AI BriefingHow AI Watermarks Hide in Plain Text
Key point
Anthropic's text watermark is hidden through statistical biases in word choice rather than at the byte level.
Details
Anthropic plans to apply watermarks to all outputs generated by Claude. While file formats like images use the C2PA standard, for plain text, it introduces an 'imperceptible watermark' that persists even after copy and paste.
Simple ASCII text has no room to store additional information at the byte level, but watermarks can be embedded in the Word Choice process rather than the bytes themselves. This method guides the model to follow specific statistical patterns when determining the next word.
Text watermarks are broadly categorized into four layers:
- Encoding: Using zero-width characters or similar alphabets (most vulnerable)
- Formatting: Utilizing formatting such as line breaks (easily removed)
- Word Choice: Statistically preferring words with specific keys (persists through copy/paste)
- Meaning: Inserting signals into the semantic structure of the text (most robust)
Anthropic's approach likely corresponds to the Word Choice layer. This method ensures the text appears normal, but a verifier with a specific key can identify whether it is AI-generated through statistical biases.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.
Related
US Appeals Court Upholds Pentagon Blacklisting of Anthropic as Supply Chain Risk