Formats
Images, video and audio
PPC Land counts 20 supported formats: 11 image, 3 video and 6 audio, with bulk uploads off.
DETECTION, OCTOBER 2026
Claude watermark detector: who can check Claude text (not us)
Anthropic built a real detector for the Claude watermark, and almost nobody can use it yet. Here is who has access, what SynthID and AI checkers can and can't tell you, and what a positive result means.
Scan my text free Updated
claude.ai
Today
Write a product description for noise cancelling headphones.
Message Claude
MacBook Pro
9:41
ClaudeCleanLocal scanPASTED FROM CLAUDE
SCANNING
…hidden characters
U+200Bzero-width space1
U+00A0no-break space1
We show hidden characters only. We do not claim to detect Anthropic's watermark.
9:41
ClaudeCleanLocal scanPASTED FROM CLAUDE
SCANNING
…hidden characters
U+200Bzero-width space1
U+00A0no-break space1
We show hidden characters only. We do not claim to detect Anthropic's watermark.
Short answer
The only Claude watermark detector is Anthropic's own detection API, in private preview for eligible organizations. Google's SynthID Detector checks media, not text, and AI detectors only guess from style. ClaudeClean isn't a watermark detector either: it scans for hidden characters.
WHERE THINGS STAND
Is there a public Claude watermark detector?
Not for individuals. Three companies now run watermark detectors, and each keeps text checks behind an application.
~80%
OpenAI textGrain detection at 200 tokens, at a 1% false positive rate
Anthropic's watermark is a version of Google DeepMind's SynthID-Text, described in Nature in 2024. It nudges which near-equivalent word Claude picks, and only a detector holding Anthropic's key can measure that nudge.
The EU AI Act's Article 50 marking rules apply from August 2, 2026, per the Commission FAQ. That's why Anthropic, OpenAI and Google all moved on provenance this year.
WHO GETS THROUGH WHICH DOOR
Four detectors, and only one reads Claude text
Switching on its own.
Anthropic detection API
Claude text watermark
LockedNot open to individuals
Reads the Claude watermark
Claude Content Checker
C2PA in Claude images, video, audio
OpenAnyone, in the browser, up to 100 MB
Can't read Claude text
Google SynthID Detector
SynthID in Google and partner media
OpenAnyone with a Google sign-in
Can't read Claude text
OpenAI textGrain detector
OpenAI's own text watermark
LockedNot public
Can't read Claude text
SIX TOOLS, SIX JOBS
Which Claude detector should you use?
These get lumped together as detectors, but only one reads the Claude watermark. The rest check other models, other media or other signals.
| Tool | What it checks | Access | Cost and status |
|---|---|---|---|
| Anthropic detection APIReads it | The Claude text watermark, using Anthropic's secret key | Eligible organizations by application | Private preview since August 2026; Anthropic plans to widen access |
| Claude Content Checker | C2PA credentials in images, video and audio from Claude; not text | Anyone, in the browser | Free; files up to 100 MB stay on your device |
| Google SynthID Detector | SynthID marks in images, video and audio from Google and partners | Anyone with a Google sign-in, English | Public since October 7, 2026; no text checks |
| OpenAI textGrain detector | OpenAI's own text watermark in EU ChatGPT and Codex output | Approved researchers and expert organizations, case by case | Not public; OpenAI plans to open-source the method |
| Third-party AI classifiers | Writing style that looks AI-made; no watermark | Public web apps, free tiers and paid plans | Live; error rates vary by tool and text type |
| ClaudeClean scan | Hidden Unicode characters, plus an optional rewrite with a diff | Anyone; 2,000 characters without an account | Scan is free in the browser; Pro is $5/mo |
Swipe to compare →
Only the first row can say anything about the Claude watermark. ClaudeClean is listed because people search for it as a checker, and it checks something else.
WHAT YOU CAN CHECK TODAY
No key needed for the hidden-character layer.
While the watermark API stays in preview, you can still see every zero-width space, soft hyphen and lookalike letter in a draft. The scan runs in your browser and nothing is stored.
- 300 rewrite credits a month
- 20,000 characters per request
- .txt, .docx and .pdf files
NO KEY NEEDED
What can you check in a Claude draft right now?
The character layer. Paste text below to see zero-width spaces, soft hyphens and odd spaces; the watermark itself stays out of reach.
RUNS IN THIS PAGE
Paste a draft and see the character layer
What is hiding
Thanks for the draft.U+202FI tightened the introU+200B and moved the budget table up.
The pilot starts in March,U+00A0with a review after six weeks.
Let me know if the timeU+00ADline works for everyU+FEFFone.
5 hidden characters found
- Narrow no-break space (1)
- Zero-width space (1)
- No-break space (1)
- Soft hyphen (1)
- Byte order mark (1)
GETTING ACCESS
How do you request access to the Claude detection API?
The application is short. The useful part is knowing what to do with your text while you wait.
WHAT APPLYING LOOKS LIKE
Eight kinds of organizations, one locked button
detector-access.example
Organization access
Detection API
Request access
Organization type
RegulatorLaw enforcementMediaFact-checkerIndependent researcherEducational organizationEU civil societyEnterprise with EU AI Act duties
What will you check, and what decision does it feed?
Send request
Run detection
Private preview{
"text": "Paste the passage to check",
"result": "probability Claude was involved"
}
Locked until access is granted
Anthropic plans to expand access over time.
MacBook Pro
- 01
Check that your organization qualifies
Anthropic lists regulators, law enforcement, media, fact-checkers, independent researchers, educational organizations, EU civil society, and enterprises with their own EU AI Act duties. - 02
Write down the use case
Be specific: what content you need to verify, roughly how much, and what decision the result feeds into. A newsroom checking a viral post and a school reviewing essays are different requests. - 03
Submit Anthropic's access form
The Claude Watermark Detector Access Request Form is linked from Anthropic's help center. Anthropic says it plans to expand access over time. - 04
Check files with the free Content Checker
For images, video and audio, the Claude Content Checker reads Claude's C2PA credentials in files up to 100 MB. It states plainly that it does not check text. - 05
Treat AI classifiers as a hint
Until you have API access, tools like GPTZero or Pangram can point at passages worth a closer look. Pair any score with a human read and the writer's drafts. - 06
Clean the character layer separately
Hidden characters such as zero-width spaces are a different thing from the watermark. A local scan finds and removes them without touching the word choices.
GOOGLE'S PORTAL
Can the SynthID Detector check Claude text?
It's the most searched detector in this space, and the answer for Claude text is no. Here's what it does cover.
Partners
Google plus a short list
Google names OpenAI, NVIDIA and Kakao as partners, with Apple coming. Anthropic isn't on the list, and Claude files carry C2PA credentials instead.
Text
No text box
PPC Land describes the portal as not a general AI detector. Pasting a Claude essay into it isn't an option.
Open source
SynthID-Text is public code
Google released SynthID-Text in Hugging Face Transformers, as the video below covers. See DeepMind's SynthID page. Each provider keeps its own key, so the code alone can't read Claude's mark.
anyone can easily check if an image, video, or audio file was made with AI from Google or our partners
The portal first opened to early testers in 2025, per Google's original post. The October 2026 launch made it available to everyone, globally, in English.
WHAT THE PORTAL TAKES
20 formats. None of them text.
PARTNERS IT CHECKS
- GoogleNamed
- OpenAINamed
- NVIDIANamed
- KakaoNamed
- AppleComing
- AnthropicNot listed
Claude files carry C2PA credentials instead, which the Claude Content Checker reads.
TERMS THAT GET MIXED UP
What's the difference between a watermark detector and an AI detector?
Most confusion about Claude detection comes from these six words being used loosely.
TWO WAYS TO GET A RESULT
A key measures. A classifier guesses.
Originality.ai: 19 of 495 human passages flagged as AI
- Keyed detector
- A detector that holds the same secret key used to mark the text. It measures how often word choices line up with that key. Anthropic's API is one.
- Classifier
- A model trained on examples of human and AI writing that predicts which one a passage resembles. GPTZero, Pangram and Turnitin work this way.
- False positive
- Human text flagged as AI. In Epoch AI's July 2026 test, Originality.ai flagged 19 of 495 human passages while Pangram and GPTZero flagged none.
- False negative
- AI text that passes as human. For watermarks it rises as text gets shorter or more heavily edited; for classifiers it rises when the model imitates a style.
- Threshold
- The cut-off a score must clear to count as a hit. OpenAI reports textGrain results at a 1% false positive rate, per its technical report and coverage.
- C2PA Content Credentials
- Signed metadata attached to a file, not to the words. Anthropic uses them for Claude images and files, and stripping metadata or taking a screenshot removes them.
READING A RESULT
What does a positive watermark result prove?
Anthropic is unusually direct about this. A hit is evidence of involvement, not authorship.
What a hit can tell you
- Claude was likely involved. Anthropic's words: the watermark can only determine that Claude was likely involved with the content at some point.
- It's a probability. The API returns how likely it is that Claude was at least partly involved, not a yes or no.
- It survives light edits. Anthropic says light editing probably won't remove the mark.
- Translations count. When Claude translates text, its output is watermarked like any other.
What it can't tell you
- Who wrote it. It cannot distinguish “Claude wrote this” from “Claude heavily edited this,” per Anthropic.
- That a human wrote it. A missing mark doesn't confirm human authorship, and Anthropic says so.
- Anything about other AIs. It can't detect GPT, Gemini or any other model's text.
- Intent or rule-breaking. Claude may have only proofread, translated or summarized. The help center calls a detected mark not fully conclusive.
READ A RESULT
Same check, five histories
The passage
Our survey found that most readers skim long reports before they decide.
Signal: Spread through the passage
The check can say
Claude was likely involved at some point
It can't tell you
Whether Claude wrote it or heavily edited it
WHERE IT GOES QUIET
When does the Claude watermark detector miss?
These limits come from Anthropic's own explainer and help center, not from outside testing.
Length
Short text
Anthropic says the signal is unreliable on short passages. A tweet or a two-line reply rarely carries enough word choices to score.
Content
Fact-dense writing
Dates, names and figures leave few interchangeable words, so the mark gets sparse in tables, lists and reference text.
Code
Less marking in code
Code has fewer safe word swaps than prose, and Anthropic says it watermarks code less.
Mixed text
Only Claude's edits carry it
When Claude edits human writing, only the words Claude changed are marked. The rest of the document stays unmarked.
Rewrites
A full rewrite removes it
Anthropic says a rewrite that replaces every word removes the mark. OpenAI reports a similar curve: swapping 25% of words dropped textGrain detection to 17%.
Older models
Pre-August models lag
Models launched before August 2, 2026 get the mark over the coming months. The help center table shows which ones are marked today.
WHAT CLAUDECLEAN CHECKS
A hidden-character checker, not a watermark detector.
ClaudeClean looks at the characters in your text, which is where copy-paste junk hides. It can't see the watermark. No detector result is promised.
01
> source text
Paste text or drop a file
Works with pasted text and .txt, .docx or .pdf files. Anonymous use covers 2,000 characters at a time.
02
scan report
3 zero-width1 soft hyphencleanable signals2See every invisible character
Zero-width spaces, soft hyphens, odd whitespace and lookalike letters get highlighted in your browser, then removed in one click.
03
before → after
clearusefulanswerfocusedhelpfulresponse
Copied ✓Rewrite only if you want to
An optional rewrite runs on OpenAI, not Claude, in up to three passes. A diff shows every word that changed.
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.
PRIMARY SOURCES
Sources for every claim on this page
Company posts for how each detector works, and independent coverage for the numbers companies published elsewhere.
Anthropic
Claude text watermark explainer
The method, the private preview, eligible groups and the limits listed above.
View sourceClaude Help Center
How Claude marks AI-generated content
Model table, reasons a mark can be missing, and the link to the access request form.
View sourceClaude
Claude Content Checker
Free C2PA checker for Claude files up to 100 MB. It does not check text.
View sourceSynthID Detector opens to everyone
October 7, 2026: partners, scale numbers and who can use the portal.
View sourceOpenAI
textGrain technical report
The keyed detection method behind OpenAI's EU text watermark.
View sourceNature
Scalable watermarking for identifying LLM outputs
The 2024 SynthID-Text paper that Anthropic's watermark builds on.
View sourceAlso used: PPC Land on SynthID Detector formats, BleepingComputer on textGrain's numbers, and OpenAI's announcement.
WATCH
How SynthID text watermarking works
Google for Developers walks through SynthID for text, the method Anthropic's watermark is based on, in under eight minutes.
FAQ
Claude watermark detector FAQ.
Yes, but it isn't public. Anthropic's detection API is in private preview for eligible organizations such as regulators, media, researchers and educational groups. Individuals can't run a watermark check yet.
Not through the watermark unless your organization has API access. AI detectors like GPTZero or Pangram guess from style, and every one of them gets some passages wrong in both directions.
No. Google's public SynthID Detector checks images, video and audio from Google and its partners. It doesn't check text, and Anthropic isn't among the partners Google named in October 2026.
No. A watermark detector checks word choices against Anthropic's secret key. A Claude AI detector is a classifier trained on writing style, so it never sees the watermark and can flag human text.
Fill in Anthropic's access request form for the Claude watermark detector. Anthropic lists regulators, law enforcement, media, fact-checkers, researchers, educational organizations, EU civil society and some enterprises.
No. ClaudeClean scans for hidden characters in your browser and offers a best-effort rewrite with a non-Claude model. It can't read Anthropic's watermark, and no detector result is promised.