Hello, Europeans working with (around) the watermark. What have your experience been?
retained text
Hello, I am have been using watermarked claude for the last weeks (mainly opus 5.5) mostly for research and translating my articles to english. I post on a website that has a zero tolerance for AI content writing, but using for translating is okay. I have been havig to rewrite each paragraph changing words so as to not get my articles flagged in the future. Since this is super inconvenient I did some research in how the watermark works and it raised a few concerns. My (limited) understanding is that it works a lot like a chess bot playing suboptimal moves that are still good, but not the best. In the context of chess this is often done to artificialy decrease the chess bot strenght. For example, if you go on [lichess.org](http://lichess.org) and plays against stockfish (the strongest chess computer we have) at full strenght, it will play the strongest move it finds everytime. But if you decrease the level and play against the weakest level of stockfish, it will be able to find the strongest move but will intentionally play a weaker move to simulate a lower rank player. Anyways, tying this with the watermark and claude: I had to translate an article from Portuguese to english and I also used [arena.ai](http://arena.ai) to compare other outcomes. It is clear that Claude is picking more unusual words and more varied sentence structures than open LLMs. Most of the time is almost impossible to notice, but with gemini and chatgpt aswell, sometimes it's a bit top suboptimal. I haven't researched much about it and my experience is limited, so if anyone here has some experience with he watermark i would appreciate to hear your take.
document fetched 2026-11-01 00:22 · canonical https://reddit.com/r/ClaudeAI/comments/1xk4sum/hello_europeans_working_with_around_the_watermark