Anthropic starts watermarking everything Claude writes and draws
Anthropic said Claude models released after 2nd Aug’26 will embed an invisible watermark in the text they generate, and will tag generated images with C2PA provenance metadata that records which model made the file and when. This is to comply with the EU AI Act's voluntary Code of Practice, which Anthropic and OpenAI have both signed. OpenAI, which sits on the C2PA steering committee, is expected to follow with watermarking of its own. The mark is designed to survive copy-paste and some editing, and detection tooling is coming. Although Anthropic admits it won't reliably catch short or heavily-edited snippets.
What it actually means
For two years you could pass AI-written copy or an AI image off as hand-made and nobody could prove otherwise. Now the model-maker is stamping provenance into the output itself, and the biggest one is about to do the same. So, if any part of your workflow depends on AI content reading as human, assume it's now detectable, and that a client, platform or regulator can check. The move isn't to panic, but to decide your disclosure line on purpose. Where leaning on AI is fine, say so. Where it truly can't show, keep a human genuinely on the work.
You can't assume AI content passes as human anymore. Model’s now sign their own work.
Source: SiliconANGLE · Wednesday 12 August 2026