How Text AI Watermarks Work and How to Bypass Them
Anthropic
Google/DeepMind
DeepSeek
The article explains the mechanism behind text watermarks in LLM outputs, required by recent EU AI regulations, and offers practical methods to remove them. The author details the probabilistic token distribution shift used in watermarking and suggests a multi-step bypass approach involving translation and rewriting with open-weight models. They also introduce a custom tool built with Streamlit and LangChain to automate this process.
The article begins by noting that EU regulations now require LLM providers to watermark AI-generated text, citing Anthropic's Claude as an example. The author explains that watermarking works by shifting the probability distribution of tokens during generation, making certain rare tokens appear more often. This is imperceptible to humans but detectable by machines that know the real distribution. To bypass watermarks, the author suggests preprocessing the text to remove obvious artifacts like em dashes and zero-width spaces, then translating the text to an intermediate language and back, and finally rewriting with an open-weight LLM like DeepSeek-v4 that does not add watermarks. The author has built a tool based on Streamlit and LangChain that automates these steps, allowing users to choose and configure each step, and it is available on GitHub.
- Abbreviations
- EU = European Union — Европейский союз
- LLM = Large Language Model — Большая языковая модель
Source: Habr — хаб ИИ —
original