Anthropic reveals Claude's watermark uses statistical word patterns, no hidden characters, but harder to remove than expected
Anthropic explains the watermarking mechanism in Claude-generated text, confirming it uses statistical techniques via the SynthID-Text standard with no hidden special characters, persisting even after being copied.
Following Anthropic's announcement of its policy to embed watermarks in text and files generated by AI to comply with the EU AI Act, immediate criticism arose as the measure applies globally, not just in Europe. Meanwhile, many questioned whether the watermark could be easily removed.
Recently, Anthropic provided further explanation via its Help Center, stating that the company uses Statistical Watermarking by adjusting token selection probabilities during text generation. This creates a detectable statistical pattern while keeping the text readable like normal content, confirming that no hidden special characters or Unicode are inserted.
The standard used is SynthID-Text, developed by Google DeepMind. This mechanism ensures the watermark "sticks to the text" even if users copy-paste it elsewhere. However, detection capabilities may decrease or disappear if the text undergoes heavy editing or paraphrasing.
The rationale behind this is compliance with Article 50 of the EU AI Act, which mandates transparency for AI-generated content. Anthropic will begin applying this to models released from August 2, 2026, onward, enforcing it globally. This was jointly reported by international outlets like Engadget and Forbes, confirming Anthropic chose a statistical approach over embedding metadata or hidden characters.
Thai Claude users will be directly affected since this measure is enforced globally, even though Thailand is not under EU jurisdiction. AI text reused elsewhere can be traced back to its AI origin, impacting transparency, academic work, and commercial content usage.