按年被提到
20232023 · 4 次2026
AI 安全、伦理与社会影响›模型窃取、记忆与水印·arXiv:2303.11156 · cs.CL

Can AI-Generated Text be Reliably Detected?

AI 生成的文本能被可靠检测吗?
Vinu Sankar Sadasivan, Aounon Kumar, Sriram Balasubramanian, Wenxiao Wang, Soheil Feizi · 2023

递归改写攻击可击溃主流 AI 文本检测器,检测存在理论上限

开始阅读 ↗PDFalphaXiv 讨论
被推荐 · 6 人首次 2023-03-21 · 最近 2026-08-17
6不同的人
6活跃月份
8讨论串
465被引(OpenAlex)

在攻击者存在的情况下,对水印、神经网络、零样本和检索式等各类 AI 文本检测器做压力测试:提出递归改写攻击,能显著降低检测率而只轻微降低文本质量。作者还研究了对加水印 LLM 的欺骗攻击,并给出理论框架,把最优检测器的 AUROC 与人类和 AI 文本分布之间的总变差距离联系起来,说明模型越强检测越难。

HN 为什么提它

评论者以它为证据,说明 AI 文本检测器并不可靠:改写攻击是目前最好的规避手段,且检测器受总变差距离限制,模型越强检测越难。

实证研究难度 3 / 5约 40 分钟经得起时间
读之前最好知道
1大语言模型文本生成
2水印与零样本检测
3AUROC 与总变差距离
4对抗攻击

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0x_rs · 2026-08

“I'd encourage reading this paper, and literature on scaling laws in autoregressive models: Can AI-Generated Text be Reliably Detected?”

推荐读这篇论文,以及自回归模型 scaling law 的相关文献
Anthropic's 'Watermark' Text Adulteration in Claude Is a Perversion of Writing原评论 ↗
引用Veen · 2023-03

“do so with any reliability [0].”

引用它说明 AI 文本检测器并不可靠
Research shows we can only accurately identify AI writers about 50% of the time原评论 ↗
引用thewataccount · 2023-04

“Most likely not Can AI-Generated Text be Reliably Detected?”

用它回答“能否可靠检测”:很可能不能
Show HN: Pledge Your Human-Made Content原评论 ↗
引用thewataccount · 2023-07

“The current best countermeasure is likely paraphrasing attacks Can AI-Generated Text be Reliably Detected?”

引用它说明释义攻击可能是目前最好的反制手段
OpenAI Quietly Shuts Down Its AI Detection Tool原评论 ↗
引用simonw · 2024-04

“Here's a paper about how unreliable they are from last year: Can AI-Generated Text be Reliably Detected?”

贴出去年的论文,说明 AI 文本检测器有多不可靠
OpenAI's comment to the NTIA on open model weights原评论 ↗
引用mendeza · 2025-09

“This paper Can AI-Generated Text be Reliably Detected? Sadasivan et. al. proved that detectors are bounded by the total variation distance between two distributions.”

引用它证明检测器受两个分布间总变差距离的限制
Show HN: AI Code Detector – detect AI-generated code with 95% accuracy原评论 ↗
看全部 7 条 →
引用0x_rs · 2026-08

“Can AI-Generated Text be Reliably Detected? may already be a good enough answer to the paper you cited”

认为这篇已足够回应对方所引的论文:递归释义可击败水印检测
How Claude marks AI-generated content原评论 ↗
被提交为 story · 1 次