Anthropic's Claude Watermarking Approach

Anthropic is implementing an imperceptible watermark into Claude's text outputs by statistically biasing word choices during generation, making patterns detectable over enough content without affecting quality or readability.

The top 3

  1. Claude's Watermark: Top 3 Technical Secrets: Anthropic's Claude uses a statistical watermarking method, a version of Google DeepMind's SynthID-Text, that subtly biases word choices during text generation using a secret key and preceding words, rather than adding visible or hidden characters.
  2. AI Text Watermarking: Who Leads Among Top 3 Labs?: Anthropic is notable for applying its invisible text watermark globally across all new Claude models, including API outputs, to comply with the EU AI Act. Google's SynthID also watermarks Gemini's text, while OpenAI currently focuses its watermarking efforts on images and audio with C2PA metadata.
  3. Claude's Watermark: Tracing Its Top 3 Tech Ancestors: Claude's text watermarking technique is based on Google DeepMind's SynthID-Text approach, published in 2024, which itself evolved from Scott Aaronson's proposal in 2022. The broader concept of digital watermarking dates back to Emil Hembrooke's patent for embedding codes in music in 1954.

Sources

Open the full topic