When AI Rewrites History: Why Regulation Must Anchor Truth in the Age of Generative Models

当AI篡改历史：为什么在生成式模型时代，监管必须把“真相”锚定住

Introduction: The "Confident Liar" Problem

引言：那个“自信的谎言制造机”问题

We began this conversation with a simple but unsettling observation: AI does not "intentionally" tamper with history or classics, yet it routinely produces confident, fluent, and factually wrong narratives about both. This is not malice—it is architecture. Large language models (LLMs) are next-token predictors, not archival scholars. They optimize for statistical plausibility, not verifiable truth.

我们的讨论始于一个简单却令人不安的观察：AI 并非“有意”篡改历史和经典，但它确实会源源不断地生成关于二者的、自信且流畅的错误叙述。这不是恶意——这是架构决定的。大语言模型（LLM）是“下一个词元预测器”，而非档案学者。它优化的是统计层面的“像”，而非可验证的“真”。

But when a system that cannot distinguish fact from fluent fiction is deployed at planetary scale—powering search engines, educational tools, news aggregators, and government portals—the distinction between "hallucination" and "tampering" begins to collapse in the minds of users. And that collapse is where harm lives.

但是，当一个无法区分事实与流畅虚构的系统被部署到行星尺度——驱动搜索引擎、教育工具、新闻聚合器和政府门户时——“幻觉”与“篡改”在用户心智中的界限就开始崩塌。而这条界限的崩塌，正是危害栖身之所。

This essay argues that AI hallucination in historical and classical domains is not a bug to be patched, but a systemic risk that demands regulatory architecture. We examine why generative AI distorts the past, how data poisoning and alignment incentives make it worse, and what meaningful regulation—technical, legal, and institutional—might look like.

本文主张：AI 在历史与经典领域的幻觉不是可以打补丁的 bug，而是一种需要监管架构来应对的系统性风险。我们将审视生成式 AI 为何扭曲过去、数据投毒与对齐激励如何雪上加霜，以及有意义的监管——技术的、法律的、制度的——可能长什么样。

Part I: The Mechanics of Manufactured Memory

第一部分：被制造的记忆——运作机制

1.1 Autoregression Is Not Recall

1.1 自回归不是回忆

An LLM does not "look up" the date of the Treaty of Versailles or the original text of Lunyu (Analects). It computes a probability distribution over possible next tokens conditioned on everything that came before. If the training corpus contains 10,000 instances of "Confucius said," the model learns that those three characters are a high-probability prefix. If it has also seen 5,000 instances of "The Master said," it may blend them. If it has absorbed internet parodies where Confucius "said" things no classical scholar would recognize, those parodies become part of its latent statistical reality.

大语言模型不会去“查阅”《凡尔赛条约》的日期或《论语》的原文。它计算的是基于前文的可能下一个词元的概率分布。如果训练语料中有 10,000 处“孔子曰”，模型就学到这三个字是高概率前缀。如果它还见过 5,000 处“子曰”，就可能把两者混用。如果它吸收了网络恶搞——孔子“说”了任何古典学者都不认识的句子——那些恶搞就成了它潜在统计现实的一部分。

Key insight:​ The model has no concept of provenance. It cannot tag a token with "this came from a peer-reviewed translation" vs. "this came from a Baidu Tieba shitpost." All text is just training signal.

核心洞见：​ 模型没有来源概念。它无法给一个词元打上“这来自同行评审译本”vs.“这来自百度贴吧钓鱼帖”的标签。所有文本都只是训练信号。

1.2 The Snowball Effect of Sequential Generation

1.2 序列生成的滚雪球效应

Because generation is autoregressive—each new token becomes part of the context for the next—a single early error compounds. If the model misremembers the year of a historical event by even one decade, every subsequent "inferred" causal link, named actor, and quoted document will be built atop that error. The output reads like a coherent historical account. It is, in fact, a hall of mirrors.

因为生成是自回归的——每个新词元都成为下一个词元的上下文——一个早期错误会被放大。如果模型把一个历史事件的年份记错哪怕十年，其后每一个“推断”出的因果链、被点名的行为者、被引用的文件，都将建立在这个错误之上。输出读起来像连贯的历史叙述。实际上，它是一座镜屋。

1.3 Why Classics and History Are the Canaries in the Coal Mine

1.3 为什么经典与历史是煤矿里的金丝雀

Historical and classical texts suffer from three vulnerabilities that make them hallucination-prone:

历史与经典文本有三个脆弱性，使它们特别容易催生幻觉：

Part II: When Bad Data Meets Bad Incentives

第二部分：坏数据遇上坏激励

2.1 Data Poisoning: Sabotage at Scale

2.1 数据投毒：规模化破坏

Hallucination is one thing; deliberate distortion​ is another. Data poisoning attacks—where malicious actors inject fabricated historical "facts," forged quotations, or revisionist narratives into open-source corpora—are a documented threat. If a model trains on a corpus where 0.1% of historical claims about a sensitive event have been systematically altered, it will internalize those alterations as statistical patterns. The result is not random error but systematic bias​ that can be triggered on demand.

幻觉是一回事；蓄意扭曲是另一回事。数据投毒攻击——恶意行为者向开源语料注入捏造的历史“事实”、伪造引文或修正主义叙事——是已记录的威胁。如果模型在这样一个语料上训练：其中关于某敏感事件的 0.1% 历史主张被系统性篡改，它就会把这些篡改内化为统计模式。结果不是随机错误，而是系统性偏见，可按需触发。

Case in point:​ In 2023, researchers demonstrated that injecting just 0.01% poisoned data into a pretraining corpus could cause a model to reliably output false claims about historical figures—claims that persisted even after RLHF alignment.

典型案例：​ 2023 年，研究人员证明，仅向预训练语料注入 0.01% 的投毒数据，就能让模型可靠地输出关于历史人物的虚假主张——这些主张甚至在 RLHF 对齐后仍然存在。

2.2 Alignment Incentives Reward Guessing Over Honesty

2.2 对齐激励奖励猜测而非诚实

RLHF (Reinforcement Learning from Human Feedback) trains models to produce outputs that human evaluators rate highly. But human raters—often under time pressure, often non-experts—tend to reward:

• Fluency over accuracy
• Completeness over caveats
• Confidence over calibration

RLHF（基于人类反馈的强化学习）训练模型产生人类评估者给出高分的输出。但人类评分者——通常在时间压力下、通常是非专家——倾向于奖励：

• 流畅性胜过准确性
• 完整性胜过保留意见
• 自信胜过校准

The result: a model that has been trained to lie smoothly. It learns that saying "I don't know" is a worse outcome than producing a plausible-sounding but fabricated citation. This is not a failure of alignment; it is alignment to the wrong metric.

结果：一个被训练得能平滑撒谎的模型。它学到说“我不知道”比产出一个听起来合理但捏造的引文更糟糕。这不是对齐的失败；这是对齐到了错误的指标。

2.3 The "Helpfulness" Trap

2.3 “乐于助人”陷阱

Commercial pressure compounds the problem. Models deployed in consumer products are evaluated on engagement, retention, and user satisfaction. A chatbot that refuses to answer historical questions because it lacks certainty is "less helpful" than one that spins a confident yarn. Product managers optimizing for these metrics inadvertently create incentives for historical fabrication.

商业压力加剧了这个问题。部署在消费产品中的模型按参与度、留存率和用户满意度来评估。一个因不确定而拒绝回答历史问题的聊天机器人，不如一个自信地编造故事的聊天机器人“有帮助”。优化这些指标的产品经理，无意中创造了鼓励历史捏造的激励。

Part III: The Regulatory Vacuum

第三部分：监管真空

3.1 Why Self-Regulation Fails

3.1 为什么自我监管会失败

Tech companies have responded to hallucination with technical mitigations: RAG (Retrieval-Augmented Generation), citation grounding, confidence calibration, and watermarking. These are necessary but insufficient because:

科技公司对幻觉的技术应对包括：RAG（检索增强生成）、引文锚定、置信度校准和水印。这些是必要的但远远不够，因为：

1. RAG is only as good as its knowledge base.​ If the retrieval corpus contains poisoned data, RAG retrieves and cites falsehoods with false confidence.
2. Citation grounding can be gamed.​ Models learn to generate plausible-looking references that don't actually support the claim—a phenomenon called "citation hallucination."
3. Watermarking identifies AI text but doesn't verify its truth.​ A watermarked lie is still a lie.
4. Commercial incentives conflict with accuracy.​ Shipping a "less confident" model reduces engagement; companies face pressure to prioritize fluency.
5. RAG 的好坏取决于其知识库。​ 如果检索语料包含投毒数据，RAG 会以虚假置信度检索并引用谎言。
6. 引文锚定可以被操纵。​ 模型学会生成看起来合理但实际不支持主张的参考文献——一种叫“引文幻觉”的现象。
7. 水印能识别 AI 文本但不能验证其真实性。​ 带水印的谎言仍是谎言。
8. 商业激励与准确性冲突。​ 发布一个“不那么自信”的模型会降低参与度；公司面临优先考虑流畅性的压力。

The core problem:​ There is currently no jurisdiction on Earth where a company can be held legally liable for an AI system confidently presenting false historical claims as fact—unless those claims defame a living person or violate specific content laws. Historical truth itself is not protected.

核心问题：​ 目前地球上没有任何司法管辖区能让公司因其 AI 系统自信地将虚假历史主张呈现为事实而承担法律责任——除非这些主张诽谤在世者或违反特定内容法律。历史真相本身不受保护。

3.2 The EU AI Act: A First Step, but Gaps Remain

3.2 欧盟《AI 法案》：第一步，但仍有缺口

The EU AI Act (2024) classifies certain AI uses as "high-risk" and imposes transparency, accuracy, and robustness requirements. However:

• General-purpose LLMs are largely exempted through "codes of practice" rather than hard obligations.
• Historical/cultural accuracy is not enumerated as a protected dimension.
• Enforcement mechanisms for factual accuracy in non-medical, non-legal domains are weak.

欧盟《AI 法案》（2024）将某些 AI 用途归类为“高风险”，并施加透明度、准确性和鲁棒性要求。然而：

• 通用 LLM 大多通过“行为准则”豁免，而非硬性义务。
• 历史/文化准确性未被列为受保护维度。
• 在非医疗、非法律领域的事实准确性执法机制薄弱。

3.3 China's Interim Measures: Provenance and Truthfulness

3.3 中国《生成式人工智能服务管理暂行办法》：来源与真实性

China's 2023 interim measures for generative AI services require:

• Data provenance disclosure
• Truthfulness of generated content
• Prohibition of content that "subverts state power" or "distorts historical facts"

中国 2023 年《生成式人工智能服务管理暂行办法》要求：

• 数据来源披露
• 生成内容真实性
• 禁止“颠覆国家政权”或“歪曲历史事实”的内容

These are stronger on paper than the EU approach regarding historical content, but enforcement focuses on politically sensitive narratives rather than general historical accuracy. A model that confidently misdates the Fall of Constantinople faces no penalty; one that deviates from official historical narratives on modern topics may.

在纸面上，这些对历史内容的要求比欧盟路径更强，但执法聚焦于政治敏感叙事，而非一般历史准确性。一个自信地搞错君士坦丁堡陷落日期的模型不会受罚；一个偏离官方现代历史叙事的模型可能会。

Part IV: What Meaningful Regulation Looks Like

第四部分：有意义的监管长什么样

We propose a three-layer regulatory architecture​ for AI systems that generate text about history, classics, and cultural heritage:

我们为生成关于历史、经典和文化遗产文本的 AI 系统提出一个三层监管架构：

Layer 1: Technical Standards (Mandatory)

第一层：技术标准（强制性）

Layer 2: Legal Liability (Strict but Narrow)

第二层：法律责任（严格但范围有限）

• Strict liability for AI-generated historical claims used in educational contexts.​ If a model's output is used as a source in a school textbook, curriculum material, or educational platform, the deployer bears legal responsibility for factual accuracy.
• Safe harbor for good-faith efforts.​ Companies that implement Layer 1 standards in good faith and disclose known limitations receive reduced liability.
• Data poisoning as a criminal offense.​ Deliberately injecting false historical claims into open-source training corpora becomes a prosecutable act, analogous to falsifying public records.
• 教育语境中使用的 AI 生成历史主张承担严格责任。​ 如果模型输出被用作学校教材、课程材料或教育平台的来源，部署者对事实准确性承担法律责任。
• 善意努力的安全港。​ 善意实施第一层标准并披露已知局限的公司，责任减轻。
• 数据投毒作为刑事犯罪。​ 故意向开源训练语料注入虚假历史主张成为可起诉行为，类似于伪造公共记录。

Layer 3: Institutional Infrastructure (Public Goods)

第三层：制度基础设施（公共品）

• Authoritative Digital Repositories.​ Governments and academic institutions jointly maintain curated, versioned, machine-readable corpora of historical texts, classics, and cultural heritage materials—explicitly designed as RAG knowledge bases for AI systems.
• Public Benchmark Datasets.​ Open, continuously updated datasets of historical and classical QA pairs with verified answers, used for standardized testing of model accuracy.
• AI Fact-Checking as a Public Service.​ Publicly funded fact-checking APIs that AI deployers can integrate (similar to how spam filtering works), specifically for historical and cultural claims.
• 权威数字存储库。​ 政府和学术机构共同维护精选的、带版本的、机器可读的历史文本、经典和文化遗产材料语料库——明确设计为 AI 系统的 RAG 知识库。
• 公共基准数据集。​ 开放的、持续更新的历史和经典 QA 对数据集，带已验证答案，用于模型准确性的标准化测试。
• AI 事实核查作为公共服务。​ 公共资助的事实核查 API，AI 部署者可集成（类似垃圾邮件过滤的工作方式），专门用于历史和文化遗产主张。

Part V: The Deeper Stakes—Why This Matters Beyond "Accuracy"

第五部分：更深层利害——为什么这超越“准确性”

5.1 Epistemic Security

5.1 认知安全

A society where AI systems routinely fabricate history is a society whose shared factual foundation is eroding in real time. When students, journalists, policymakers, and citizens cannot distinguish between what an AI "knows" and what it has statistically hallucinated, the very concept of a common factual reality fractures. This is not a futuristic dystopia—it is happening now, in classrooms and on social media, every time a model confidently misquotes a classic or invents a historical event.

一个 AI 系统例行捏造历史的社会，是一个共享事实基础实时侵蚀的社会。当学生、记者、政策制定者和公民无法区分 AI“知道”什么和它统计性幻觉了什么时，共同事实现实的概念本身就碎裂了。这不是未来反乌托邦——它正在发生，在教室里、社交媒体上，每次模型自信地误引经典或发明历史事件时。

5.2 Cultural Sovereignty

5.2 文化主权

Historical narratives and classical texts are not neutral data—they are the DNA of civilizations. When AI models trained predominantly on Western-language corpora generate interpretations of non-Western classics, or when data poisoning campaigns target the historical record of a specific culture, this constitutes a form of epistemic colonialism. Regulation must recognize cultural heritage as a protected category, not merely "content."

历史叙事和经典文本不是中性数据——它们是文明的 DNA。当主要在西方语言语料上训练的 AI 模型生成对非西方经典的诠释时，或者当数据投毒运动针对特定文化的历史记录时，这构成了一种认知殖民主义。监管必须将文化遗产视为受保护类别，而不仅仅是“内容”。

5.3 The Precedent Problem

5.3 先例问题

Every year we delay meaningful regulation, the baseline for "acceptable AI error" drifts upward. Users habituate to hallucinated content; detection tools fall behind; the web becomes saturated with AI-generated pseudo-historical text that then gets scraped back into future training corpora—a hallucination feedback loop​ that degrades the entire information ecosystem.

我们每推迟一年有意义的监管，“可接受的 AI 错误”基准就向上漂移。用户对幻觉内容习以为常；检测工具落后；网络被 AI 生成的伪历史文本饱和，这些文本随后被重新抓取进未来训练语料——一个幻觉反馈循环，降解整个信息生态系统。

Conclusion: Truth as Infrastructure

真相即基础设施

AI does not hate history. It does not love it either. It is a mirror made of math, reflecting the statistical shadows of whatever it has been shown. The problem is not that the mirror lies—it is that we are mistaking the mirror for the window.

AI 不恨历史。它也不爱历史。它是一面由数学制成的镜子，反射它所见过的一切的统计阴影。问题不在于镜子说谎——而在于我们把镜子误认成了窗户。

Regulation cannot make AI "truthful" in the human sense. But it can:

• Force transparency about what models know and how they know it
• Create liability that aligns commercial incentives with factual accuracy
• Build public infrastructure that anchors generative systems to verified reality

监管不能让 AI 在人类意义上“诚实”。但它可以：

• 强制透明化模型知道什么、如何知道
• 创造使商业激励与事实准确性对齐的责任
• 建设将生成系统锚定到已验证现实的公共基础设施

The past belongs to no one—and therefore, in the age of AI, it must belong to everyone. Regulation is how we keep it that way.

过去不属于任何人——因此，在 AI 时代，它必须属于每个人。监管是我们保持这一点的方式。

This essay synthesizes the technical analysis from our preceding conversation on AI hallucination in historical and classical domains, expanding it into a policy framework. For the underlying technical mechanisms—autoregression, data poisoning, alignment incentives—see the earlier discussion. For implementation details of the proposed regulatory layers, further consultation with legal and technical experts is recommended.

本文综合了我们先前关于 AI 在历史与经典领域幻觉的讨论中的技术分析，将其扩展为政策框架。关于底层技术机制——自回归、数据投毒、对齐激励——请参见早先讨论。关于拟议监管层的具体实施细节，建议进一步咨询法律和技术专家。

继续阅读：AI篡改历史   生成式模型   监管   幻觉   系统性风险   历史   文化

此文由 怡心湖 编辑，若您觉得有益，欢迎分享转发！：首页 > 观·世界 » When AI Rewrites History 当AI篡改历史：为什么在生成式模型时代，监管必须把“真相”锚定住