← sources

A thought on watermarks and how to circumvent them

source: reddit (reddit:ClaudeAI)published 2026-08-30 17:26ingested 2026-08-31 02:36

retained text

Heres what ive understood about the watermark technology and how its applied. First is a “green” list of words claude can pick from when constructing sentences. When this list is nudged into a body of work, statistically it shows up if you know the key.

Here is the part that, to me seems like it should be easy to break. Each sentence is constructed word by word and the statistically likelihood of the next word grows as the sentence unfolds. This means the green words, as they have been presented to us in examples from anthropic come at the end of sentences. They have to come at the end otherwise there isnt enough statistical certainty for the green payload to be useful. This could also apply to the last word before a comma, semi colon or colon. It cannot apply to the first words in a sentence at least this is my understanding. 

So really, all we need to do is create a skill that asks a model to swap a percentage of last words before periods, commas, semicolons or colons to something similar. Use this skill in another LLM entirely. Bring your work from claude over there and apply. Bam, bobs your unkle and the watermark is defeated.

Help me understand why this is wrong  
   

document fetched 2026-08-31 00:21 · canonical https://reddit.com/r/ClaudeAI/comments/1w2nmla/a_thought_on_watermarks_and_how_to_circumvent_them