Simple Unmark guide
Provider Text Watermark Status Tracker
- Published
- Last updated
- Sources checked
- Primary references: 8
Short answer
Google documents SynthID Text for generated text. Anthropic documents a SynthID-Text-style token-choice watermark in Claude output, applied globally from August 2026. OpenAI documents textGrain for eligible ChatGPT and Codex text in the EU, rolling out from October 2026, with API watermarking opt-in and off by default. None of them offers a public detector for arbitrary text today.
Provider behaviour changes, and most content about it goes stale silently. This page records only what primary sources support, with the date each claim was last verified.
Choose the privacy boundary before every rewrite.
Private is the lowest-cost path and stores no text in our application database. Confidential verifies Google Confidential Space in your browser and encrypts to a key the workload created for that one request, so our web backend receives no text — though DeepInfra still performs the rewrite. Confidential AI, with the model inside a confidential GPU, is coming soon.
Compare the three modes and their limitsHow to read this table
"Documented" means a first-party page from the provider supports the claim. "Reported" means credible coverage exists without first-party confirmation, and "unverified" means neither. We do not upgrade a claim on the strength of press coverage alone.
The verification date is when a person last re-read the linked sources, not when this page was deployed. Provider status is reviewed monthly, and within 72 hours of a major announcement.
What is true across all of them
Three patterns hold regardless of provider, and they are more useful than any single row.
- Where text watermarking is documented, it is statistical and adds no characters. Google, Anthropic and OpenAI all describe word-choice methods. Anthropic states explicitly that nothing is added and there are no hidden characters. OpenAI says the same of textGrain.
- No public detector exists for arbitrary text. Google publishes its approach and a reference implementation. Anthropic says a detection API is planned. OpenAI is opening textGrain detection first to approved researchers and expert organizations. Third-party claims to verify removal should be treated as unsupported.
- Rewriting is the documented lever, and the bar differs. Google and Anthropic say light edits survive. OpenAI reports that replacing 25% of words with synonyms cut textGrain detection from about 92% to 17%.
What would change these rows
OpenAI's 5 October 2026 post is why the ChatGPT row now says textGrain is rolling out in the EU, rather than that no text watermark is documented. A public detector from any of the three would, for the first time, let anyone measure a claim about removal — including ours, which we would then benchmark rather than describe.
Our methodology page explains how we would test that, and what we would refuse to claim without it.
Provider status
| Field | Google Gemini | Anthropic Claude | OpenAI ChatGPT |
|---|---|---|---|
| Text watermark | In production | Rolling out | Rolling out |
| Evidence | Documented by the provider | Documented by the provider | Documented by the provider |
| Mechanism | A logits processor applied during generation. Google describes it as augmenting the model's logits with a pseudorandom g-function, so the watermark lives in which tokens were chosen rather than in any added symbol. | A key-based pattern in low-stakes token choices. Anthropic describes using the key plus the preceding words to settle which word the model picks, and states the approach is a version of the SynthID-Text method published in Nature. | textGrain adds an invisible statistical signal to the model's word choices. API customers worldwide can opt in, and watermarking stays off by default. Eligible ChatGPT and Codex output in the EU receives it over the coming weeks. It is not a global default. |
| Hidden characters involved? | No. SynthID Text does not insert hidden characters; nothing is appended to the visible text. | No. Anthropic states plainly that nothing is added to the text and that there are no hidden characters. | No. textGrain is a statistical signal in word choices, not hidden characters. Invisible characters in copied ChatGPT text still usually come from the copy path. |
| Public detector | Google publishes the watermarking and detection approach and has open-sourced a reference implementation, but a public detector for arbitrary Gemini text is not offered as a general consumer service. | Anthropic offers its watermark detection API in restricted private preview to eligible organizations and enterprises with EU compliance obligations. It is not a general public checker, and Simple Unmark does not have access. | No public detector. OpenAI is opening access first to approved researchers and expert organizations. Image and audio verification stays public and separate. |
| Last verified | 29 October 2026 | 29 October 2026 | 29 October 2026 |
Google Gemini: what weakens the signal
- Google states detector confidence "can be greatly reduced when an AI-generated text is thoroughly rewritten"
- Translation into another language
- Factual responses, where there is less opportunity to augment generation
- Short excerpts, though the method is robust to cropping and mild paraphrasing
Anthropic Claude: what weakens the signal
- Anthropic states light editing probably will not remove it, while a complete rewrite in which every word is replaced will
- Text Claude only lightly edited, because few of the words are Claude's
- Factual passages and code, where there are fewer free choices to encode
- Small samples, where detection does not work well
OpenAI ChatGPT: what weakens the signal
- OpenAI's evaluation: replacing 10% of words with synonyms cut detection from about 92% to 66% on 400-token passages
- Replacing 25% of words with synonyms cut that detection to about 17%
- Shorter text, and text with little freedom in word choice such as mathematics, is harder to detect
- OpenAI says text provenance has to account for how easily text can be rewritten, translated, or edited
Sources
- Our approach to EU text provenance rules
OpenAI (5 October 2026) · First-party source
- textGrain: Entropy-Calibrated Watermarking for Language Model Text
OpenAI technical report (5 October 2026) · First-party source
- SynthID text watermarking and detection
Google AI for Developers · First-party source
- SynthID overview
Google DeepMind · First-party source
- How Claude's text watermarking works
Anthropic · First-party source
- Transparency Hub
Anthropic · First-party source
- Advancing content provenance
OpenAI · First-party source
- Scalable watermarking for identifying large language model outputs
Nature (2024)
Published by Simple Unmark. Our Methodology explains how we source claims, what we test, and what we do not claim.
Apply this to your own text.
Guests get three free cleans without an account. Choose Private or browser-verified Confidential processing before each rewrite. Deterministic character inspection is free and runs only in your browser.
Keep reading
- How AI Text Watermarks WorkHow hidden-character marks and statistical token watermarks differ, how detection works, and why removal claims should always be bounded.
- What Is SynthID Text?SynthID Text explained: how Google's text watermark is applied during generation, how detection works, what weakens it, and how it differs from image watermarking.
- Hidden Unicode vs Statistical Text WatermarksThe two mechanisms confused most often: invisible Unicode characters versus keyed token-choice watermarks. What each one is, what removes it, and what neither proves.
- AI Watermarks vs AI Detectors vs Provenance MetadataThree different things routinely confused: keyed watermarks added at generation, statistical AI detectors that guess, and C2PA metadata attached to files.
