Anthropic is adding invisible watermarks to all of its Claude text

Anthropic will be inserting hidden watermarks into all texts produced by Claude, as well as embedding C2PA metadata into images (e.g., .png, .jpg, .svg). This will affect all versions of Claude (including Claude Code and their API) starting in August 2026 (immediately for new models, later for legacy models). The text won’t be affected (and may even survive “some editing”), but if you do a lot of editing or translation, the watermark will be removed.

This will fundamentally shift the landscape of AI content creation, since we’ll finally have a way to automatically check whether something was written by an AI or not. Anthropic has committed to the EU AI Act Code of Practice on Transparency, which means we’re starting to see regulations drive the move from theoretical to practical uses of invisible watermarking. They say they’re working on creating third party verification tools, but don’t have a timeline yet.