AI在招聘、信贷中如何造成不公平？
How can AI create unfairness in hiring and lending?
若训练数据带有历史歧视，模型会延续甚至放大偏见，导致某些群体被自动拒掉，且理由常不透明、难以申诉。
If training data carries historical bias, models repeat or amplify it, auto-rejecting groups with opaque, hard-to-contest reasons.
难度 专业 · 适合 高中(15-18)
CC BY 4.0 · 已核验 · Verified 2026-08-23 · 2 条独立证据
来源 · Source
Russell & Norvig,《Artificial Intelligence: A Modern Approach》第4版,Pearson,2020
S. Russell and P. Norvig, Artificial Intelligence: A Modern Approach, 4th ed., Pearson, 2020
独立核验依据 · Independent verification: 核验依据 2 ARCHIVE 标识 isbn_9788177583670：Artificial Intelligence: A
#算法偏见 #公平 #招聘
什么是 AI 对齐？
What is AI alignment?
AI 对齐指让系统的目标和行为与人类的真实意图和价值观一致，避免它为完成任务而采取有害手段。难点在于把模糊的人类偏好准确转化为可优化的目标。这是安全部署强大 AI 的核心研究课题。
AI alignment means making a system's goals and behavior match human intent and values, avoiding harmful shortcuts taken just to finish a task. The hard part is turning vague human preferences into an optimizable objective. It is central to safely deploying powerful AI.
难度 专业 · 适合 初中(12-15)
CC BY 4.0 · 已核验 · Verified 2026-08-23 · 2 条独立证据
来源 · Source
Russell & Norvig,《Artificial Intelligence: A Modern Approach》第4版,Pearson,2020
S. Russell and P. Norvig, Artificial Intelligence: A Modern Approach, 4th ed., Pearson, 2020
独立核验依据 · Independent verification: 核验依据 2 ARCHIVE 标识 isbn_9788177583670：Artificial Intelligence: A
#AI对齐 #目标 #安全
为什么人们担心 AI 武器化？
Why are people concerned about weaponized AI?
自主武器可让 AI 决定攻击目标，降低了使用暴力的门槛并可能失控。误判和算法偏见会让平民面临更大风险。国际社会正讨论是否应限制致命性自主武器系统。
Autonomous weapons let AI choose targets, lowering the threshold for violence and risking loss of control. Misjudgment and bias can endanger civilians more. The international community debates limiting lethal autonomous weapons systems.
难度 专业 · 适合 初中(12-15)
CC BY 4.0 · 已核验 · Verified 2026-11-04
来源 · Source
SPARROW,《Killer Robots》,Journal of Applied Philosophy,2007 (DOI: 10.1111/j.1468-5930.2007.00346.x)
SPARROW, 'Killer Robots', Journal of Applied Philosophy, 2007, DOI: 10.1111/j.1468-5930.2007.00346.x
独立核验依据 · Independent verification: 核验依据 1 Crossref DOI 10.1111/j.1468-5930.2007.00346.x
#自主武器 #安全 #治理
公平性在算法上如何度量？
How is fairness measured for algorithms?
公平性常用统计指标衡量，如不同群体间的准确率、假阳性率是否接近（均等机会），或预测阳性比例是否一致（人口均等）。不同定义之间常互相冲突，难以同时满足。选择哪种取决于具体场景的价值观。
Fairness is often measured by statistics like equal accuracy or false-positive rates across groups (equal opportunity), or equal positive-prediction rates (demographic parity). These definitions frequently conflict and cannot all hold at once. The choice depends on the values of the specific context.
难度 专业 · 适合 初中(12-15)
CC BY 4.0 · 已核验 · Verified 2026-08-23 · 2 条独立证据
来源 · Source
Russell & Norvig,《Artificial Intelligence: A Modern Approach》第4版,Pearson,2020
S. Russell and P. Norvig, Artificial Intelligence: A Modern Approach, 4th ed., Pearson, 2020
独立核验依据 · Independent verification: 核验依据 2 ARCHIVE 标识 isbn_9788177583670：Artificial Intelligence: A
#公平性 #指标 #群体
为什么“AI 为什么这么判断”是个难题？
Why is “why did the AI decide this” so hard?
深度神经网络把输入经过成千上万参数的非线性变换后输出结果，内部是高维中间表示，没有一个变量对应“可读的理由”，所以被称为黑箱。可解释性方法分两类：全局的（如注意力权重、特征重要性）和局部的（对单个预测给出归因，如 LIME、SHAP）。在医疗、信贷等高风险场景，可解释性直接关系到责任认定和监管合规。
A deep network pushes input through thousands of nonlinear parameters into high-dimensional intermediate representations where no single variable reads as a “reason,” which is why it is called a black box. Explainability splits into global methods (attention weights, feature importance) and local ones attributing individual predictions (LIME, SHAP). In high-stakes settings like medicine and credit, explainability ties directly to accountability and regulatory compliance.
难度 专业 · 适合 高中(15-18)
CC BY 4.0 · 已核验 · Verified 2026-11-04
来源 · Source
Takefuji,《Beyond the black box: Enhancing feature explainability in machine learning with SHAP and complementary approaches》,Computers and Electronics in Agriculture,2026 (DOI: 10.1016/j.compag.2025.111226)
Takefuji, 'Beyond the black box: Enhancing feature explainability in machine learning with SHAP and complementary approaches', Computers and Electronics in Agriculture, 2026, DOI: 10.1016/j.compag.2025.111226
独立核验依据 · Independent verification: 核验依据 1 Crossref DOI 10.1016/j.compag.2025.111226
#可解释性 #黑箱 #SHAP
AI生成的内容版权归谁？
Who owns the copyright of AI-generated content?
各国规定不一，多数认为纯AI生成因缺少“人类作者”不受版权保护，而用AI训练他人作品也引发版权纠纷。
Rules differ; many say purely AI-made output lacks a human author and is unprotected, while training on others’ work also sparks disputes.
难度 专业 · 适合 高中(15-18)
CC BY 4.0 · 来源待核验 · Source pending review
来源 · Source
出处备注：著录含书名，但缺版次 / 出版社 / 年份中的一部分，需编辑补全后才能做独立比对。
美国版权局,《Copyright and Artificial Intelligence: Part 1, Creativity》(2023)
U.S. Copyright Office, Copyright and Artificial Intelligence: Part 1, Creativity (2023)
#版权 #AI生成 #法律
训练大模型的环境成本有多高？
How high is the environmental cost of training big models?
大规模训练耗电巨大，数据中心冷却还要消耗大量水，碳排放可观，因此绿色AI和模型蒸馏受到关注。
Large training uses huge electricity and water for cooling, with notable carbon emissions, so green AI and distillation get attention.
难度 专业 · 适合 高中(15-18)
CC BY 4.0 · 已核验 · Verified 2026-08-23 · 2 条独立证据
来源 · Source
Russell & Norvig,《Artificial Intelligence: A Modern Approach》第4版,Pearson,2020
S. Russell and P. Norvig, Artificial Intelligence: A Modern Approach, 4th ed., Pearson, 2020
独立核验依据 · Independent verification: 核验依据 2 ARCHIVE 标识 isbn_9788177583670：Artificial Intelligence: A
#碳排放 #能耗 #绿色AI
大模型的“幻觉”在医疗或法律上为什么特别危险？
Why are LLM hallucinations especially dangerous in medicine or law?
它们会编造看似专业的文献或法条，外行很难识别，可能误导治疗或案件，造成严重后果。
They fabricate plausible-sounding papers or statutes that laypeople can’t easily spot, misguiding treatment or cases with serious harm.
难度 专业 · 适合 高中(15-18)
CC BY 4.0 · 已核验 · Verified 2026-08-23 · 2 条独立证据
来源 · Source
Russell & Norvig,《Artificial Intelligence: A Modern Approach》第4版,Pearson,2020
S. Russell and P. Norvig, Artificial Intelligence: A Modern Approach, 4th ed., Pearson, 2020
独立核验依据 · Independent verification: 核验依据 2 ARCHIVE 标识 isbn_9788177583670：Artificial Intelligence: A
#幻觉 #准确性 #风险
什么是“提示词注入”，为什么有风险？
What is prompt injection and why is it risky?
攻击者在输入里藏入指令，让大模型执行非预期操作，比如泄露数据或代发消息，是LLM应用的新攻击面。
Attackers hide instructions in input to make an LLM do unintended things like leak data or send messages—a new attack surface for LLM apps.
难度 专业 · 适合 高中(15-18)
CC BY 4.0 · 已核验 · Verified 2026-08-23 · 2 条独立证据
来源 · Source
Russell & Norvig,《Artificial Intelligence: A Modern Approach》第4版,Pearson,2020
S. Russell and P. Norvig, Artificial Intelligence: A Modern Approach, 4th ed., Pearson, 2020
独立核验依据 · Independent verification: 核验依据 2 ARCHIVE 标识 isbn_9788177583670：Artificial Intelligence: A
#提示注入 #安全 #LLM
为什么医生不敢完全相信AI的诊断？
Why do doctors hesitate to fully trust AI diagnosis?
模型常是“黑箱”，难以解释判断依据，而误诊后果严重，所以可解释AI和“人在回路”更受信赖。
Models are often black boxes with unclear reasoning, and misdiagnosis is costly, so explainable AI and human-in-the-loop are preferred.
难度 专业 · 适合 高中(15-18)
CC BY 4.0 · 已核验 · Verified 2026-10-04
来源 · Source
E. Topol,《High-performance medicine: the convergence of human and artificial intelligence》,Nature Medicine,2019,25(1):44–56
E. Topol, "High-performance medicine: the convergence of human and artificial intelligence", Nature Medicine, 2019, 25(1):44-56
独立核验依据 · Independent verification: 核验依据 1 DOI 10.1038/s41591-018-0300-7
#可解释 #黑箱 #医疗
自动驾驶出事故，责任算谁的？
Who is liable when a self-driving car crashes?
在L3及以上，人与车的责任边界模糊，可能涉及车主、厂商或算法提供方，保险与法律仍在探索中。
At L3 and above the line between driver and car blurs, involving owner, maker, or algorithm provider; law and insurance are still evolving.
难度 专业 · 适合 高中(15-18)
CC BY 4.0 · 已核验 · Verified 2026-08-23 · 2 条独立证据
来源 · Source
Russell & Norvig,《Artificial Intelligence: A Modern Approach》第4版,Pearson,2020
S. Russell and P. Norvig, Artificial Intelligence: A Modern Approach, 4th ed., Pearson, 2020
独立核验依据 · Independent verification: 核验依据 2 ARCHIVE 标识 isbn_9788177583670：Artificial Intelligence: A
#自动驾驶 #责任 #伦理
用 AI 生成的内容反复训练下一代 AI，会发生什么？
What happens when each new generation of AI is trained mainly on the previous generation's output?
会出现模型崩塌：每一代都从上一代的采样输出中学习，采样误差与拟合误差逐代累积，分布的尾部（罕见事件、少数群体的表达）先被抹平，最终连主体部分也退化成方差极小的少数模式。Nature 2024 年发表的实验显示，即使每一代数据量充足，纯用递归生成数据训练的语言模型也会在几代之内开始输出无意义的重复内容。这意味着可溯源的真实人类语料具有长期价值，实践上需要保留原始语料、每轮混入真实数据，并对训练集做来源标识与过滤。
The outcome is model collapse: each generation learns from the previous one's samples, so sampling and approximation errors compound, first erasing the tails of the distribution — rare events and minority phrasings — and eventually degenerating even the bulk into a few low-variance modes. Experiments published in Nature in 2024 showed that language models trained purely on recursively generated data begin emitting meaningless repetition within a handful of generations, even with plentiful data each round. The implication is that provenance-tracked human data holds lasting value, so practitioners preserve original corpora, blend real data into every round, and filter training sets by source.
难度 专业 · 适合 高中(15-18)
CC BY 4.0 · 已核验 · Verified 2026-11-02
来源 · Source
Shumailov 等《AI models collapse when trained on recursively generated data》(Nature, 2024)
Shumailov et al., "AI models collapse when trained on recursively generated data," Nature (2024)
独立核验依据 · Independent verification: 核验依据 1 CROSSREF 标识 10.1038/s41586-024-07566-y：AI models collaps
#模型崩塌 #合成数据 #数据来源 #训练退化
大模型训练会涉及个人隐私吗？
Can large model training involve personal privacy?
公开网页里常含姓名、邮件等个人信息，模型可能“记住”并在回答中泄露，需要用去标识化和差分隐私来防护。
Public web pages hold names and emails; models may memorize and leak them, so de-identification and differential privacy are used for protection.
难度 专业 · 适合 高中(15-18)
CC BY 4.0 · 已核验 · Verified 2026-08-23 · 2 条独立证据
来源 · Source
Russell & Norvig,《Artificial Intelligence: A Modern Approach》第4版,Pearson,2020
S. Russell and P. Norvig, Artificial Intelligence: A Modern Approach, 4th ed., Pearson, 2020
独立核验依据 · Independent verification: 核验依据 2 ARCHIVE 标识 isbn_9788177583670：Artificial Intelligence: A
#隐私 #训练数据 #泄露
深度伪造（Deepfake）是怎么做出来的？为什么难防范？
How are deepfakes made, and why are they hard to stop?
深度伪造用生成对抗网络（GAN）等模型训练出能生成逼真假视频、假声音的网络：生成器不断造假，判别器不断鉴别，互相逼对方变强。它难防范是因为人眼和耳朵难以分辨，且生成成本越来越低；应对靠检测模型、内容来源认证和平台治理，本质上是一场“造假—检测”的军备竞赛。
Deepfakes are made with models like generative adversarial networks (GANs): a generator keeps forging, a discriminator keeps detecting, and each pushes the other to improve until fake video and audio look and sound real. They are hard to stop because humans struggle to tell, and the cost keeps dropping; defenses rely on detection models, provenance verification, and platform governance—an arms race between forgery and detection.
难度 进阶 · 适合 大众
CC BY 4.0 · 来源待核验 · Source pending review
来源 · Source
出处备注：此条未指向带版次的可引用著作（通识性内容、机构文件或未展开的英文书名），本库不提供独立书目出处；如需精确引用，需由编辑补全具体篇目与版次。
arXiv（生成对抗网络论文）
arXiv (GAN paper)
#深度伪造 #GAN #内容安全
AI能可靠地鉴别AI生成的内容吗？
Can AI reliably detect AI-generated content?
生成与检测是场军备竞赛，模型迭代很快，水印也能被去除，目前还没有完全可靠的鉴别方法。
Generation and detection are an arms race; models evolve fast and watermarks can be removed, so no fully reliable detector exists yet.
难度 进阶 · 适合 高中(15-18)
CC BY 4.0 · 已核验 · Verified 2026-08-23 · 2 条独立证据
来源 · Source
Russell & Norvig,《Artificial Intelligence: A Modern Approach》第4版,Pearson,2020
S. Russell and P. Norvig, Artificial Intelligence: A Modern Approach, 4th ed., Pearson, 2020
独立核验依据 · Independent verification: 核验依据 2 ARCHIVE 标识 isbn_9788177583670：Artificial Intelligence: A
#检测 #水印 #对抗
和聊天机器人聊天会让人产生情感依赖吗？
Can chatbots create emotional dependence?
拟人化的对话能满足陪伴需求，孤独的人群可能过度依赖，反而影响现实中的真实人际关系。
Human-like chat satisfies companionship, and lonely users may over-rely on it, hurting real-life relationships.
难度 进阶 · 适合 初中(12-15)
CC BY 4.0 · 已核验 · Verified 2026-08-23 · 2 条独立证据
来源 · Source
Russell & Norvig,《Artificial Intelligence: A Modern Approach》第4版,Pearson,2020
S. Russell and P. Norvig, Artificial Intelligence: A Modern Approach, 4th ed., Pearson, 2020
独立核验依据 · Independent verification: 核验依据 2 ARCHIVE 标识 isbn_9788177583670：Artificial Intelligence: A
#情感 #依赖 #聊天机器人
自主武器（杀手机器人）有什么伦理争议？
What are the ethical concerns about autonomous weapons?
把致命决定交给算法缺乏人道判断，也难以界定责任主体，因此国际社会呼吁对致命自主武器加以限制。
Letting algorithms make lethal choices lacks human judgment and clear accountability, prompting international calls to limit such weapons.
难度 进阶 · 适合 高中(15-18)
CC BY 4.0 · 已核验 · Verified 2026-08-23 · 2 条独立证据
来源 · Source
Russell & Norvig,《Artificial Intelligence: A Modern Approach》第4版,Pearson,2020
S. Russell and P. Norvig, Artificial Intelligence: A Modern Approach, 4th ed., Pearson, 2020
独立核验依据 · Independent verification: 核验依据 2 ARCHIVE 标识 isbn_9788177583670：Artificial Intelligence: A
#自主武器 #伦理 #责任
训练 AI 使用的数据涉及哪些隐私问题？
What privacy issues arise from AI training data?
训练数据可能包含个人信息，若未脱敏会泄露隐私，模型甚至可能“记住”并复现敏感样本。成员推断攻击能从模型反推某人是否在训练集中。做法上需要脱敏、授权与差分隐私等技术。
Training data may contain personal information that leaks privacy if not anonymized, and models can even memorize and reproduce sensitive samples. Membership inference can tell whether someone was in the training set. Techniques include anonymization, consent, and differential privacy.
难度 进阶 · 适合 大众
CC BY 4.0 · 已核验 · Verified 2026-08-23 · 2 条独立证据
来源 · Source
Russell & Norvig,《Artificial Intelligence: A Modern Approach》第4版,Pearson,2020
S. Russell and P. Norvig, Artificial Intelligence: A Modern Approach, 4th ed., Pearson, 2020
独立核验依据 · Independent verification: 核验依据 2 ARCHIVE 标识 isbn_9788177583670：Artificial Intelligence: A
#数据隐私 #训练数据 #差分隐私
被AI“遗忘权”为什么重要？
Why is the 'right to be forgotten' by AI important?
个人数据被用于训练后很难撤回，遗忘权让公民有权要求删除，在创新与隐私之间取得平衡。
Once personal data is used for training it is hard to retract; the right to be forgotten lets people demand deletion, balancing innovation and privacy.
难度 进阶 · 适合 大众
CC BY 4.0 · 已核验 · Verified 2026-08-23 · 2 条独立证据
来源 · Source
Russell & Norvig,《Artificial Intelligence: A Modern Approach》第4版,Pearson,2020
S. Russell and P. Norvig, Artificial Intelligence: A Modern Approach, 4th ed., Pearson, 2020
独立核验依据 · Independent verification: 核验依据 2 ARCHIVE 标识 isbn_9788177583670：Artificial Intelligence: A
#遗忘权 #隐私 #数据
生成式 AI 的作品版权归谁？
Who owns the copyright of generative AI output?
多数司法辖区要求作品有“人类作者”，纯 AI 自动生成的内容往往难以获得版权保护。训练数据若未经授权使用他人作品，还可能涉及侵权。相关规则仍在快速演化。
Most jurisdictions require a human author, so purely AI-generated output often cannot be copyrighted. Training on unauthorized works may also raise infringement issues. The rules are still evolving quickly.
难度 进阶 · 适合 大众
CC BY 4.0 · 已核验 · Verified 2026-08-23 · 2 条独立证据
来源 · Source
Russell & Norvig,《Artificial Intelligence: A Modern Approach》第4版,Pearson,2020
S. Russell and P. Norvig, Artificial Intelligence: A Modern Approach, 4th ed., Pearson, 2020
独立核验依据 · Independent verification: 核验依据 2 ARCHIVE 标识 isbn_9788177583670：Artificial Intelligence: A
#版权 #生成式AI #法律
什么是“负责任 AI”？
What is 'responsible AI'?
负责任 AI 是一套原则和实践，要求系统在公平、透明、隐私、安全和可问责等方面达标。它强调从设计到部署全程考虑社会影响。许多公司和监管机构都发布了相应准则。
Responsible AI is a set of principles and practices requiring systems to meet fairness, transparency, privacy, safety, and accountability. It stresses considering social impact from design through deployment. Many companies and regulators have published such guidelines.
难度 进阶 · 适合 大众
CC BY 4.0 · 已核验 · Verified 2026-08-23 · 2 条独立证据
来源 · Source
Russell & Norvig,《Artificial Intelligence: A Modern Approach》第4版,Pearson,2020
S. Russell and P. Norvig, Artificial Intelligence: A Modern Approach, 4th ed., Pearson, 2020
独立核验依据 · Independent verification: 核验依据 2 ARCHIVE 标识 isbn_9788177583670：Artificial Intelligence: A
#负责任AI #治理 #准则
算法推荐怎么把人困在“信息茧房”里？
How do recommendation algorithms trap people in filter bubbles?
为留住用户，系统持续推送相似内容，视野越来越窄、观点越来越极化，削弱了公共讨论。
To keep users engaged, systems keep feeding similar content, narrowing views and polarizing opinions, weakening public discussion.
难度 进阶 · 适合 初中(12-15)
CC BY 4.0 · 已核验 · Verified 2026-08-23 · 2 条独立证据
来源 · Source
Russell & Norvig,《Artificial Intelligence: A Modern Approach》第4版,Pearson,2020
S. Russell and P. Norvig, Artificial Intelligence: A Modern Approach, 4th ed., Pearson, 2020
独立核验依据 · Independent verification: 核验依据 2 ARCHIVE 标识 isbn_9788177583670：Artificial Intelligence: A
#信息茧房 #推荐 #算法
什么是大模型的“幻觉”？
What is a large model's 'hallucination'?
幻觉指模型自信地生成看似合理却错误或无依据的内容，例如编造不存在的论文或事实。它源于模型在预测下一个词而非检索事实。缓解办法包括检索增强和人工核查。
Hallucination is when a model confidently produces plausible but false or unsupported content, like inventing a nonexistent paper. It comes from predicting the next word rather than retrieving facts. Mitigations include retrieval-augmented generation and human review.
难度 进阶 · 适合 大众
CC BY 4.0 · 已核验 · Verified 2026-08-23 · 2 条独立证据
来源 · Source
Russell & Norvig,《Artificial Intelligence: A Modern Approach》第4版,Pearson,2020
S. Russell and P. Norvig, Artificial Intelligence: A Modern Approach, 4th ed., Pearson, 2020
独立核验依据 · Independent verification: 核验依据 2 ARCHIVE 标识 isbn_9788177583670：Artificial Intelligence: A
#幻觉 #大语言模型 #准确性
C2PA 这类「内容凭证」标准能可靠证明一张图片的来龙去脉吗？
Can content-provenance standards such as C2PA reliably prove where an image came from?
C2PA 的做法是给媒体文件附加一份经数字签名的清单，记录拍摄设备、编辑软件和每一步修改，任何后续改动都要追加新的签名断言，形成可验证的链条。它能证明内容自某个签名点以来未被篡改，也能标注是否由 AI 生成，但对完全没有签名的文件无话可说，而截图、转码或社交平台的重压缩都会把凭证剥掉。因此它是提高伪造成本的工具，而不是鉴伪的万灵药，只有相机厂商、编辑软件和分发平台端到端支持才真正有效。
C2PA attaches a cryptographically signed manifest to a media file recording the capture device, editing tools and each modification, with every later change appending another signed assertion to build a verifiable chain. It can prove that content is unaltered since a given signing point and can flag AI generation, but it says nothing about unsigned files, and screenshots, transcoding or platform recompression strip the credentials entirely. So it raises the cost of undetected manipulation rather than detecting fakes, and only works when cameras, editors and distribution platforms all support it end to end.
难度 进阶 · 适合 初中(12-15)
CC BY 4.0 · 来源待核验 · Source pending review
来源 · Source
出处备注：著录仅给书名，无版次与年份；同一书名常对应多个版次，不臆测具体版本。
Coalition for Content Provenance and Authenticity《C2PA Technical Specification, Version 2.1》
Coalition for Content Provenance and Authenticity, "C2PA Technical Specification, Version 2.1"
#C2PA #内容溯源 #数字签名 #AI标识
人脸识别技术带来哪些伦理担忧？
What ethical concerns does face recognition raise?
人脸识别可悄无声息地大规模采集身份信息，容易引发监控过度和隐私侵蚀。它在不同人群上的准确率差异还可能加剧歧视。是否使用需要透明的法律和公众同意。
Face recognition can silently and massively collect identity, risking over-surveillance and privacy erosion. Its uneven accuracy across groups can also worsen discrimination. Its use needs transparent law and public consent.
难度 进阶 · 适合 大众
CC BY 4.0 · 已核验 · Verified 2026-08-23 · 2 条独立证据
来源 · Source
Russell & Norvig,《Artificial Intelligence: A Modern Approach》第4版,Pearson,2020
S. Russell and P. Norvig, Artificial Intelligence: A Modern Approach, 4th ed., Pearson, 2020
独立核验依据 · Independent verification: 核验依据 2 ARCHIVE 标识 isbn_9788177583670：Artificial Intelligence: A
#人脸识别 #隐私 #监控
什么是机器学习中的“模型卡”（model card）？它如何帮助组织落实 AI 的透明度与问责要求？
What is a model card in machine learning, and how does it help organizations fulfill transparency and accountability obligations for AI systems?
模型卡是随训练好的模型一同交付的标准化简短文档，内容包括预期用途与禁用场景、训练数据概况、分人群子集的性能评估、已知局限和伦理风险提示。下游使用者在接入模型前就能查到它在哪些群体上表现欠佳、哪些场景不该用，也为审计与监管提供可核查的书面记录，使事后问责有据可依。
A model card is a standardized short document delivered together with a trained model, covering intended use and out-of-scope uses, an overview of training data, performance evaluated per demographic slice, known limitations, and ethical caveats. Downstream users can check before adoption where the model underperforms, and auditors get a verifiable written record that grounds accountability.
难度 进阶 · 适合 初中(12-15)
CC BY 4.0 · 已核验 · Verified 2026-11-03
来源 · Source
Margaret Mitchell 等，《Model Cards for Model Reporting》，ACM FAT* 会议论文集，2019年，DOI: 10.1145/3287560.3287596
M. Mitchell et al., Model Cards for Model Reporting, Proceedings of ACM FAT* 2019, DOI: 10.1145/3287560.3287596.
独立核验依据 · Independent verification: 核验依据 1 CROSSREF 标识 10.1145/3287560.3287596：Model Cards for Mode
#模型卡 #AI治理 #透明度
算法偏见主要来自哪里？为什么删掉种族、性别等敏感属性并不能让模型变公平？
Where does algorithmic bias mainly come from, and why does dropping sensitive attributes like race or gender fail to make a model fair?
算法偏见主要有三个来源：训练数据记录了历史上的不平等决策，采样偏差让部分群体代表性不足，人工标注又掺入主观判断。即便删去敏感字段，邮编、消费记录等代理变量仍可能还原这些信息，模型照样习得歧视性模式。因此公平治理需要覆盖数据、特征、训练目标与上线后监测的全流程，仅靠字段脱敏远远不够。
Algorithmic bias arises mainly in three places: training data records historically unequal decisions, sampling bias leaves some groups under-represented, and human annotation injects subjective judgment. Even with sensitive fields removed, proxies such as postal codes or purchase histories can reconstruct them, so the model still learns discriminatory patterns. Fairness governance therefore has to span data collection, feature design, training objectives, and post-deployment monitoring; field redaction alone falls far short.
难度 进阶 · 适合 大众
CC BY 4.0 · 已核验 · Verified 2026-11-03
来源 · Source
N. Mehrabi 等，《机器学习中的偏见与公平性综述》，ACM Computing Surveys，2021，DOI: 10.1145/3457607
N. Mehrabi et al., 'A Survey on Bias and Fairness in Machine Learning', ACM Computing Surveys, 2021, DOI: 10.1145/3457607
独立核验依据 · Independent verification: 核验依据 1 CROSSREF 标识 10.1145/3457607：A Survey on Bias and Fairnes
#算法偏见 #代理变量 #公平性 #ML伦理
自动驾驶出事时，责任应该归谁？
When a self-driving car crashes, who is responsible?
责任归属尚无统一答案，可能分散在车主、厂商、软件供应商和监管方之间。关键看事故发生时是由人监控还是系统完全控制。明确的法律与保险框架正在逐步形成。
There is no single answer; responsibility may spread across the owner, maker, software vendor, and regulator. It hinges on whether a human was monitoring or the system was fully in control. Clear legal and insurance frameworks are still emerging.
难度 进阶 · 适合 大众
CC BY 4.0 · 已核验 · Verified 2026-08-23 · 2 条独立证据
来源 · Source
Russell & Norvig,《Artificial Intelligence: A Modern Approach》第4版,Pearson,2020
S. Russell and P. Norvig, Artificial Intelligence: A Modern Approach, 4th ed., Pearson, 2020
独立核验依据 · Independent verification: 核验依据 2 ARCHIVE 标识 isbn_9788177583670：Artificial Intelligence: A
#自动驾驶 #责任 #法律
AI 的偏见是从哪来的？
Where does AI bias come from?
偏见主要来自训练数据：如果历史数据里某种群体被系统性地低估或排除，模型会把这当作规律学走并放大。算法目标本身也可能有偏，比如只优化点击量就天然偏向吸睛内容。技术只能减轻（数据再平衡、公平性约束），要根治还得审视数据采集和社会历史的不公。
Bias mainly comes from training data: if historical data systematically underrepresents or devalues some group, the model learns and amplifies that pattern. Objectives themselves can be biased too—optimizing purely for clicks favors sensational content. Techniques (data rebalancing, fairness constraints) mitigate it, but fixing the root cause means examining data collection and social inequities.
难度 进阶 · 适合 初中(12-15)
CC BY 4.0 · 来源待核验 · Source pending review
来源 · Source
出处备注：著录含书名，但缺版次 / 出版社 / 年份中的一部分，需编辑补全后才能做独立比对。
Buolamwini 与 Gebru《Gender Shades》(Conference on Fairness, Accountability and Transparency, 2018)；NIST 报告 NISTIR 8280《Face Recognition Vendor Test Part 3: Demographic Effects》(2019)
Buolamwini & Gebru, "Gender Shades," Conference on Fairness, Accountability and Transparency (2018); NIST, NISTIR 8280, "Face Recognition Vendor Test Part 3: Demographic Effects" (2019)
#AI偏见 #数据公平 #算法歧视
什么是可解释 AI？
What is explainable AI?
可解释 AI 指让模型的决策过程和依据对人可理解，而不只是给一个结果。在医疗、信贷等高风险场景，解释能帮助发现错误和偏见。常用方法包括特征重要性和局部近似解释。
Explainable AI makes a model's decisions and reasoning understandable to people, not just a final answer. In high-stakes areas like medicine and lending, explanations help catch errors and bias. Common methods include feature importance and local approximations.
难度 进阶 · 适合 大众
CC BY 4.0 · 已核验 · Verified 2026-08-23 · 2 条独立证据
来源 · Source
Russell & Norvig,《Artificial Intelligence: A Modern Approach》第4版,Pearson,2020
S. Russell and P. Norvig, Artificial Intelligence: A Modern Approach, 4th ed., Pearson, 2020
独立核验依据 · Independent verification: 核验依据 2 ARCHIVE 标识 isbn_9788177583670：Artificial Intelligence: A
#可解释性 #透明度 #信任
一个「总体准确率 99%」的人脸识别系统，为什么对不同人群可能极不公平？
Why can a face recognition system with 99% overall accuracy still be grossly unfair across demographic groups?
总体准确率是各子群误差按其在数据中的比例加权的平均值，样本多的群体主导这个数字，样本少的群体即使误差高十倍也几乎不影响总分。Buolamwini 与 Gebru 2018 年的审计发现，商用性别分类在浅肤色男性上错误率不到 1%，在深肤色女性上高达 34.7%；NIST 对上百个算法的大规模评测同样确认，多数系统在不同人种、性别、年龄间的假阳性率相差一到两个数量级。因此评估必须分组报告误差，并区分假阳性与假阴性——手机解锁的假阴性只是不便，而比对检索中的假阳性可能把无关的人牵连进来。
Overall accuracy is an average of per-group error rates weighted by how common each group is in the data, so well-represented groups dominate the figure and a small subgroup can carry ten times the error without moving it. Buolamwini and Gebru's 2018 audit found commercial gender classifiers erring under 1% on lighter-skinned men but 34.7% on darker-skinned women, and NIST's evaluation of over a hundred algorithms confirmed false-positive rates differing by one to two orders of magnitude across race, sex and age. Fair evaluation therefore requires disaggregated error reporting and separating false positives from false negatives — a false negative at phone unlock is an inconvenience, while a false positive in an identification search can implicate the wrong person.
难度 进阶 · 适合 高中(15-18)
CC BY 4.0 · 来源待核验 · Source pending review
来源 · Source
出处备注：著录含书名，但缺版次 / 出版社 / 年份中的一部分，需编辑补全后才能做独立比对。
Buolamwini 与 Gebru《Gender Shades》(Conference on Fairness, Accountability and Transparency, 2018)；NIST 报告 NISTIR 8280《Face Recognition Vendor Test Part 3: Demographic Effects》(2019)
Buolamwini & Gebru, "Gender Shades," Conference on Fairness, Accountability and Transparency (2018); NIST, NISTIR 8280, "Face Recognition Vendor Test Part 3: Demographic Effects" (2019)
#人脸识别 #算法公平性 #分组误差 #假阳性
“深伪”（deepfake）为什么危险？
Why are deepfakes dangerous?
深伪能伪造名人或官员讲话的视频，用来制造谣言、敲诈或干预选举，让“眼见为实”的视频证据不再可靠。
Deepfakes can forge videos of celebrities or officials to spread lies, extort, or sway elections, making video evidence no longer trustworthy.
难度 入门 · 适合 大众
CC BY 4.0 · 已核验 · Verified 2026-08-23 · 2 条独立证据
来源 · Source
Russell & Norvig,《Artificial Intelligence: A Modern Approach》第4版,Pearson,2020
S. Russell and P. Norvig, Artificial Intelligence: A Modern Approach, 4th ed., Pearson, 2020
独立核验依据 · Independent verification: 核验依据 2 ARCHIVE 标识 isbn_9788177583670：Artificial Intelligence: A
#深伪 #虚假 #安全
AI生成内容泛滥会带来什么问题？
What problems come from flooding the web with AI content?
假新闻和水军可以低成本量产，真实信息被稀释，公众的判断成本和被误导的风险都上升了。
Fake news and bots can be mass-produced cheaply, diluting real information and raising the cost and risk of being misled.
难度 入门 · 适合 大众
CC BY 4.0 · 已核验 · Verified 2026-08-23 · 2 条独立证据
来源 · Source
Russell & Norvig,《Artificial Intelligence: A Modern Approach》第4版,Pearson,2020
S. Russell and P. Norvig, Artificial Intelligence: A Modern Approach, 4th ed., Pearson, 2020
独立核验依据 · Independent verification: 核验依据 2 ARCHIVE 标识 isbn_9788177583670：Artificial Intelligence: A
#假信息 #内容 #AI
什么是算法偏见？
What is algorithmic bias?
算法偏见指 AI 系统因训练数据或设计选择而系统性地偏向某些群体、产生不公平结果。例如用人脸数据训练不足时，对某些肤色识别更差。缓解需要从数据和评估两端同时改进。
Algorithmic bias means an AI system systematically favors some groups or yields unfair outcomes due to its training data or design. For example, with insufficient face data it may recognize some skin tones worse. Mitigation needs work on both data and evaluation.
难度 入门 · 适合 大众
CC BY 4.0 · 已核验 · Verified 2026-08-23 · 2 条独立证据
来源 · Source
Russell & Norvig,《Artificial Intelligence: A Modern Approach》第4版,Pearson,2020
S. Russell and P. Norvig, Artificial Intelligence: A Modern Approach, 4th ed., Pearson, 2020
独立核验依据 · Independent verification: 核验依据 2 ARCHIVE 标识 isbn_9788177583670：Artificial Intelligence: A
#算法偏见 #公平 #数据
学生用AI写作业算不算作弊？
Is it cheating for students to use AI for homework?
取决于用途：用AI辅助理解、查资料通常可以，但让AI代写代考会损害学习，学校需要明确规则。
It depends: using AI to understand or research is fine, but having it write or sit exams harms learning, so schools need clear rules.
难度 入门 · 适合 初中(12-15)
CC BY 4.0 · 已核验 · Verified 2026-08-23 · 2 条独立证据
来源 · Source
Russell & Norvig,《Artificial Intelligence: A Modern Approach》第4版,Pearson,2020
S. Russell and P. Norvig, Artificial Intelligence: A Modern Approach, 4th ed., Pearson, 2020
独立核验依据 · Independent verification: 核验依据 2 ARCHIVE 标识 isbn_9788177583670：Artificial Intelligence: A
#教育 #AI #学术
什么是深度伪造（Deepfake）？
What is a deepfake?
深度伪造是用 AI 生成高度逼真的假视频或音频，让人看起来说了或做了从未发生的事。它可被用于诈骗或散布谣言。检测水印和来源认证是应对方向之一。
A deepfake uses AI to fabricate highly realistic video or audio of people saying or doing things they never did. It can enable scams or rumors. Watermarking and provenance authentication are among the countermeasures.
难度 入门 · 适合 大众
CC BY 4.0 · 已核验 · Verified 2026-08-23 · 2 条独立证据
来源 · Source
Russell & Norvig,《Artificial Intelligence: A Modern Approach》第4版,Pearson,2020
S. Russell and P. Norvig, Artificial Intelligence: A Modern Approach, 4th ed., Pearson, 2020
独立核验依据 · Independent verification: 核验依据 2 ARCHIVE 标识 isbn_9788177583670：Artificial Intelligence: A
#深度伪造 #生成式AI #误导
AI会取代哪些人的工作？
Which jobs could AI replace?
重复性、规则明确的任务如客服、翻译、制图更容易被辅助或替代，而需要创造力和共情的工作相对更稳。
Repetitive, rule-based tasks like support, translation, and drawing are easier to assist or replace, while creative and empathetic work is safer.
难度 入门 · 适合 小学低段(6-8)
CC BY 4.0 · 已核验 · Verified 2026-08-23 · 2 条独立证据
来源 · Source
Russell & Norvig,《Artificial Intelligence: A Modern Approach》第4版,Pearson,2020
S. Russell and P. Norvig, Artificial Intelligence: A Modern Approach, 4th ed., Pearson, 2020
独立核验依据 · Independent verification: 核验依据 2 ARCHIVE 标识 isbn_9788177583670：Artificial Intelligence: A
#就业 #自动化 #AI
AI 会取代哪些工作？
Which jobs might AI replace?
AI 更可能替代重复性、规则明确的任务，如数据录入、基础翻译和客服初筛。需要创造、共情和复杂判断的工作较难被取代。多数情况下 AI 是辅助人而非完全替代。
AI is more likely to take over repetitive, rule-based tasks like data entry, basic translation, and first-line support. Jobs needing creativity, empathy, and complex judgment are harder to replace. In most cases AI assists rather than fully replaces people.
难度 入门 · 适合 大众
CC BY 4.0 · 已核验 · Verified 2026-08-23 · 2 条独立证据
来源 · Source
Russell & Norvig,《Artificial Intelligence: A Modern Approach》第4版,Pearson,2020
S. Russell and P. Norvig, Artificial Intelligence: A Modern Approach, 4th ed., Pearson, 2020
独立核验依据 · Independent verification: 核验依据 2 ARCHIVE 标识 isbn_9788177583670：Artificial Intelligence: A
#就业 #自动化 #人机协作