Anthropic opens Claude AI text detection to regulators, media, fact-checkers, and others
decoder ·
Anthropic has released a watermark verification API that allows approved organizations to detect invisible watermarks in text generated by Claude. Access is open to regulators, law enforcement, media, fact-checkers, researchers, educational groups, EU civil society, and enterprises needing compliance verification. The feature responds to the EU AI Act, which has required invisible watermarks in new Claude models since August 2025. Anthropic's system adapts Google's SynthID method, adjusting word-selection randomness to create a statistically detectable pattern that may survive some editing, and claims it is more reliable than tools like Pangram. The company asserts the watermark contains no user data and does not affect quality or content, but critics argue that using a watermark key for synonym choice can degrade text quality. Artificial Lawyer also notes transparency risks in contexts where AI use is banned or during fee negotiations, though it deems the watermarks likely harmless overall.
decoder