How the Claude Text Watermark Works
This page summarizes Anthropic's published explanation of Claude's SynthID-Text watermark. It separates current documentation from details that Anthropic has not yet released.
The Mechanism
Claude generates text by selecting one next token at a time. When several choices are similarly suitable, Anthropic says its version of SynthID-Text uses a secret key plus preceding words to choose the source of randomness. Across a long enough passage, those choices form a statistical pattern that a detector with the key can test. The system does not add hidden characters or extra tokens.
What's Publicly Known
- Models launched in the EU on or after August 2, 2026 support marking at launch; older models are in a transition rollout.
- The method is a version of SynthID-Text, published by Google DeepMind in 2024.
- Marks apply across supported Claude products, API access, cloud partners, and regions.
- The key does not encode a user, organization, or chat identity.
- Code, factual passages, and light proofreading may carry a weaker signal because fewer word choices are available.
What's Not Public
Anthropic has not published its production key, confidence thresholds, detector implementation, or public detection API. It says a detection API and further technical documentation are forthcoming. This site therefore cannot perform official Claude watermark detection.
Limitations
Detection is probabilistic. Short samples offer less evidence, and heavily edited, paraphrased, translated, or mixed text may lose a detectable signal. A positive result only suggests Claude may have processed the text; a negative result does not prove that no AI was involved.
Can I detect the watermark myself?
Not with this site's checker. Anthropic says its official detection API is still in development.
Is the mechanism public?
The SynthID-Text design and research paper are public. Anthropic has not published its production key, detector, or operating thresholds.
Sources
- Anthropic: How Claude's text watermark works, August 14, 2026
- Claude Help Center: How Claude marks AI-generated content
- Nature: Scalable watermarking for identifying large language model outputs, 2024