Claude's watermark, in Anthropic's own words
Anthropic began watermarking Claude's text output on 2 August 2026 and published an explanation on 14 August. The single most important sentence in it: 'Nothing is added to the text and there are no hidden characters.'
The watermark is applied during generation by biasing the model's choice between statistically equivalent next tokens, using a key held by Anthropic. It is the SynthID-Text approach published by Google DeepMind.
This is why tools that strip zero-width characters have no effect on it, and why no third party can tell you whether a given piece of text carries the mark.
Frequently asked questions
Does Claude add invisible characters to its output?
No. Anthropic states directly that nothing is added to the text and there are no hidden characters. The watermark is a statistical pattern in which words were chosen.
Can Turnitin or my university detect Claude's watermark?
Not the watermark itself, which requires Anthropic's private key. Academic integrity tools use separate stylistic classifiers that are unrelated to watermarking and have well-documented false positive rates.
When did Claude watermarking start?
2 August 2026, aligned with the EU AI Act Article 50 transparency obligations that took effect for newly placed models on that date.