Anthropic Details Claude’s New Watermarking System Amid User Backlash

Key Points

• Anthropic will embed invisible watermarks in all Claude‑generated text to meet the EU AI Act Transparency Code.
• The watermark uses the SynthID‑Text method and is undetectable to readers but verifiable with a secret key.
• Anthropic plans to launch a detection API that lets third parties confirm Claude’s output.
• Light editing won’t erase the watermark; only a full rewrite could remove it, potentially voiding the AI‑generated label.
• Code snippets will carry a negligible watermark, limited to optional elements like comments.
• User backlash on Reddit and X led to dozens of subscription cancellations.
• Other major AI model developers have also pledged to implement watermarks under the same Code of Practice.