Simple Unmark guide
How AI Text Watermarks Work
- Published
- Last updated
- Sources checked
- Primary references: 5
Short answer
There are two unrelated mechanisms. Hidden-character marks are literal invisible code points that can be found and deleted exactly. Statistical watermarks bias which words a model chooses, so the signal is the wording itself; it survives copying, weakens under thorough rewriting, and can only be reduced rather than deleted.
AI text watermarking is not one technology, and most confusion online comes from treating it as one. The useful split is between deterministic marks hidden in Unicode and probabilistic signals distributed across word choices.
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 limitsProbabilistic token watermarks
At each step a language model has several acceptable next words. A statistical watermark uses a secret key to nudge that choice, consistently, in a way a reader cannot perceive. Google describes SynthID Text as a logits processor that augments the model's logits with a pseudorandom g-function; Anthropic describes using the key plus the preceding words to settle which word gets picked. OpenAI's textGrain, announced on 5 October 2026, is the same kind of signal: an invisible statistical bias in word choices for eligible ChatGPT and Codex text in the EU.
Detection then measures whether a passage's choices align with the key more often than chance allows. This is why the signal is cumulative: more text means more evidence. It is also why nothing is hidden in the characters — Anthropic states plainly that nothing is added to the text and there are no hidden characters.
What weakens a statistical signal
The providers document this themselves. Rewriting is the lever that matters, and the more of the original wording that survives, the more of the signal survives with it. They do not set the same bar for a light edit.
- Thorough rewriting: Google says detector confidence can be greatly reduced; Anthropic says a complete rewrite in which every word is replaced will remove the mark, while light editing probably will not.
- OpenAI's textGrain: in a 400-token evaluation, replacing 10% of words with synonyms cut detection from about 92% to 66%, and replacing 25% cut it to 17%.
- Translation into another language.
- Density of free choices: factual passages, code and quotations carry less watermark, because the model had fewer acceptable alternatives to encode into.
- Length: short samples carry little evidence. Anthropic notes detection does not work well on small samples, and OpenAI reports the same for shorter or more constrained text.
Light editing probably won't remove the watermark completely; a complete rewrite where every word is replaced will.
Why guarantees are misleading
Detection is statistical, keys and methods are not static, and for most providers there is no public detector to test against. Anyone promising that text is now permanently undetectable is describing a measurement they cannot make.
The defensible claim is narrower. Rewriting is what the providers say weakens the signal, and how far it weakened cannot be checked without their detector. It also helps to separate the mechanical question from the ethical one: a watermark indicates a model was involved, not who holds authorship, and Anthropic says explicitly that its mark cannot distinguish text Claude wrote from text Claude heavily edited.
Sources
- SynthID text watermarking and detection
Google AI for Developers · First-party source
- How Claude's text watermarking works
Anthropic · First-party source
- Our approach to EU text provenance rules
OpenAI (5 October 2026) · First-party source
- Scalable watermarking for identifying large language model outputs
Nature (2024)
- UTR #36: Unicode Security Considerations
Unicode Consortium · First-party source
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
- 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.
- Does Gemini Watermark Text?Yes. Gemini text carries SynthID Text, a word-choice watermark with no hidden characters. What Google documents, what weakens it, and why you cannot check it.
