Anthropic adds watermarks to Claude text output
The AI company revealed a technique to mark its model's generated text without altering quality.
What to know
- Anthropic is adding watermarks to Claude text that identify AI-generated content without degrading quality, based on Google DeepMind's SynthID technique.
- The watermark is embedded in the model's word-selection process during text generation, not added as invisible characters or metadata.
- The approach addresses the challenge that text watermarking is difficult — changing even one word can alter meaning — by working within the model's probability sampling mechanism.
- Older Claude models will receive the watermarking technology gradually over coming months, with new models launching with it built-in.
The watermarking approach is technically sound and preserves output quality by embedding marks in word choices.
-
“The key thing to keep in mind is that it is in fact possible to watermark LLM-generated text without degrading output quality”
@random_walker · X
“Change a single word and you might change the meaning of an entire sentence. If the AI wants to say, for example, "The capital of France is Paris," you obviously can't watermark it by changing Paris to London.”
Video explainer, Technical explanation · YouTube
Anthropic AI companyGoogle DeepMind AI research lab
The record 13 articles and posts · last 30 days
-
Hacker News · Anon84
- summary covers to here · Aug 19, 10:47 PM · 9 pieces above arrived after
-
Hacker News · aymar_99
What people are saying 3 voices from 2 sites · verbatim
- Aug 23
-
N
This explainer shows how AI text watermarking works. commons.every.to/eb6c504586b3a7… #AI #Claude #watermarking
-
N
This explainer shows how AI text watermarking works. https:// commons.every.to/eb6c504586b3a 71e/ # AI # Claude # watermarking
- Aug 17
-
Something amazing is happening in the debate about text watermarking in Claude. This is my best attempt to make sense of the quickly evolving situation. The key thing to keep in mind is that it is in fact possible to watermark LLM-generated text without degrading output quality