Anthropic has launched a watermark verification API allowing regulators, media outlets, fact-checkers, and approved enterprises to confirm whether text originates from its Claude models. This move follows the European Union AI Act requirement, effective from August 2, 2025, that new Claude versions must embed invisible digital markers in their output. Eligible users include law enforcement bodies, independent researchers, and EU civil society groups, alongside companies needing to prove compliance with their own internal policies.
The system adapts Google’s SynthID method by adjusting word-selection randomness to create patterns that may survive some editing. Anthropic claims the watermark contains no user data and does not degrade writing quality or content integrity. Critics argue the technique might force the model to select synonyms based on a key rather than meaning, potentially lowering quality. The Artificial Lawyer also highlights transparency risks where detectable fingerprints could complicate contracts banning AI use or affect fee negotiations.
* Access expands gradually over time
* Traditional detectors like Pangram are considered less reliable
* The watermark is intended to persist through minor edits
