% of words were swapped

BleepingComputer on OpenAI

The 16 models run from the newest Fable, Mythos, Opus, Sonnet and Haiku releases back to Opus 4.5, Sonnet 4.5 and Haiku 4.5. Older models such as Opus 4.1 and the Claude 3 family aren't on the list.

The 17% figure comes from OpenAI's own tests on its textGrain mark, not Anthropic's. It shows the general shape of statistical watermarks: a 10% synonym swap took detection on 400-token passages from about 92% to 66%, and a 25% swap took it to 17%.

CHECKERS COMPARED

Who can detect the Claude watermark today?

Five tools people mention in the same breath, and what each one can read.

Tools that are said to detect Claude text and what each can read
ToolReads the text watermark?Who can use itWhat it checks
Anthropic detection APIHolds the keyYes, with Anthropic's keyPrivate preview: regulators, media, researchers, educators, some enterprisesLikelihood that Claude was involved at some point
Claude Content CheckerNo, it states it does not check textAnyone, freeC2PA credentials in image, video and audio files up to 100 MB
SynthID DetectorNoAnyone since October 7, 2026Watermarks in images, video and audio from Google and partners
Turnitin, GPTZero, PangramNo, they don't hold Anthropic's keySchools, paid usersTheir own classifiers for AI writing style
ClaudeCleanNoAnyone, 2,000 characters without an accountHidden Unicode characters, plus an optional style rewrite

Swipe to compare →

Anthropic detection API

Holds the key
Reads the text watermark?
Yes, with Anthropic's key
Who can use it
Private preview: regulators, media, researchers, educators, some enterprises
What it checks
Likelihood that Claude was involved at some point

Claude Content Checker

Reads the text watermark?
No, it states it does not check text
Who can use it
Anyone, free
What it checks
C2PA credentials in image, video and audio files up to 100 MB

SynthID Detector

Reads the text watermark?
No
Who can use it
Anyone since October 7, 2026
What it checks
Watermarks in images, video and audio from Google and partners

Turnitin, GPTZero, Pangram

Reads the text watermark?
No, they don't hold Anthropic's key
Who can use it
Schools, paid users
What it checks
Their own classifiers for AI writing style

ClaudeClean

Reads the text watermark?
No
Who can use it
Anyone, 2,000 characters without an account
What it checks
Hidden Unicode characters, plus an optional style rewrite

Anthropic's explainer names Pangram as an example of a detector that lacks the key and looks at stylistic tells instead. No public tool reads the Claude text watermark as of October 9, 2026.

WHO HOLDS THE KEY

A watermarked passage walks past five checkers

Each checker reads something. Only one of them reads the text watermark.

Illustration

Sample: pasted Claude text

The mark sits in word choice, so the passage looks like any other text to every checker without the key.

Starting the walk

  1. ClaudeClean

    Checks: Hidden Unicode characters, plus an optional style rewrite

    Access: Anyone, 2,000 characters without an account

    No key

  2. Turnitin, GPTZero, Pangram

    Checks: Their own classifiers for AI writing style

    Access: Schools, paid users

    No key

  3. SynthID Detector

    Checks: Watermarks in images, video and audio from Google and partners

    Access: Anyone since October 7, 2026

    No key

  4. Claude Content Checker

    Checks: C2PA credentials in image, video and audio files up to 100 MB

    Access: Anyone, free

    No key

  5. Anthropic detection API

    Checks: Likelihood that Claude was involved at some point

    Access: Private preview: regulators, media, researchers, educators, some enterprises

    Reads the mark
Drawn from the comparison table above, as of October 9, 2026. Brand dots are color markers only.

Where ClaudeClean fits

ClaudeClean works on the layer you can inspect: it scans for hidden characters in your browser and can rewrite style through a non-Claude model with a before and after diff. It doesn't read or verify any provider's watermark. No detector result is promised.

WHAT YOU CAN CHECK

Scan the character layer while the watermark stays a black box

Three steps, and the first two never leave your browser.

01

> source text

Paste or drop a file

Paste text or load a .txt, .docx or .pdf. Anonymous use covers 2,000 characters per check.

02

scan report

3 zero-width1 soft hyphencleanable signals2

See every hidden character

Zero width spaces, no-break spaces, soft hyphens, bidi controls and Cyrillic or Greek lookalikes get flagged and cleaned locally.

03

before → after

clearusefulanswer

Copied ✓

Rewrite if the style is the issue

An optional Light, Medium or Strong rewrite runs through a non-Claude model, with a diff so you approve every change.

PRICING

Clean the text free. Rewrite more with Pro.

The scan and hidden-character cleanup are free. Pro adds room for rewrites and files.

PRO

CLAUDECLEAN PRO

ClaudeClean Pro

$5/ month

Cancel anytime.

  • 300 credits every month (≈300,000 characters of rewriting)
  • 20,000 characters per request
  • Clean .txt, .docx and .pdf files
  • Light / medium / strong intensity
  • Dashboard with history
Checking account…

Free: 10 credits a month with an account, or try 2,000 characters right here without one. Strong rewrites cost 2× credits.

UNDER THE HOOD

How does the Anthropic watermark work?

It's a nudge on sampling, not a stamp on the output.

A language model picks each next word by sampling from a list of likely candidates. Normally a random number decides; with the watermark, a secret key combined with the preceding words stands in for that randomness.

Anthropic's example: after "The weather today was cold and", words like overcast or grey are likely, and the key tilts the pick among them. A rare word like nubilous doesn't get favored, so the text reads the same.

Its analogy is swapping Monopoly dice for the digits of pi. The rolls still look random, but anyone who knows the sequence can check them.

The method is a version of SynthID-Text, published by Google DeepMind in Nature in October 2024, and the family goes back to Scott Aaronson's 2022 proposal. That paper reported a latency increase of only 0.57% in its tests.

For a slower walk through the math, Sebastian Raschka's write-up pairs with the video further down this page.

Nothing is added to the text and there are no hidden characters
Source: Anthropic, Claude text watermark explainer

Cost and quality

Anthropic says marking adds no tokens and no price change, has a negligible effect on speed, and showed no quality difference in its internal tests or a side-by-side rater study.

WATCH THE NUDGE

Same words for the reader, a pattern for the key

Anthropic's own example: after "cold and", overcast and grey are both likely. The key tilts the pick between them and never favors a rare word like nubilous.

Illustration
Tinted words carry the keyed lean.
Illustration of the mechanism Anthropic describes in its explainer. Tint strength and bar lengths are drawn for clarity, they are not detector output.

WHAT IT CAN'T DO

Where is the Claude AI watermark strong, and where is it weak?

Anthropic is upfront about the limits. These come from its explainer and help center.

Where the signal holds

  • Longer passages with room for word choice
  • Copying and pasting into other apps
  • Translations, which Anthropic says are watermarked
  • Light editing, which Anthropic says probably won't remove it
  • Code comments, even when the code itself is unmarked

Where it gets thin

  • Short samples, where it's unreliable
  • Factual text with little freedom in wording
  • Proofreading, which may leave too few Claude-chosen words
  • Code where exact output is needed, which gets no mark
  • A full rewrite that replaces every word

BUILD A CASE

Set the conditions, see where the signal lands

Pick who wrote it, how long it is, what kind of text it is and what happened to it afterwards.

Illustration

Who chose the words

Length

Kind of text

What happened next

No markThinHolds

Signal holds

  • Longer passages with room for word choice hold the signal.
  • Copying and pasting keeps it.
Qualitative illustration of the limits Anthropic lists in its explainer and help center. It doesn't predict any detector result.

Two limits matter most for anyone reading a result. The watermark only applies to words Claude chooses, and Anthropic says it cannot distinguish "Claude wrote this" from "Claude heavily edited this".

It also can't spot other AIs. A missing mark doesn't prove a human wrote the text, and the help center lists older models, heavy paraphrasing, short passages and unsupported platforms as reasons a mark may not show.

TERMS

Claude text watermark terms, defined

The vocabulary that shows up across Anthropic, Google and OpenAI documents.

Statistical watermark
A pattern in which words a model picks, invisible to readers and only testable with the key used to make it.
Secret key
The value Anthropic combines with the preceding words to steer sampling. Without it, the pattern looks like normal text.
SynthID-Text
Google DeepMind's text watermark, described in Nature in 2024 and tested on nearly 20 million Gemini responses.
C2PA content credential
A signed record in a file's metadata that says how it was made. Claude adds one to supported files it generates.
Article 50 Code of Practice
The EU's voluntary code for marking AI-generated content under the AI Act. Anthropic signed it in July 2026.
Detection API
Anthropic's private-preview service that returns how likely it is that Claude was involved in a piece of text.
Hidden characters
Unicode code points like U+200B that take no visible space. A separate layer from the statistical watermark.
Stylistic detector
A classifier that guesses AI authorship from phrasing habits. It doesn't use any provider's watermark key.

PICK YOUR PATH

Which Claude watermark guide do you need?

Every guide in this set answers one narrower question. Start with the one that matches yours.

REWRITING

You want a human voice back

What a humanizer can change, and the claims to ignore. Claude humanizer

CHATGPT

Your text came from ChatGPT

OpenAI's EU watermark and the hidden spaces people found in o3 output. ChatGPT watermark remover

PRIMARY SOURCES

Where do these Claude watermark facts come from?

Everything on this page traces back to these documents. Read them directly when the details matter.

Anthropic

Claude text watermark explainer

Method, cost, limits, the detection API preview and the global scope decision. Updated September 1, 2026.

View source

Claude Help Center

How Claude marks AI-generated content

The model table, covered products and cloud partners, plus reasons a mark may be missing.

View source

Claude

Claude Content Checker

Free file checker for C2PA credentials. It states plainly that it does not check text.

View source

Nature

Scalable watermarking for identifying LLM outputs

Dathathri et al., 2024. The SynthID-Text paper that Anthropic's method builds on.

View source

BleepingComputer

OpenAI adds invisible watermarks to ChatGPT and Codex text in the EU

OpenAI's textGrain rollout and its robustness numbers under word replacement.

View source

Fedil

AI Act update: enforcement begins

Where the December 2, 2026 deadline for systems already on the market comes from.

View source
A watermark can only determine that Claude was likely involved with the content at some point.
Source: Anthropic, Claude text watermark explainer

WATCH IT EXPLAINED

See how Claude text watermarking works in 48 minutes

A long, clear walkthrough from next-token sampling to the keyed nudge that makes a watermark.

How Claude's Text Watermarking Works by Sebastian Raschka. Slides and a transcript are on his newsletter

FAQ

Claude watermark FAQ.

No. Anthropic says nothing is added to the text and there are no hidden characters. The mark lives in which words Claude picks, so copying, pasting or changing fonts leaves it in place and nothing shows on screen.
Not with a public tool yet. Anthropic's detection API is in private preview for regulators, media, researchers, educators and some enterprises. The free Claude Content Checker only checks files, not text.
Anthropic's help center lists no opt-out. It says embedded watermarks apply to all generated text, worldwide, across the API, the Claude apps, Claude Code and cloud partners, with some older models still in transition.
No. Anthropic says the watermark carries no identifying information and can't be traced to a person, organization or chat. The key only answers how likely it is that Claude was involved in the text.
Light editing probably won't, per Anthropic. A full rewrite that replaces every word will. Short samples, factual passages and lightly proofread text carry a weaker signal from the start.
ClaudeClean strips hidden Unicode characters in your browser and offers a best-effort rewrite through a non-Claude model with a before and after diff. It cannot read the watermark. No detector result is promised.

Keep reading

Does Claude watermark text?Watermark vs hidden charactersClaude watermark detectorChatGPT vs Claude vs Gemini watermarksWhat is SynthID?Open the cleaner