Which AI models watermark their text?
A sourced, dated record of which AI providers watermark the text they generate — since when, where, how the mark works and who can detect it. Every row links to the provider’s own announcement.
Last updated · 11 providers tracked
| Provider | Watermarks text? | Since | Where | How | Who can detect it |
|---|---|---|---|---|---|
| OpenAIChatGPT, Codex, API | Yes — EU, rolling out | Announced Oct 5, 2026 | Eligible ChatGPT and Codex text for EU users; API opt-in worldwide (off by default) | Statistical (textGrain): nudges word choice into a detectable pattern | OpenAI’s detector — approved researchers and expert organisations only; academic-integrity checks are excluded |
| GoogleGemini app and web | Yes | May 14, 2024 | Text generated in the Gemini app and on the web (SynthID also covers images, audio and video) | Statistical (SynthID-Text): biases token sampling with a secret key | No public text detector — the SynthID Detector, open to everyone since Oct 7, 2026, checks images, video and audio only |
| AnthropicClaude, Claude Platform (API), Claude Code, Cowork, Tag — and Claude on AWS, Google Cloud and Microsoft Foundry | Yes — worldwide | Claude models launched from Aug 2, 2026 | Global. Text watermarks on Fable 5 and 5.1, Mythos 5.1, Opus 4.8, 5 and 5.5, Sonnet 5 and 5.5, and Haiku 5.5; Anthropic is adding them to older models such as Opus 4.7 and Sonnet 4.6, with no date published. Supported files get C2PA credentials | Statistical: low-stakes word choices made with a secret key and the preceding words; no hidden characters | Detection API in private preview for eligible organisations (regulators, researchers, media, educators…) |
| MicrosoftWindows AI text APIs (Phi Silica) | Yes — hidden characters | October 2026 | Text from the Windows AI generate, rewrite and summarise APIs, inserted zero or more times per response | Invisible characters: the sequence U+200B U+034F U+2060 U+200B (zero-width space, combining grapheme joiner, word joiner, zero-width space) | Anyone — they are ordinary Unicode characters (our checker flags the exact sequence) |
| MicrosoftFoundry (Azure OpenAI) · Copilot | Opt-in (Foundry) · not announced (Copilot) | Foundry docs Oct 17, 2026 | Planned for selected Azure OpenAI gpt-5.1–gpt-5.6 deployments, switched on per deployment. Microsoft 365 Copilot watermarks images, audio and video, not text | Invisible watermark from the underlying model (details not published) | Microsoft’s provenance checker covers images, audio and video, not text |
| MetaMeta AI · Llama | No text watermark announced | — | Signed the EU Code of Practice (July 28, 2026); its announcement covers images and C2PA labels | — | — |
| MistralVibe (formerly Le Chat) | No text watermark announced | — | Code of Practice signatory; no public statement on text marking | — | — |
| DeepSeekDeepSeek app · API | Visible label only | Under China’s labelling rules (Sep 1, 2025) | Visible notices in DeepSeek’s own app: on the welcome page, at the end of generated text and at the bottom of the interface. The disclosure says nothing about labelling API output. Not an EU Code signatory | Visible notice, no machine-readable text mark announced | — |
| xAIGrok API · Grok app | Opt-in (API) · not announced (Grok app) | Docs updated Nov 9, 2026 | Text from grok-4.7, 4.6, 4.5, 4.3 and the grok-4.20 models through the API, when a team admin turns it on for a model; off by default. Nothing published for text in the Grok app or Grok on X | Statistical (SynthID-Text): shapes which tokens the model picks; no characters added. An x-watermark: true response header flags marked output | xAI’s detection API, by application; it answers watermarked, not watermarked or uncertain |
| PerplexityPerplexity | No text watermark announced | — | No official statement; not an EU Code signatory | — | — |
| AmazonNova · Bedrock | No text watermark announced | — | Nova Canvas and Reel watermark images and video; Nova text models’ service cards don’t mention one. Not an EU Code signatory | — | — |
Provider details
OpenAI ChatGPT, Codex, API
OpenAI says textGrain adds no hidden characters or unusual punctuation. Its tests at a 1% false-positive rate: ~80% of 200-token and ~95% of 400-token passages detected (lower for maths); on 400-token passages, ~66% after swapping 10% of words for synonyms and 17% after 25%. Detection also varies by language, from 69.0% (Spanish) to 42.2% (Romanian) before OpenAI strengthened the weaker ones.
Google Gemini app and web
SynthID-Text was open-sourced in October 2024 (Hugging Face Transformers). Google notes detection confidence drops when text is thoroughly rewritten or translated, and the mark is weaker on factual answers.
Anthropic Claude, Claude Platform (API), Claude Code, Cowork, Tag — and Claude on AWS, Google Cloud and Microsoft Foundry
Anthropic says light editing probably won’t remove the watermark completely, but a complete rewrite that replaces every word will. A detection indicates likely Claude involvement, not authorship.
Microsoft Windows AI text APIs (Phi Silica)
Unlike the statistical marks above, this one is made of characters, so a character scan finds it and stripping the characters removes it completely. Microsoft reminds developers to consider their own marking obligations before removing it.
Microsoft Foundry (Azure OpenAI) · Copilot
Claude models on Foundry carry Anthropic’s own watermark: Anthropic says supported Claude models accessed through AWS, Google Cloud or Microsoft Foundry are watermarked.
Meta Meta AI · Llama
Meta’s FAIR lab has published research on text watermarking, but nothing says it is deployed in Meta AI or Llama.
Mistral Vibe (formerly Le Chat)
As a signatory, Mistral has committed to the Code’s marking measures, but has not said publicly how or when.
DeepSeek DeepSeek app · API
DeepSeek’s disclosure mentions no watermark, metadata or hidden identifier, so text copied out of the app carries no announced mark.
xAI Grok API · Grok app
xAI’s documentation, which now uses the company name SpaceXSI, says text watermarking is off by default and that some models can be locked on for a team. News coverage in early November reported it would be on by default from November 7; the documentation, last updated November 9, says otherwise, so this entry follows it. The watermark covers ordinary replies, structured outputs and tool-call arguments. The Grok app’s FAQ mentions a watermark on generated images and videos only, and says xAI’s acceptable use policy prohibits removing watermarks or other provenance signals.
Perplexity Perplexity
We found no official Perplexity statement either way.
Amazon Nova · Bedrock
Claude models used through AWS do carry Anthropic’s own watermark, according to Anthropic.
Why now: the EU AI Act
Article 50(2) of the EU AI Act requires providers of AI systems that generate text, images, audio or video to mark the output so it can be detected as AI-generated. It applies from August 2, 2026; systems already on the market before that date have until December 2, 2026. AI used only for assistive or standard editing, or that doesn’t substantially change what you give it, is exempt.
The Commission’s Code of Practice on transparency of AI-generated content (final version June 10, 2026) spells out what that means for text: because plain text can’t carry metadata, an imperceptible watermark is the one marking layer; text shorter than about 200 tokens counts as “very short” and is exempt; and detectors may be restricted to verified expert users — which is why none of the detectors above is public. Around 190 organisations have signed it, including OpenAI, Anthropic, Google, Meta, Microsoft and Mistral; xAI, DeepSeek, Perplexity and Amazon are not on the list.
Separately, Article 50(4) puts a duty on whoever publishes: AI-generated text published to inform the public on matters of public interest must be disclosed as such, unless it has gone through human review or editorial control and someone holds editorial responsibility for it.
“Providers of AI systems, including general-purpose AI systems, generating synthetic audio, image, video or text content, shall ensure that the outputs of the AI system are marked in a machine-readable format and detectable as artificially generated or manipulated.”EU AI Act, Article 50(2)
What it means in practice
- You can’t see these watermarks. They live in word choice, so neither reading nor a character scan reveals them — and no provider offers a public text detector.
- Light edits rarely remove them. OpenAI’s own figures show detection still around 66% after a tenth of the words change; Anthropic says light editing probably won’t remove its mark.
- A thorough rewrite does. Changing most of the wording breaks the pattern — that is what Unmark’s rewrite does, best-effort, while keeping your meaning.
- Hidden characters are a separate issue. They come from formatting and copy-paste, not from these watermarks. Check your text for them.
How this tracker is maintained
Entries are checked against each provider’s own announcement or documentation; news coverage is used only to find those primary sources. When a provider has not said anything public, the entry says so rather than guessing. Spotted a change or an error? The changelog below records every update.
Changelog
To hear about the next change without coming back, follow the changelog as a feed in any feed reader.
Cite this page
Unmark. “AI Text Watermark Tracker.” Updated November 10, 2026. https://unmark.dev/ai-watermark-tracker