How Claude's 'Watermark' Distorts Writing

The text watermark Anthropic plans to introduce to Claude models does not use invisible character insertion but instead leverages steganography.

Core Mechanism: Probabilistic Token Selection

• When generating the next token, the model classifies words in real time into 'green' and 'red' lists.
• To embed the watermark, the model intentionally increases the probability of selecting words from the green list over the red list.
• As these subtle probabilistic biases accumulate across the text, a detector with the secret key can statistically identify them.

Concerns Over Quality Degradation and Distortion

• Anthropic claims this method does not alter the meaning, quality, or readability of the text.
• However, the author argues that the probabilistic word selection process interferes with the model's natural language choices, inevitably distorting the meaning and quality of the text.