01
> source text
Paste the exact copy
Use the text as it will be submitted, or load the .docx, .pdf or .txt. Up to 2,000 characters work without an account.
SCOPE CHECK
Does Claude watermark text? Yes, on 16 models so far
The answer depends on which model wrote it and where. Here's Anthropic's own list, by model and by surface, plus what a mark can and can't tell anyone.
Scan my text free Updated
claude.ai
Today
Explain photosynthesis to a 10 year old.
Message Claude
MacBook Pro
9:41
ClaudeCleanLocal scanPASTED FROM CLAUDE
SCANNING
…hidden characters
U+200Bzero-width space1
U+00A0no-break space1
Character scan is clean. The token-level signal needs a rewrite to change.
9:41
ClaudeCleanLocal scanPASTED FROM CLAUDE
SCANNING
…hidden characters
U+200Bzero-width space1
U+00A0no-break space1
Character scan is clean. The token-level signal needs a rewrite to change.
Short answer
Yes. Anthropic's help center lists 16 Claude models whose generated text carries an embedded watermark, from Opus 4.5 and Sonnet 4.5 up to the 5.x releases. It covers the Claude apps, the API, Claude Code and cloud partners, worldwide. The mark sits in word choice, so you can't see it, and only Anthropic's key can read it.
ANTHROPIC'S OWN TABLE
Which Claude models watermark text?
The help center groups models by what they mark today. We grouped the 16 rows the same way so the table stays readable.
| Models | Text, first-party | Text, cloud partners | C2PA on files |
|---|---|---|---|
| Fable 5.1, Mythos 5.1, Opus 5.5, Opus 5, Sonnet 5.5, Haiku 5.5Fully marked | Yes | Yes | Yes |
| Fable 5, Opus 4.8, Sonnet 5 | Yes | Yes, Bedrock rollout completes by October 17 | Yes |
| Mythos 5, Opus 4.7, Opus 4.6, Opus 4.5, Sonnet 4.6, Sonnet 4.5, Haiku 4.5 | Yes | Not listed | Not listed |
| Opus 4.1, Sonnet 4, Claude 3 family | Not in the table | Not in the table | Not in the table |
Swipe to compare →
Source: Anthropic help center, read October 9, 2026. Anthropic says older models are in a transition and marking rolls out over the coming months, so this list will grow.
THE 16, AT A GLANCE
Which models light up for each kind of marking
Pick a marking type. Tiles light up for every model Anthropic lists with it.
0of 16 listed models
Claude apps, the API, Claude Code
Claude apps, the API, Claude Code
- Fable 5.1
- Mythos 5.1
- Opus 5.5
- Opus 5
- Sonnet 5.5
- Haiku 5.5
- Fable 5
- Opus 4.8
- Sonnet 5
- Mythos 5
- Opus 4.7
- Opus 4.6
- Opus 4.5
- Sonnet 4.6
- Sonnet 4.5
- Haiku 4.5
- Opus 4.1Not in the table
- Sonnet 4Not in the table
- Claude 3 familyNot in the table
bars: first-party text, cloud text, C2PA not listed for this type
The rule of thumb from Anthropic's help center: models launched in the EU on or after August 2, 2026 support marking at launch. Models from before that date are in a transition.
The help center states that embedded watermarks will apply to all generated text. It names no opt-out and no plan-level exemption.
APP, API, CODE
Does Claude watermark text in the app, the API and Claude Code?
Same answer for every product Anthropic runs, with a few gaps on partner clouds and file types.
| Where you use Claude | Text watermark | File credential (C2PA) | Worth knowing |
|---|---|---|---|
| Claude appsCovered | Yes, for listed models | Yes, where file generation is offered | Web, desktop and mobile chats all count as first-party |
| Claude Platform (API)Covered | Yes, for listed models | Yes | No extra tokens and no price change, per Anthropic |
| Claude Code | Yes, for prose and comments | Not described | Code where exact output is needed gets no watermark |
| Claude Cowork and Claude Tag | Yes, for listed models | Not described | Named in the help center's covered products |
| AWS, Google Cloud, Microsoft Foundry | Yes, for 9 of the 16 models | Claude Platform on AWS and Claude in Microsoft Foundry | Bedrock rollout for some models completes by October 17 |
Swipe to compare →
Anthropic's help center adds that some platforms or features may not support certain marking types. For code specifics, see does Claude watermark code.
ONE MODEL, FIVE EXITS
Where marked text comes out
Follow a response from a listed model to each surface Anthropic names.
A listed Claude model
Word choice is marked at generation
Claude apps
Marked, listed modelsC2PA where file generation is offered
Web, desktop and mobile
Claude Platform (API)
Marked, listed modelsC2PA: yes
No extra tokens, no price change
Claude Code
Prose and comments markedExact code: no watermark
Files: not described
Cowork and Claude Tag
Marked, listed modelsFiles: not described
Named as covered products
AWS, Google Cloud, Microsoft Foundry
Marked for 9 of the 16Bedrock: some models by October 17
C2PA on Claude Platform on AWS and Foundry
ONE QUESTION YOU CAN ANSWER
The watermark is hidden. The characters don't have to be.
Nobody outside Anthropic's preview can test the mark. You can still see every zero width space and lookalike letter in the exact text you copied, checked in your browser.
- 300 rewrite credits a month
- 20,000 characters per request
- .txt, .docx and .pdf files
THE FOLLOW-UP QUESTIONS
Is Claude text watermarked after you copy, translate or edit it?
The questions people ask right after the first yes, answered from Anthropic's own documents.
Does the watermark survive copy and paste?
Yes. Anthropic's help center says the mark will travel with the text when it's copied and pasted elsewhere, because no characters are involved. Changing the font, the app or the file format doesn't touch word choice.
Is translated text watermarked?
If Claude did the translation, yes. Anthropic's explainer lists translations as marked text, since Claude is still choosing the words.
Does a quick edit remove it?
Anthropic says light editing probably won't, while a full rewrite that replaces every word will. The help center adds that the mark may persist through some editing, so it treats heavy edits as a gray zone.
Is it only switched on in the EU?
No. Article 50 of the EU AI Act drove the change, but Anthropic marks worldwide because it found no durable way to scope it by region.
Does watermarking change Claude's answers?
Anthropic reports no quality difference in internal tests or in a side-by-side rater study, and a negligible effect on speed. The mark tilts choices between words that were already likely.
Does it carry my name or account?
No. Anthropic says the watermark carries no identifying information and can't be traced to a specific person, organization or chat.
FROM CHAT TO NOTES
What happens to the mark when you copy a reply
A reply goes from a chat app into a notes app and gets one small edit. The words carry the mark the whole way.
9:41
Claude writes the reply
MarkedThe mark sits in which words were picked, with no characters added.
You tap Copy
TravelsThe help center says the mark travels with copied and pasted text.
It lands in another app
TravelsA new app, font or file format doesn't touch word choice.
You change a word or two
Probably staysLight editing probably won't remove it. Only a rewrite of every word will.
- Not in this loop: a full rewrite that replaces every word removes the mark, and a Claude translation keeps it.
READ THIS BEFORE A VERDICT
Does no watermark mean a human wrote it?
No, and Anthropic says so directly. A missing mark has plenty of ordinary explanations.
A missing mark proves nothing on its own
Anthropic's help center lists the reasons a mark may not show: the model predates marking, the text was heavily edited, paraphrased or translated, the passage is very short, metadata was stripped, or the platform or file type isn't supported.
A detected mark is narrower than it sounds too. It means the text may have been generated or processed by Claude, which includes proofreading, translating, summarizing or converting.
The watermark also can't name other AIs. Text from ChatGPT or Gemini carries those providers' own marks, if any, and Anthropic's key says nothing about them.
For classroom tools, the short version is that they don't hold Anthropic's key. The longer version is in can Turnitin detect the Claude watermark.
THE REGULATION BEHIND IT
Why does Claude watermark text now?
An EU transparency rule set the date. Anthropic chose to apply it everywhere.
Article 50 of the EU AI Act requires providers of generative AI to mark synthetic content in a machine-readable way. Anthropic's explainer says that as of August 2, the EU requires AI providers serving its market to mark AI-generated content.
Anthropic signed the EU's Code of Practice on the transparency of AI-generated content in July 2026, which its help center ties to Article 50(2). Signing it is one way a provider can show it meets the rule.
Older systems get a grace period. Per Fedil, systems already on the market before 2 August 2026 have until 2 December 2026 to integrate machine-readable marking, which lines up with Anthropic's transition for older models.
Why it isn't EU-only
Anthropic's explainer says there's no durable way to scope the watermark by region. Text written in one country gets read, pasted and published in others, so the marking runs for every user.
The same explainer says the watermark doesn't change ownership, authorship or legal responsibility for the text.
JULY TO DECEMBER 2026
The dates behind the watermark
Pick a date. The orange band is the transition window for models that were on the market before August 2.
October 9
16 models in the table
The help center lists 16 models with text marking. Older models are still in a transition.
Source: Anthropic help center, read October 9, 2026
THE PART YOU CAN CHECK
See what's in your Claude text before anyone else scores it
ClaudeClean leaves the watermark question to Anthropic and shows you the layer that's visible to code.
02
scan report
3 zero-width1 soft hyphencleanable signals2Read the scan report
Each zero width character, odd space, soft hyphen or lookalike letter gets listed, then cleaned, without the text leaving your browser.
03
before → after
clearusefulanswerfocusedhelpfulresponse
Copied ✓Decide on a rewrite
If the wording itself is the concern, run a best-effort rewrite through a non-Claude model and review the diff line by line.
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.
WHAT A RESULT MEANS
What can a Claude text watermark tell someone?
Anthropic describes the detection result as one question with one kind of answer.
It can tell
- Likely Claude involvement. The key answers how likely it is that Claude wrote part of the text.
- Across copies. The mark travels with pasted text into other apps and formats.
- Across languages. Claude translations carry it, since Claude chose the words.
- In longer passages. More Claude-chosen words means a stronger signal.
It can't tell
- Who you are. No account, organization or chat is encoded.
- Wrote vs edited. It cannot separate Claude writing from Claude heavily editing.
- Other AIs. It says nothing about ChatGPT, Gemini or any other model.
- Short or fixed text. Short samples are unreliable and exact code isn't marked.
EIGHT QUESTIONS, ONE KEY
Ask the watermark key a question
Pick a question someone might put to a detection result. The key answers some and is silent on the rest.
Anthropic's key
Only this key can read the mark
Did Claude likely write part of this?
It can tell
This is the one question it answers: how likely it is that Claude was involved with the text.
Anthropic's wording is precise here: a watermark can only determine that Claude was likely involved with the content at some point. Treat any claim stronger than that with suspicion.
SIX STEPS
How to check whether your Claude text is likely watermarked
You can't run the detector yourself, but you can work out what you're holding with Anthropic's public information.
- 01
Find the model that wrote it
In the Claude apps, the model picker shows which model you used. In an API response, the model field names it. Write it down along with the date. - 02
Match it against Anthropic's table
Compare that model with the help center table at the top of this page. If it's one of the 16, first-party text from it is marked; if it isn't listed, it may still be in transition. - 03
Note the surface it came through
Claude apps, the API, Claude Code and cloud partners are all covered, but cloud partner marking is listed for fewer models. Bedrock finishes rolling out some models by October 17. - 04
Check files in the Content Checker
If Claude generated an image, audio or video file, the free Claude Content Checker reads its C2PA credential. It does not check pasted text. - 05
Ask about detection API access
For text, only Anthropic's detection API can test the mark. It's in private preview, so the realistic path is an institution that already has access. - 06
Scan the character layer separately
Hidden characters are a different question with a checkable answer. A local scan shows any zero width spaces, odd spaces or lookalike letters in the exact string you have.
What the Content Checker accepts
JPG, PNG, GIF, WEBP, TIFF, HEIC, AVIF, SVG, DNG, JXL, MP4, MOV, AVI, WAV, MP3, M4A and FLAC files up to 100 MB. Its own page states: "The tool does not check text."
THE NEWS CLIP
Watch the original Claude watermark announcement coverage
ABC News covered the plan the day it was announced, framed around student use.
Written coverage the same day, such as the Northeast Times, reported invisible watermarks in all Claude AI text. Anthropic's detailed explainer followed on August 14.
CITED ON THIS PAGE
Where does the Claude model list come from?
The model table and surface details come from Anthropic's help center. The rest is linked here.
Claude Help Center
How Claude marks AI-generated content
The 16-model table, covered products, cloud partners and the reasons a mark may be missing.
View sourceAnthropic
Claude text watermark explainer
Method, limits, translations, code, and why marking runs worldwide instead of only in the EU.
View sourceClaude
Claude Content Checker
The free file checker for C2PA credentials, with its supported formats and 100 MB limit.
View sourceAI Act Explorer
Article 50: transparency obligations
The EU provision that requires providers to mark AI-generated content in a machine-readable way.
View sourceFedil
AI Act update: enforcement begins
Systems already on the market before August 2, 2026 have until December 2, 2026 to add marking.
View sourceNortheast Times
Anthropic will embed invisible watermarks in all Claude AI text
Same-day coverage of the August 11, 2026 announcement.
View sourceFor the bigger picture, including the timeline and who can read the mark, start at the Claude watermark guide.
FAQ
Does Claude watermark text? FAQ.
Anthropic's help center says embedded watermarks will apply to all generated text from listed models. It also says some platforms or features may not support every marking type, and older models are still in a transition period.
No. An EU rule triggered it, but Anthropic marks text worldwide. Its explainer says there's no durable way to scope the watermark by region, so the same models are marked everywhere.
Opus 4.1, Sonnet 4 and the Claude 3 family don't appear in Anthropic's model table as of October 9, 2026. Anthropic says marking for older models rolls out over the coming months, so recheck the help center.
Anthropic says nothing is added to the text and there are no hidden characters. Zero width spaces or odd spaces in a draft are a separate issue, and they can come from editors, web pages or copy paths.
Not by yourself today. Only Anthropic's detection API reads text marks, and it's in private preview for groups like regulators, media and researchers. The free Claude Content Checker handles files only.
No. ClaudeClean scans hidden Unicode characters in your browser and can rewrite style through a non-Claude model with a diff. It can't read Anthropic's watermark. No detector result is promised.