ANTHROPIC
Claude: global, always on
Applied in every region, with no hidden characters. A private preview API serves regulators, media, researchers and obligated enterprises via an interest form. Translations by Claude are marked too.
UPDATED FOR OCTOBER 2026
ChatGPT vs Claude watermark, plus Gemini SynthID Text
Three of the biggest chatbots now hide a statistical watermark in their text, but they do it in different places, by default or not, and with different rules on who can check. Here is the side by side, vendor by vendor.
Scan my text free Updated
chatgpt.com
Today
Compare two laptops for a student in one paragraph.
Message ChatGPT
MacBook Pro
9:41
ClaudeCleanLocal scanPASTED FROM CHATGPT
SCANNING
…hidden characters
U+202Fnarrow no-break space1
U+200Bzero-width space1
Same scan for ChatGPT, Claude and Gemini text. Statistical marks need a rewrite.
9:41
ClaudeCleanLocal scanPASTED FROM CHATGPT
SCANNING
…hidden characters
U+202Fnarrow no-break space1
U+200Bzero-width space1
Same scan for ChatGPT, Claude and Gemini text. Statistical marks need a rewrite.
Short answer
Claude watermarks text worldwide with a version of Google's SynthID Text. ChatGPT and Codex watermark eligible text only in the EU, using OpenAI's textGrain from October 2026; the API is opt-in. Gemini uses SynthID Text in its app. Meta, Mistral and DeepSeek have announced no chatbot text watermark.
THE FULL MATRIX
Which AI models watermark text in 2026?
Seven products, five questions. Every cell comes from the vendor's own page or the EU signatory list, checked October 9, 2026.
| Product | Method | Where it applies | Default | Who can check | Open source |
|---|---|---|---|---|---|
| Claude (Anthropic) | Statistical, a version of SynthID Text | All regions, models launched from Aug 2, 2026; older models over the coming months | Always on | Private preview API for eligible organizations | Method is public; key is private |
| ChatGPT and Codex (OpenAI) | Statistical, textGrain | EU only, rolling out over weeks from Oct 5, 2026 | On for eligible EU output, all plans | Approved researchers and expert organizations by application | Technical report out; open source planned |
| OpenAI API | Statistical, textGrain | Worldwide, select models | Off; customer opts in | Enabling it does not grant detector access | Same as above |
| Gemini (Google) | Statistical, SynthID Text | Gemini app and web experience | On | Public SynthID Detector covers image, video, audio; text not listed | Yes, Transformers v4.46.0+ and GitHub |
| Meta AI and Llama | No public text watermark on Meta AI; TextSeal research toolkit | Not announced | Not announced | None public | TextSeal is Apache 2.0 on GitHub |
| Mistral (Le Chat) | No public text watermark announced | Not announced | Not announced | None public | Not applicable |
| DeepSeek and Grok (xAI) | No public text watermark announced | Not announced | Not announced | None public | Not on the EU signatory list we checked |
Swipe to compare →
Not announced means we found no vendor statement of a text watermark on that chatbot. It does not prove the absence of one.
VENDOR BY VENDOR
How do Claude, ChatGPT and Gemini watermarks differ?
Same family of math, three different policies. The other big names are mostly silent.
SAME PROMPT, THREE APPS
Flip the region and watch which reply stays marked
Switching on its own.
9:41
ClaudeEU
Why is the sky blue? Two sentences.
Sunlight scatters off air molecules, and blue light scatters far more than red. That scattered blue reaches your eyes from every part of the sky.
A version of SynthID Text
Watermarked
Private preview API for eligible organizations
Reply to Claude
9:41
ChatGPTEU
Why is the sky blue? Two sentences.
Air molecules scatter short blue wavelengths much more strongly than long red ones. So when you look up, that scattered blue light fills the sky.
textGrain
Watermarked in the EU
Approved researchers, by application
Reply to ChatGPT
9:41
GeminiEU
Why is the sky blue? Two sentences.
The atmosphere scatters blue light more than other colors because its waves are shorter. Our eyes catch that light coming from all directions.
SynthID Text
Watermarked in the app
Public Detector lists media, not text
Reply to Gemini
9:41
ChatGPTEU
Why is the sky blue? Two sentences.
Air molecules scatter short blue wavelengths much more strongly than long red ones. So when you look up, that scattered blue light fills the sky.
textGrain
Watermarked in the EU
Approved researchers, by application
Reply to ChatGPT
OPENAI
ChatGPT: EU only, for now
OpenAI says it is not making text watermarking a global default at launch. Outside the EU, ChatGPT text stays unmarked unless an API customer turns textGrain on.
Gemini: the original SynthID
SynthID Text runs in the Gemini app and is open source, with a Bayesian detector that returns watermarked, not watermarked or uncertain.
META
Meta: tools, not a rollout
Meta signed the EU marking section and maintains TextSeal for research. We found no statement that Meta AI chat replies carry a text watermark.
MISTRAL
Mistral: signed, quiet
Mistral appears among the EU code signatories, but has published no text watermark method or rollout for Le Chat that we could find.
DEEPSEEK AND XAI
DeepSeek and Grok: nothing public
Neither showed up on the EU signatory list we searched, and neither has announced a text watermark. Their EU obligations still apply if they serve the EU market.
WHATEVER THE MODEL
One cleaner for ChatGPT, Claude and Gemini drafts.
Watermarks differ by vendor, but hidden characters and AI writing habits look the same everywhere. Scan any draft in your browser, then rewrite it only if you want to.
- 300 rewrite credits a month
- 20,000 characters per request
- .txt, .docx and .pdf files
SHARED MATH
Is Claude's watermark the same as Gemini's SynthID?
Same technique, different keys. That one detail decides who can detect what.
Claude’s text watermark is a version of the SynthID-Text approach published by Google DeepMind in a Nature paper in 2024.
WHO CAN READ WHOSE MARK
Same math, three locks
Reads it. Anthropic's key marked this text.
A matching key is the only way to measure the pattern in which words won.
SHARED
Same trick with randomness
Both pick each next word from the usual probabilities, but seed the random choice from a secret key and the preceding words. The text reads the same; the pattern sits in which equally good word won.
SEPARATE
Different keys, no cross-reading
Anthropic notes that another AI using watermarking would have a different key. So Google's detector cannot vouch for Claude text, and Claude's key says nothing about Gemini.
OPENAI'S CHOICE
Why OpenAI built its own
OpenAI says textGrain gives it more control than SynthID or TextSeal over the balance between detectability and the variety of answers to the same prompt.
FOUR YEARS IN TEN STEPS
When did each AI company start watermarking text?
Google went first in 2024. The EU AI Act pulled Anthropic and OpenAI in during 2026.
2022
Scott Aaronson's proposal
Anthropic traces the whole family of statistical text watermarks back to Aaronson's 2022 idea: change only the source of randomness used to pick words.May 14, 2024
Google adds SynthID to Gemini text
DeepMind announced SynthID watermarking for AI text in the Gemini app and web experience, alongside video.Aug 4, 2024
OpenAI holds its method back
OpenAI said its watermark resisted paraphrasing but not translation or rewording by another model, and could stigmatize non-native English speakers.Oct 2024
SynthID Text goes open source
The Nature paper reported a live test on about 20 million Gemini responses, and Google released the code in Hugging Face Transformers and on GitHub.2025
SynthID Detector portal
Google launched a verification portal for journalists, then opened it to everyone for images, video and audio. Text is not on its list.Dec 2025
Meta publishes TextSeal
Meta's Text Seal repo covers generation-time and post-hoc text watermarking, but no rollout to Meta AI has been announced.Jul 2026
About 190 sign the EU transparency code
The European Commission lists Anthropic, Google, Meta, Microsoft, Mistral and OpenAI among signatories of the marking section.Aug 2026
Claude watermarked worldwide
EU AI Act Article 50 applies from August 2. Anthropic marks Claude text in every region because it has no durable way to scope by region yet.Oct 5, 2026
OpenAI ships textGrain in the EU
OpenAI starts EU ChatGPT and Codex watermarking, opens a worldwide API opt-in and takes detector applications.Dec 2, 2026
Legacy systems must comply
iubenda summarizes the omnibus deadline: Article 50(2) watermarking for systems launched before August 2 applies from this date.
NUMBERS VENDORS PUBLISHED
How reliable are AI text watermarks?
Every figure here is a vendor's own test. No independent body has benchmarked all three side by side.
LENGTH
Short text barely registers
Anthropic and OpenAI both say confidence grows with length. The EU code does not even require marks below 200 tokens, about 150 English words.
FACTS AND CODE
Exact answers carry little signal
When only one word is right, there is nothing to nudge. Google, Anthropic and OpenAI all flag factual text, math and code as weak spots.
TRANSLATION
Rewriting and translation erode it
Google says detector confidence can drop greatly after thorough rewriting or translation. Anthropic says a full rewrite removes it, though a Claude translation is itself marked.
WHERE THE SIGNAL THINS OUT
Three weak spots, playing on a loop
Third-party detectors are a different test
GPTZero, Turnitin, Pangram and similar tools run classifiers on writing style; none of them holds a vendor key. ClaudeClean only cleans characters and rewrites style on request. No detector result is promised.
SAME FLOW, ANY CHATBOT
Paste text from ChatGPT, Claude or Gemini and clean it the same way
The vendor changes the watermark, not the cleanup. Hidden characters and AI habits are model-agnostic.
01
> source text
Paste from any chatbot
Drop in a ChatGPT, Claude or Gemini answer, or a .txt, .docx or .pdf. 2,000 characters free without an account.
02
scan report
3 zero-width1 soft hyphencleanable signals2Scan for hidden characters
Zero-width marks, odd spaces, soft hyphens and homoglyphs are flagged and removed in your browser, never stored.
03
before → after
clearusefulanswerfocusedhelpfulresponse
Copied ✓Rewrite if you choose
Light, Medium or Strong passes remove AI habits, with a diff to review. Rewriting is best effort and runs through OpenAI.
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.
PLAIN ENGLISH
What do SynthID Text, textGrain and C2PA mean?
The terms vendors use in their announcements, defined once.
WHERE A MARK CAN LIVE
An image file has two homes for a signal. Text has one.
- SynthID Text
- Google DeepMind's text watermark. It works as a logits processor during generation, described in the Nature paper and documented for developers by Google AI for Developers.
- textGrain
- OpenAI's own text watermark. It creates several reweighted sets of word probabilities that average back to the original, and a secret key picks which set each word comes from.
- TextSeal
- Meta's open toolkit for generation-time and post-hoc text watermarking plus contamination detection. A research codebase, not a product rollout.
- C2PA Content Credentials
- Signed metadata from the C2PA standard. OpenAI adds it to images and Anthropic adds it to .png, .jpg and .svg files Claude makes. Plain text has no metadata slot, which is why text needs a statistical mark.
- Article 50(2)
- The EU AI Act clause requiring providers to mark generated content in a machine-readable way. It applies from August 2, 2026 for new systems.
- False positive rate
- How often a detector flags unmarked text as marked. OpenAI reports its numbers at a 1% target, meaning roughly 1 in 100 unmarked passages could be flagged.
WHO CAN CHECK WHAT
Can you tell if text came from ChatGPT, Claude or Gemini?
Short answer: not yourself, not today. Here is who holds each key.
Can I check Claude text?
Only through Anthropic's detection API, which is in private preview for eligible organizations such as regulators, media, fact-checkers, researchers and educational bodies, not to individual users.
Can I check ChatGPT text?
Not publicly. OpenAI's text detector is limited to approved researchers and expert organizations. openai.com/verify handles images and audio only.
Can I check Gemini text?
Google's public SynthID Detector lists images, video and audio. Developers can run the open-source SynthID Text detector, but only on text from their own watermarked models with their own key.
Does a negative result prove a human wrote it?
No. OpenAI says text may be too short, edited, translated, made before watermarking began or produced by another company's model.
WATCH
A six-minute news explainer on all three watermarks
ThePrint's walkthrough, published the day after OpenAI's announcement.
SOURCES
Primary sources for each vendor
Vendor documentation first, then the EU record.
Anthropic
Claude text watermark
SynthID-based method, global scope, code and translation behavior, private preview detection API.
View sourceOpenAI
Our approach to EU text provenance rules
Oct 5, 2026: EU ChatGPT and Codex watermarking, API opt-in, detector applications, edit robustness.
View sourceOpenAI Help Center
Provenance signals in OpenAI-generated content
Signals per content type, language results, why textGrain over SynthID or TextSeal.
View sourceGoogle DeepMind
SynthID
Gemini app text watermarking and the SynthID Detector portal for image, video and audio.
View sourceNature
Scalable watermarking for identifying large language model outputs
The 2024 SynthID Text paper with the 20 million response live experiment.
View sourceEuropean Commission
Strong backing for the Code of Practice on Transparency of AI-generated Content
Signatory list and counts for the Code of Practice on Transparency of AI-Generated Content.
View sourceFAQ
ChatGPT vs Claude watermark FAQ.
Similar idea, different reach. Claude marks text worldwide. ChatGPT marks eligible text only in the EU, rolling out from October 5, 2026, with an opt-in for API customers. Both hide the signal in word choice, not in characters.
Yes. Google DeepMind says SynthID watermarks text generated by the Gemini app and web experience. Google open-sourced SynthID Text in 2024, and Anthropic's Claude watermark is a version of the same approach.
No. Each vendor uses its own secret key, so only that vendor's detector can read its mark. Anthropic and OpenAI restrict text detection to approved organizations, and Google's public portal lists images, video and audio, not text.
None has announced a text watermark on its chatbot as of October 9, 2026. Meta and Mistral signed the EU transparency code and Meta publishes the TextSeal toolkit; DeepSeek and xAI were not on the EU signatory list we checked.
Nobody has published a fair head-to-head. OpenAI says textGrain matched or beat SynthID in its own tests. All three weaken with heavy rewording or translation, and all are weak on short or factual text.
Yes. It scans any pasted text for hidden characters in your browser and offers an optional rewrite that sends text to OpenAI. It cannot read any vendor's watermark key. No detector result is promised.