How Claude’s text watermark works
anthropic ·
Anthropic is implementing a text watermarking system across future Claude models to comply with European Union regulations requiring transparency around AI-generated content. The technique subtly alters the random number generator used during word selection, embedding an undetectable pattern that verification tools can later identify using a cryptographic key. According to internal testing and external research, this process does not degrade output quality, alter readability, increase costs, or slow down response times. However, the watermark only indicates the probability of Claude’s involvement rather than providing absolute confirmation, and its effectiveness diminishes with short passages, highly factual statements, code, or heavily edited human writing. Importantly, the system contains no personal or organizational identifiers, ensuring that individual users and their conversations remain completely anonymous. While a detection API is currently available in private preview for regulated entities like researchers and media outlets, visual files will instead utilize standard C2PA metadata credentials to signal AI processing. Anthropic applied the feature globally at launch to simplify deployment, though they continue to evaluate region-specific implementations moving forward.
anthropic