Statistical watermarking is about token choices
Text watermarking research generally nudges the probability of selected next tokens, creating a pattern that a compatible detector can test across enough text. That pattern is fundamentally different from inserting zero-width spaces or special punctuation.
Editing, translating, shortening, or mixing text can weaken statistical signals. Reliable interpretation also needs the provider’s exact method and calibrated detector.
What Anthropic makes available
Anthropic’s public transparency material says it continues to explore watermarking. Until Anthropic publishes an official, supported Claude detection mechanism, third-party claims of definitive Claude detection should be treated skeptically.
Primary sources and further reading
For technical and policy detail, consult the C2PA specification, Content Authenticity Initiative developer documentation, Anthropic transparency material, Google DeepMind’s SynthID overview, and the European Commission’s AI transparency guidance.