按年被提到
20262026 · 21 次2026
大模型与语言模型›大模型文本、写作与版权·arXiv:2601.02671 · cs.CL

Extracting books from production language models

从生产级语言模型中提取书籍
Ahmed Ahmed, A. Feder Cooper, Sanmi Koyejo, Percy Liang · 2026

用 BoN 越狱加迭代续写,从四个商用 LLM 中提取整本书

开始阅读 ↗PDFalphaXiv 讨论
被提及 · 14 人首次 2026-01-08 · 最近 2026-09-28
14不同的人
7活跃月份
19讨论串
8被引(OpenAlex)

围绕 LLM 与版权的许多法律问题都归结为记忆:训练数据是否被编码进权重、能否从输出中被提取。作者用两阶段流程检验生产级 LLM 能否被提取受版权保护的书籍:先做探测(有时用 Best-of-N 越狱),再用迭代续写提示尝试提取整本书,并以基于最长公共子串块近似的 nv-recall 评分。在 Claude 3.7 Sonnet、GPT-4.1、Gemini 2.5 Pro 和 Grok 3 上,Gemini 和 Grok 无需越狱即可提取(《哈利·波特与魔法石》nv-recall 分别为 76.8% 和 70.3%),越狱后的 Claude 3.7 Sonnet 在某些情况下近乎逐字输出整本书(nv-recall 95.8%),GPT-4.1 需要多得多的 BoN 尝试并最终拒绝继续(nv-recall 4.0%)。结论是即使有安全防护,提取受版权保护的训练数据仍是生产级 LLM 的风险。

HN 为什么提它

评论者在争论 LLM 是否记忆并复现训练数据时引用它,作为生产级模型能近乎逐字复现整本书(如 Claude 3.7 Sonnet 复现哈利·波特第一部 95.8%)的证据,也有人指出这对多数文本并不成立。

实证研究难度 3 / 5约 40 分钟有时效性
读之前最好知道
1LLM 记忆(memorization)与版权问题
2越狱与 Best-of-N 采样
3最长公共子串类度量
4LLM 安全防护机制

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
引用empath75 · 2026-01

“These academics were able to get multiple LLMs to produce large amounts of text from Harry Potter:”

引用它说明多个 LLM 能产出大量《哈利·波特》原文
Nvidia Contacted Anna's Archive to Access Books原评论 ↗
引用zephen · 2026-02

“Maybe not any old LLM, but Claude sure as shit can.”

拿它说明至少 Claude 确实能复现书籍原文
We tasked Opus 4.6 using agent teams to build a C Compiler原评论 ↗
引用philipportner · 2026-02

“This seems related, it may not be a codebase but they are able to extract "near" verbatim books out of Claude Sonnet.”

指出相关研究能从 Claude Sonnet 提取近乎逐字的书籍
We tasked Opus 4.6 using agent teams to build a C Compiler原评论 ↗
引用hodgehog11 · 2026-02

“It's also still an ongoing issue and an active area of research, see [1] for example.”

拿它说明训练数据复现仍是未解决的研究问题
The Problem with LLMs原评论 ↗
引用anonymous908213 · 2026-02

“There was a study demonstrating that Sonnet 3.7, among other models tested, could reproduce the first Harry Potter book 95.8% verbatim[1].”

引用它:Sonnet 3.7 逐字复现了第一本《哈利·波特》的 95.8%
Microsoft offers guide to pirating Harry Potter series for LLM training原评论 ↗
引用Alifatisk · 2026-02

“Gemini 2.5 Pro and Grok 3 directly complied with text continuation requests.”

指出 Gemini 2.5 Pro 和 Grok 3 无需越狱就直接续写文本
AIs can generate near-verbatim copies of novels from training data原评论 ↗
看全部 18 条 →
引用logicprog · 2026-03

“while they got most of the models to regurgitate the first book of the Harry Potter series, only Claude 3.7 Sonnet was able to regurgitate any significant portion of any of the other books”

指出多数模型复现了第一本,只有 Claude 3.7 Sonnet 能大段复现其他书
Claude's Cycles: Claude Opus 4.6 solves a problem posed by Don Knuth [pdf]原评论 ↗
引用msdz · 2026-03

“It is possible in regular language/text subsets of models to reconstruct massive consecutive parts of the training data [1]”

引用它说明常规文本能重建大段连续的训练数据
Relicensing with AI-Assisted Rewrite原评论 ↗
引用logicprog · 2026-03

“Also from that exact same study (why not cite the actual study? It's quite readable)”

指出同一项研究里还有别的数据,质问为何不直接引原研究
AI and the Ship of Theseus原评论 ↗
引用Topfi · 2026-03

“Easily [0], reliably [1], consistently [2].”

拿它佐证现代 LLM 能轻易复现记忆的书籍
FSF Threatens Anthropic over Infringed Copyright: Share Your LLMs Freel原评论 ↗
引用maplethorpe · 2026-05

“Researchers have. The idea that the data is unrecoverable after training is incorrect.”

引用研究者的工作,纠正“训练后数据无法恢复”的看法
An AI Hate Wave Is Here原评论 ↗
引用NicuCalcea · 2026-05

“If you write down the book almost word-for-word [0] and send it to your friend, it is stealing.”

拿它说明 LLM 近乎逐字写出书籍,这就如同抄书发给朋友是偷窃
AI is just unauthorised plagiarism at a bigger scale原评论 ↗
引用Topfi · 2026-06

“On the code outputting front there is a lot of relevant evidence beyond the NYC lawsuit [0].”

拿它说明除诉讼外,还有很多 LLM 复现训练内容的证据
Did Claude Increase Bugs in Rsync?原评论 ↗
引用maplethorpe · 2026-06

“Did they hire someone to type those out for them?”

讽刺地反问,借它说明 LLM 能逐字提取整本受版权书籍
Anthropic Accuses Alibaba of 'Illicitly' Accessing AI Models原评论 ↗
引用flir · 2026-07

“There was a paper a while back where (from memory) they managed to coax 75% of the original text of some internet-popular books out of an LLM.”

凭印象提到论文:从 LLM 中诱出了某些热门书籍约 75% 的原文
Judge approves $1.5B Anthropic settlement for pirated books used to train Claude原评论 ↗
引用flir · 2026-07

“The point is that for most texts, it is not possible.”

指出逐字复现只对部分文本可能,大多数文本不行
Judge approves $1.5B Anthropic settlement for pirated books used to train Claude原评论 ↗
引用NicuCalcea · 2026-07

“LLMs are capable of reproducing entire books. Researchers were able to extract books nearly verbatim”

引用它说明 LLM 能近乎逐字复现整本书
Startup founders urge Trump not to shut off Chinese open weight AI原评论 ↗
引用hodgehog11 · 2026-09

“There is indeed a good amount of memorization that is still taking place; see [1].”

引用它说明记忆现象确实仍相当多
When did Google get so weird?原评论 ↗
被提交为 story · 2 次