Event ·Multi-source

文本水印原理详解：KGW方法、破解与工程难点

本文以Claude的水印为例，详解了KGW方法的原理：生成时用密钥和已生成token计算哈希，将词表分为绿组和红组，并提高绿组词被选中的概率。检测时用同一密钥还原分组，若绿组词占比显著超50%则判定带水印。改写可破解水印，但统计模型生成的哈希（如SIR、Adaptive Watermark）能增强鲁棒性。Google在2万条文本的人类反馈实验中称质量下降难以感知，但短文本或高确定性输出（如1+1=2）会失效。工程难点包括流式输出、密钥轮换和代码类输出的限制。

Current read

想搞懂AI文本水印到底怎么运作、能不能被破解？这篇把KGW方法、改写攻击和工程难点讲得明明白白，还对比了Anthropic和Google的做法。

11 sources73° AI heatLast updated 2026/8/11 13:40:18

Evidence

11 sources

Primary source宝玉

2026/8/11 13:40:18View source

Related source 1IT之家

2026/8/11 04:52:16View source

Related source 2Paul Graham

2026/8/10 21:16:30View source

Related source 3Geek

2026/8/11 06:12:13View source

Related source 4Yangyi

2026/8/12 03:19:22View source

Related source 5techcrunch

2026/8/9 19:20:32View source

Related source 6Simon Willison’s Weblog

2026/8/9 23:31:39View source

Related source 7量子位

2026/8/10 05:37:51View source

Related source 8AI Will

2026/8/10 06:49:32View source

Related source 9The Rundown AI

2026/8/10 14:30:21View source

Related source 10arXiv cs.AI

2026/8/10 17:24:50View source

Conflict check

Dedup data only shows that these sources discuss the same event, not that they agree. There is no structured supporting or refuting evidence yet, so no conflict is inferred automatically.

Open the story