Anthropic Embeds Watermark in Claude Text Starting August 2; Open-Source Removal Projects Emerge

According to Anthropic's support article, starting August 2, 2026, Claude models began embedding an imperceptible watermark in all generated text worldwide. The watermark applies to Claude's chatbot, API, Claude Code, and cloud partners including AWS and Google Cloud. Anthropic stated the mark is woven directly into text, survives copy-paste, and will not alter readability or meaning.

The technique remains undisclosed, though researchers infer it relies on statistical bias in word choice, similar to Google's SynthID Text. Within days, open-source projects emerged to remove the watermark: mikiane/claude-watermark-cleaner strips invisible Unicode and rewrites text with alternative models, while guillaumemeyer/watermarks-remover (4.6k GitHub stars) targets Claude marks, C2PA, and other watermarking signals across multiple file formats.