Anthropic embeds invisible watermarks across Claude text and files globally

On August 11, Anthropic quietly updated a Claude Help Center article confirming that Claude models released on or after August 2, 2026 will globally embed machine-readable watermarks in generated text and attach C2PA-based digitally signed provenance metadata to supported files (PNG, JPG, SVG). The move is Anthropic's concrete action to comply with EU AI Act Article 50 transparency obligations.

Article 50(2) of the EU AI Act took effect August 2, requiring providers of AI systems that generate synthetic content to ensure outputs are marked in a machine-readable format and detectable as artificially generated or manipulated. Non-compliance can lead to fines of up to 15 million euros or 3% of global annual turnover. Anthropic has signed the EU Code of Practice on Transparency of AI-Generated Content; approximately 190 organizations including Google, Meta, Microsoft, OpenAI, Black Forest Labs, and Synthesia have also signed.

At the text level, Anthropic embeds a statistical bias pattern directly in model output that survives copy-paste and remains detectable; at the file level, C2PA metadata records Claude's involvement and enables tamper detection, though screenshots or re-saving can strip the metadata. The watermarking applies across Claude Code, Claude Cowork, Claude Tag, and Claude models accessed via AWS, Google Cloud, and Microsoft Foundry. Anthropic stated it will publish detection methods and APIs in the future.

Anthropic emphasized that a detected watermark only indicates Claude may have been involved in content generation or processing, not that it constitutes intellectual authorship—it is a transparency signal. Debate has emerged across academia and industry about watermark misinterpretation in education, journalism, and legal contexts, and Anthropic's global-by-default approach means non-EU users are equally affected.