Which AI models watermark their text
Two of these five assistants mark the text they generate, and neither mark is made of characters a scanner can find. This page records what each provider actually does, what the evidence is, and when it was last checked.
| Model | Text watermark | Mechanism | Last verified |
|---|---|---|---|
| ClaudeAnthropic | Marked | SynthID-Text, keyed token selection | |
| GeminiGoogle | Marked | SynthID-Text, tournament sampling | |
| ChatGPTOpenAI | Not marked | None documented for text | |
| CopilotMicrosoft | Not marked | C2PA on images, nothing for text | |
| GrokSpaceXAI | Not marked | Visible mark on media only |
What each provider does
Claude
MarkedAnthropic · verifiedMarked on models launched from August 2026, by a keyed bias in word choice rather than hidden characters.
Sources: Anthropic · Anthropic Help Center · TechCrunch
Read the Claude entry →Gemini
MarkedGoogle · verifiedMarked in the Gemini app and web experience since 2024, by the most thoroughly published scheme of the five.
Sources: Nature 634(8035), 818–823 · Google DeepMind · Google AI for Developers · Google
Read the Gemini entry →ChatGPT
Not markedOpenAI · verifiedNo text watermark shipped. The narrow no-break spaces people report are a typographic artefact.
Sources: OpenAI Help Center · Rumi · OpenAI Developer Community · European Commission
Read the ChatGPT entry →Copilot
Not markedMicrosoft · verifiedNo text watermark documented. Images always carry provenance metadata; audio and video marking is optional.
Sources: Microsoft Learn · Microsoft Learn · Microsoft Learn · C2PA
Read the Copilot entry →Grok
Not markedSpaceXAI · verifiedNothing documented for text. Images and video carry a visible label, and the maker did not sign the EU code.
Sources: SpaceXAI · SpaceXAI · European Commission · European Commission
Read the Grok entry →Invisible characters are a different thing
The eight characters below are real, and they do turn up in text that has passed through a model, an editor, a PDF or a web page. They are not watermarks. Nobody places them to identify anything, removing them defeats no provenance system, and finding one points to a text-handling step rather than to an author.
These eight are what our tools detect and clean, and nothing more. The two real text watermarks on this page work by biasing which word a model picks, which no scanner can read from text alone. We do not claim to detect or remove them.
Frequently asked questions
What is the difference between a watermark and an invisible character?
A watermark is a statistical signal placed while the model chooses its words; nothing is inserted, and reading it needs a key. An invisible character is a real code point that ended up in the text, usually by accident. Only the second can be found by scanning.
Can any tool tell me which model wrote a passage?
No public tool can, including ours. The two providers that do watermark text keep detection behind a key, and the characters that scanners find are produced by editors and web pages as readily as by models.
Why do these pages carry a verification date?
Because the answers change. Anthropic began marking text in August 2026, and Article 50 of the EU AI Act gives systems already on the market until 2 December 2026 to meet its marking obligation, so at least one entry here has a known expiry date. Each records when it was last checked against primary sources.
Related tools
How these records are checked, and what this site refuses to claim, is set out elsewhere. About GenPolish.
Working on your own draft?
GenPolish rewrites text you already have, showing every change before you accept it. It does not add or remove provenance signals, and it makes no claim about detectors.