Claude Text Watermark

Anthropic plans to include watermarks in text generated by future Claude models to comply with the EU AI Act. This technology is used to identify the likelihood that the text was written by Claude.

Watermarking utilizes 'low-stakes choices' that occur during the text generation process. When the model selects the next word, it combines a specific key with previous words instead of using complete randomness, leaving a pattern. This pattern is invisible to general readers, but someone with the key can verify the text's consistency to probabilistically determine if it was generated by Claude.

Key features are as follows:

• Quality Preservation: Does not affect the quality, creativity, or readability of the text.
• Invisibility: No hidden characters or additional tokens; indistinguishable to the human eye.
• Privacy: Contains no identifying information that can track specific individuals, organizations, or chat content.
• Cost Efficiency: No additional tokens or costs are incurred.

This approach is similar to Google DeepMind's SynthID-Text technology, and internal testing results showed no statistical difference in text quality based on the presence or absence of the watermark.