Anthropic has begun implementing machine-readable watermarks on Claude-generated text to comply with the EU AI Act, sparking debate among users regarding content ownership and the reliability of detection.

Key Points

• Anthropic is adopting SynthID-Text technology to mark AI-generated outputs in accordance with Article 50 of the EU AI Act.
• The marking requirement applies to free-form text longer than 200 tokens, while excluding source code and short sequences of numbers or symbols.
• Users cannot opt out of the watermarking process, which may trigger even when Claude is used only for minor editing or proofreading.
• Research from ETH Zurich indicates that watermarks can be bypassed or spoofed through paraphrasing attacks, raising questions about the long-term effectiveness of these signals.
• Major tech companies, including Google, OpenAI, Microsoft, and Meta, have signed a voluntary Code of Practice to standardize AI transparency and content provenance.

Why it Matters

The integration of machine-readable marks into AI-generated content creates a new layer of metadata that search engines and platforms could potentially use to evaluate or rank published material. As these signals become more prevalent, content creators face uncertainty regarding how automated detection might influence the perceived authenticity and search visibility of their work.