Claude will watermark AI-generated content
Anthropic says supported Claude models will add imperceptible watermarks to generated text. The watermark is part of the text itself, so it can travel with copied text and can even survive some editing.
For supported images and other files, Claude will attach signed C2PA provenance metadata instead. Anthropic says detection tools and technical details are still to come, and these marks are signals rather than perfect proof of origin.
I am not sure how this works, but I think it might work based on this paper. And this is similar to Google's SynthID that also embeds imperceptible watermarks in AI-generated content.
Update: Aug 12, 2026
Just saw an update on this as Thariq from Anthropic confirmed that the watermark detection API will be free to use.
I predict lots of big publishing platforms like Substack, Medium, and even search engines like Google and others will use the Claude API internally to decide what content to give more priority. And in the future, I guess, more AI providers will launch such an API.