Live Tracker
AI Watermark Status Tracker
Which AI models currently watermark their text output? Here is the live status across all major providers as of August 2026.
Quick answer
Google Gemini is the only major AI model currently watermarking text output, and has been doing so since 2024. Anthropic announced Claude text watermarking in August 2026 but it is not yet applied to current models. OpenAI, Meta, Microsoft, and Mistral have no text watermarking. No public detection API exists for any provider except Google.
1
Model currently watermarked
Gemini (all models)
1
Announced but not yet applied
Claude (models after Aug 2)
1
Announced, no date yet
ChatGPT/OpenAI
Full status by provider
| Provider | Status | Method | Detection | Removal |
|---|---|---|---|---|
| Claude (Anthropic) | Announced — not yet applied | SynthID-Text (statistical bias in word choices) | API announced, not shipped | Complete rewrite by non-watermarking model |
| Gemini (Google) | Live | SynthID-Text (statistical pattern in token distribution) | SynthID detection available | Structural rewrite by different model |
| ChatGPT (OpenAI) | Planned — not yet shipped | Statistical watermarking (details undisclosed) | No detection method available | N/A |
| Copilot (Microsoft) | No watermark | N/A | N/A | N/A |
| Llama (Meta) | No watermark | N/A | N/A | N/A |
| Mistral | No watermark | N/A | N/A | N/A |
Claude watermark: detailed status
What Anthropic announced
On August 14, 2026, Anthropic announced text watermarking for Claude, citing EU AI Act Article 50 compliance. The method is SynthID-Text style — a keyed statistical bias in word choices, no hidden characters, no steganography. The watermark is applied globally, not just in the EU.
Which Claude models are affected?
The watermark applies to models launched on or after August 2, 2026. As of August 2026, no current Claude model meets this cutoff. Opus 5, Sonnet 5, Fable 5, and Opus 4.8 all launched before August 2 and are not watermarked. Retrofits to older models are planned but have not been applied.
Can anyone detect it?
No. Anthropic announced a detection API but it has not shipped. No third-party tool can read the watermark. Once the API launches, any text generated by a watermarked Claude model becomes retroactively checkable.
What about translations and grammar fixes?
Anthropic states that translations carry the watermark “because in this case every word is chosen by Claude.” Light proofreading usually leaves too little signal, but substantive grammar correction or sentence restructuring involves enough word selection to embed the fingerprint.
Does the watermark make content uncopyrightable?
The watermark does not change copyright status. However, it makes AI provenance provable for the first time. Under current US Copyright Office guidance, provably AI-generated content may not be copyrightable. The watermark turns “AI-generated” from an accusation into a checkable fact.
Read more: Does the Claude watermark make your content uncopyrightable?
How text watermarks work
SynthID-Text (Claude, Gemini)
At generation time, the model biases its token choices using a secret key. This creates a statistical pattern across the entire document that is invisible to humans but detectable with the key. The watermark survives copy-paste, light editing, and retyping. The only acknowledged removal is a complete rewrite by a different model.
C2PA metadata (images)
Separate from text watermarks. Claude attaches “Made with Claude” C2PA metadata to generated images. This is file-level provenance information, not a text watermark. It can be stripped by re-saving the image without metadata. ClaudeRewrite handles this automatically.
Frequently asked questions
Is Claude text watermarked right now?
Not yet. Anthropic announced text watermarking in August 2026, but it applies to models launched on or after August 2, 2026. Current models (Opus 5, Sonnet 5, Fable 5, Opus 4.8) launched before that date and are not watermarked. Retrofits to older models are planned but not yet applied.
Can anyone detect AI watermarks today?
For Claude, no — Anthropic announced a detection API but it has not shipped. For Google Gemini, SynthID detection is available. For ChatGPT/OpenAI, no detection method exists yet. No independent third-party tool can reliably detect text watermarks from any provider.
Which AI models watermark text in 2026?
Google Gemini has live SynthID-Text watermarking since 2024. Anthropic Claude announced text watermarking in August 2026 but it is not yet applied to current models. OpenAI ChatGPT has no text watermark yet. Meta Llama, Microsoft Copilot, and Mistral have no text watermarking.
Does the watermark survive copy-paste?
Yes. SynthID-Text watermarks survive copy-paste, light editing, and retyping. The only acknowledged removal method is a complete rewrite by a different AI model.
Can I remove a watermark from AI-generated text?
Yes. A complete rewrite by a different AI model breaks the statistical fingerprint. ClaudeRewrite does this automatically — it uses Gemini 2.5 Flash to rewrite Claude-generated text, preserving meaning while resetting the watermark.
Related analysis
I wrote it myself — Claude only translated it. Why is it marked as AI?
The watermark marks translations because Claude chose every word. Here's why non-native speakers are disproportionately affected.
Does the Claude watermark make your content uncopyrightable?
The watermark doesn't change copyright status — but it makes AI provenance provable for the first time.
Remove the watermark before detection ships
ClaudeRewrite rewrites your text to break the statistical fingerprint. Free — 5 rewrites/month, no card.
Try it free →