按年被提到
20232023 · 11 次2026
AI 安全、伦理与社会影响›模型窃取、记忆与水印·arXiv:2301.10226 · cs.LG

A Watermark for Large Language Models

大语言模型的水印
John Kirchenbauer, Jonas Geiping, Yuxin Wen, Jonathan Katz, Ian Miers, Tom Goldstein · 2023

用绿色词表软性偏置采样,给 LLM 文本加水印

开始阅读 ↗PDFalphaXiv 讨论
被推荐 · 19 人首次 2023-01-25 · 最近 2026-08-17
19不同的人
10活跃月份
18讨论串
312被引(OpenAlex)

提出面向专有语言模型的水印框架:在生成每个词之前随机选出一组绿色 token,采样时软性地提高使用绿色 token 的概率,从而在文本中嵌入人眼不可见、但算法可检测的信号。水印对文本质量影响可忽略,检测无需访问模型 API 或参数,并配有带可解释 p 值的统计检验和信息论分析框架。作者在 OPT 系列的数十亿参数模型上做了测试,并讨论了鲁棒性与安全性。

HN 为什么提它

评论者常把它称为红绿词表方案,是讨论如何识别 AI 生成文本时的标准出处,同时也有人指出这种水印很容易被绕过。

方法难度 3 / 5约 50 分钟经得起时间
读之前最好知道
1语言模型的 token 采样与 logits
2哈希与伪随机数生成
3假设检验与 p 值
4信息论基础

HN 书评

推荐在前,同类按时间 · 每篇最多保留 20 条,统计用全部提及
推荐kimburgess · 2023-02

“On the technical front there's a good paper here: A Watermark for Large Language Models, and a nice very approachable video explaining it here: [LINK].”

推荐这篇讲 LLM 水印技术的好论文,并附带通俗易懂的视频讲解
ShareGPT: Share your ChatGPT conversations with one click原评论 ↗
推荐Ethcad · 2023-07

“some interesting work using steganography to mark generated text”

推荐它,用类似隐写术的方法给生成文本做标记
A.I. Companies Agree to Safeguards After Pressure from the White House原评论 ↗
推荐thewataccount · 2023-07

“Here's a decent paper on it.”

推荐一篇不错的论文,讲私有水印:没有密钥不可检测,且能抵御简单编辑
OpenAI Quietly Shuts Down Its AI Detection Tool原评论 ↗
推荐HNx1 · 2023-08

“A really good paper about this idea was just awarded outstanding paper at ICML - A Watermark for Large Language Models.”

推荐这篇刚获 ICML 杰出论文的水印论文,通过调整 logit 给 LLM 输出加水印
Show HN: IdentityLM, cryptographic proof of identity via language model output原评论 ↗
批评Der_Einzige · 2024-08

“is not an example of this because of how trivial it is to defeat.”

批评这种水印方案太容易被破解,算不上有效的例子
Update on Llama adoption原评论 ↗
引用stcg · 2023-03

“I think watermarking is a better solution. There is a paper on it:”

认为水印是比校验和更好的方案,并指向这篇论文
Research shows we can only accurately identify AI writers about 50% of the time原评论 ↗
看全部 13 条 →
引用Zambyte · 2024-07

“I wonder if they have a watermark built in that allows them to detect text generated by Llama easily”

好奇他们是否内置了水印,能轻易检测出 Llama 生成的文本
Meta Launches AI Studio in US原评论 ↗
引用tomsander1998 · 2024-11

“text watermarking works pretty well! A Watermark for Large Language Models”

拿它说明文本水印效果不错,通过偏置 LLM 采样来嵌入水印
Watermark Anything原评论 ↗
引用Der_Einzige · 2026-08

“closer to the original "red token, green token" paper: A Watermark for Large Language Models”

指向更接近原版的“红绿 token”水印论文
How Claude marks AI-generated content原评论 ↗
引用m-chrzan · 2026-08

“explaining a paper by John Kirchenbauer, Jonas Geiping et al. that described a method for watermarking LLM output like this.”

解释这篇论文描述了给 LLM 输出加水印的方法,Claude 的水印可能就用了它
How Claude marks AI-generated content原评论 ↗
引用Eisenstein · 2026-08

“A token is hashed and used to seed a random number generator, which produces the red list for the token after it.”

描述其机制:对 token 哈希来为随机数生成器播种,得到下一个 token 的红名单
How Claude marks AI-generated content原评论 ↗
引用krackers · 2026-08

“the easiest for people to understand is the red/green scheme by Kirchenbauer et al., which does technically distort the logits (but I'd argue only in cases where you wouldn't notice it anyway).”

称红绿方案最容易理解,虽然确实会扭曲 logit,但多数情况下不易察觉
Anthropic's 'Watermark' Text Adulteration in Claude Is a Perversion of Writing原评论 ↗
引用Majromax · 2026-08

“A reasonable guess about the algorithm is 'A Watermark for Large Language Models'.”

猜测所用算法就是它,并解释红绿名单水印的原理及可检测性取舍
Anthropic's 'Watermark' Text Adulteration in Claude Is a Perversion of Writing原评论 ↗
被提交为 story · 7 次