🤔 Nothing to see here

Anthropic has explained how its invisible watermark for Claude-generated text actually works.

Spoiler: there are no hidden characters. The model simply makes slightly different choices between equally plausible words, creating a pattern that can later be used to estimate whether a text was generated by Claude.

Here’s how it works and where it doesn’t:

↖️ https://durovscode.com/claude-text-watermark