Anthropic Details Claude Text Watermarking: SynthID Approach and Invisible Detection

APPSO · wechat · 2026-08-15

Anthropic released a detailed blog post explaining the mechanics, impact, and implementation of text watermarking for Claude.

Core Mechanism

Adopts Google DeepMind's SynthID-Text method. The watermark is applied by subtly altering the source of randomness used when the model selects the next token (shifting from arbitrary random numbers to a pseudo-random sequence derived from a key and context).
It claims no change to text content, no extra tokens, no cost increase, and is invisible to humans without the specific key.

Impact on Quality and Use Cases

States that there is no statistically significant difference in quality, creativity, or readability, citing internal tests and user ratings from Gemini.
Limitations: Detection confidence is lower for short texts. Factual texts (e.g., book titles, formulas) offer no choice for word selection, so watermarks cannot be applied there. Code contains sparse watermarks, mostly in comments.

Other Key Details

Editing & Translation: When polishing human text, watermarks exist only in modified words; full translations will contain watermarks.
Privacy & Compliance: Watermarks do not contain user identity info. The move is to comply with the EU AI Act and will be rolled out globally.
Detection & Evasion: A detection API will be released. Heavy rewriting can remove watermarks, while light edits may not.
File Watermarking: Images use C2PA standard metadata, distinct from the text watermarking mechanism.

More from Models