Anthropic has shared new details about the watermarking system coming to Claude, clarifying how the feature is designed to mark AI-generated text while keeping the reading experience unchanged.
The company says the method creates a hidden pattern in word choices during moments when several valid options exist. To readers, the output should look identical to standard text, but authorized detection tools can identify the embedded signal.
Anthropic says the approach is based on SynthID-Text, the watermarking technique introduced by Google DeepMind, and that it plans to offer a detection API for developers and organizations. The company also stressed that this is different from AI-detection tools that try to spot stylistic clues in writing.
According to Anthropic, light editing is unlikely to remove the watermark fully, while a complete rewrite would erase it. For text that Claude only helps polish, the result depends on how much of the original wording remains. In code, the watermark is expected to be much weaker, since functional code leaves fewer free choices in wording. Comments and other flexible sections may still carry the signal.
Anthropic also noted that Claude will not be the only system moving in this direction, as other major model developers have signed the same code of practice and are expected to adopt similar transparency measures. The shift points toward a future where AI-generated content becomes easier to identify without changing how it is created or read.
