Yes. Anthropic announced in August 2026 that Claude's text carries an invisible statistical watermark, applied worldwide to help meet the EU AI Act. It works through word choice using a secret key, so there are no hidden characters to find or delete. Images and similar files Claude creates get signed C2PA Content Credentials instead. Detection is limited to approved organizations through a private preview.
How Claude's watermark works
When Claude writes, it often has several words that would fit equally well. Normally a random number picks between them. With watermarking, that choice is guided by a secret key and the words that came before. Readers can't tell the difference, but someone with the key can measure the pattern and estimate how likely it is that Claude was involved.
Anthropic describes its method as a variant of Google DeepMind's SynthID Text, the same family of technique Google uses for Gemini. See SynthID explained for background.
What's covered
- Text from Claude models released from 2 August 2026, watermarked at launch. Anthropic is adding it to older models over time.
- Every region, not just the EU, across the Claude apps, the API, Claude Code and Claude through cloud providers.
- Translations and summaries, since Claude picks every word.
- Code, more lightly. Code often needs exact output, so it's watermarked less, though comments may be.
- Files such as PNG, JPG and SVG images get signed C2PA Content Credentials in their metadata instead of a watermark. You can check those with our Content Credentials checker.
What it can and can't tell you
| It can | It can't |
|---|---|
| Estimate whether Claude was likely involved | Prove a human wrote something |
| Survive copying and light editing | Survive a full rewrite |
| Work on longer passages | Work reliably on very short text |
| Tell Claude writing apart from Claude heavily editing | |
| Detect other AI systems |
Like OpenAI and Google, Anthropic stresses that a detected watermark is a signal, not proof, and a missing one doesn't show the text is human.
How the big chatbots compare
| Chatbot | Text watermark | Where | Public detector |
|---|---|---|---|
| Claude (Anthropic) | Statistical, since Aug 2026 | Worldwide | No, private preview |
| Gemini (Google) | SynthID Text, since 2024 | Gemini app and web | No |
| ChatGPT (OpenAI) | Statistical, from Oct 2026 | EU first | No, restricted |
None of them use hidden characters. More on ChatGPT in does ChatGPT watermark text?
Why Anthropic is doing this
The EU AI Act's transparency rules (Article 50) apply from 2 August 2026. They require providers of generative AI to mark AI-generated content in a machine-readable way. Anthropic signed the EU's Code of Practice on transparency of AI-generated content in July 2026 and chose to apply watermarking worldwide rather than only in the EU.
What this means if you use Claude
If you use Claude to help you write, assume the text can be identified as Claude-assisted by those with access to the detector. The straightforward approach is to edit drafts into your own words and say how you used AI wherever that's expected, at school, at work or in publishing. The AI disclosure statement generator writes that line for you. Cleaning the text here fixes formatting, but it doesn't touch the watermark, and we don't claim it does.