English

A computational system determines who can act, what information is available, and how a decision can be corrected. Law and policy define rights, duties, and procedures that may apply to those actions. Technical design affects whether those requirements can be exercised in practice; a written right to correct a record is difficult to use if the system cannot locate the record or identify its downstream copies.

The connection works in both directions. A legal requirement can motivate a technical mechanism, while a technical result can reveal that a proposed requirement is ambiguous or difficult to satisfy. Neither a mathematical guarantee nor a legal label answers every question about acceptable use.

Background: Security and privacy (§4.5), data systems (§4.4), interpretability (§5.6), and professional practice (§8.9). Legal examples must be read in their jurisdiction and context; the focus here is how requirements become design questions.

The Relationship: Code, Law, and the Governance of Computation

Privacy concerns more than hiding a name. Collection, linkage, inference, access, retention, and use can each affect a person. A location history without an explicit name may still reveal a home or workplace. Encryption can protect transmission or storage while leaving an authorized recipient able to inspect every record. State which party should learn which information under which conditions.

Data-protection rules also cover purposes and procedures. Under the EU GDPR, consent is one possible lawful basis among several, not the universal basis for all processing. Purpose limitation, data minimization, and individual rights impose additional requirements. A technical protection does not by itself establish a lawful purpose, and a consent form does not eliminate every other obligation.

Differential privacy gives a specific statistical guarantee about a randomized mechanism’s outputs. For pure ε-differential privacy, changing one person’s contribution under the stated neighboring-dataset definition changes the probability of any output event by at most a multiplicative factor exp(ε). With ε = 0.5, that factor is about 1.65. This bounds how much that person’s participation can affect the released result; it does not mean that the system reveals no population information, nor that ε is a probability of privacy failure.

A count illustrates the mechanism. If each person contributes at most one qualifying record, adding or removing one person changes the exact count by at most one. A suitable randomized release adds noise calibrated to this sensitivity and the privacy parameter. Releasing repeated answers consumes a combined privacy budget; bounding rows is not enough if one person can contribute many rows. The guarantee belongs to the mechanism, contribution bound, and release process together.

Other techniques address other threats. Secure multiparty computation and homomorphic encryption can limit exposure of inputs during specified computations, but outputs may still reveal information. Federated learning keeps raw training data decentralized while exchanging updates; those updates can leak information without additional protections. Choose the mechanism after defining the threat, rather than treating every privacy-related technique as an interchangeable guarantee.

Intellectual Property: Old Frameworks for New Objects

Software can be an expressive work and a functional artifact at the same time. Copyright generally concerns protected expression, including code, rather than an abstract idea or method itself. Patent eligibility, trade-secret protection, and contractual duties address different questions and vary by jurisdiction. Calling something “an algorithm” does not settle which rights or obligations apply.

Open-source licenses grant permissions under stated conditions. Some require preserving notices; copyleft licenses can impose source-sharing obligations when covered software is conveyed or used in specified ways. The exact license and activity matter. Publicly accessible source is not automatically permission to reuse it without conditions, and “open source” does not describe one uniform license.

For generative AI, separate the acquisition of training data, copies made during training, the resulting model, and particular outputs. Each can raise different questions. A model is not simply a directory of source files, but it can memorize and reproduce some training material; training-data extraction experiments demonstrate that “models never store examples” is too strong. Neither memorization in a particular case nor the general use of learning alone decides the legal status of every training run or output.

Copyrightability of an output is also different from permission to use the material that contributed to it. The U.S. Copyright Office’s 2025 report distinguishes human-authored contributions from material generated without sufficient human authorship; AI assistance does not automatically exclude protection for the human contribution. That is a jurisdiction-specific account, not a universal rule for every country. For an actual reuse decision, identify the material, permissions, applicable law, and intended activity rather than relying on a general claim that all AI output is either owned or unprotected.

Algorithmic Accountability: Transparency, Fairness, and Recourse

Accountability requires identifying a decision, the people responsible for it, the evidence used, and a way to investigate and correct errors. A feature-attribution chart may help inspect a prediction, but does not by itself establish that the input was accurate, the objective appropriate, or the decision lawful. A simple model can also be used within an opaque or poorly governed process.

Fairness metrics make different questions precise. Suppose two groups each contain 100 cases, with 20 actual positives in group A and 60 in group B. A classifier with 80% true-positive rate and 10% false-positive rate in each group produces 16 true and eight false positive predictions for A, versus 48 true and four false positives for B. True- and false-positive rates match, but the fraction of positive predictions that are correct is 16/24 in A and 48/52 in B: about 67% versus 92%.

The example shows why equal conditional error rates and equal predictive value are different requirements. Formal incompatibility results need stated assumptions, such as unequal base rates and non-perfect prediction; they do not say that fairness is impossible in every sense. Selecting a metric requires understanding the decision, how labels were obtained, the costs of errors, and the people affected. Meeting one metric is evidence about that metric, not a complete moral or legal judgment.

Recourse requires a working process. Preserve enough version and decision information to investigate a disputed result, subject to appropriate retention limits. Make it possible to correct relevant inputs, identify a responsible reviewer, and communicate the outcome. Review must have authority to change a decision when appropriate; merely displaying an explanation or adding an approval button may not provide meaningful correction.

Legal requirements depend on the kind of decision and applicable rules. GDPR protections concerning solely automated decisions with legal or similarly significant effects have conditions and exceptions; they are not a blanket ban on automation. Translate the applicable requirement into a concrete workflow and verify that people can actually use it. Technical interpretability, procedural accountability, and legal compliance overlap but are not synonyms.

What This Perspective Changes

For a system that handles personal data or consequential decisions, connect each requirement to a mechanism and a procedure. Identify what the mechanism guarantees, who can invoke the procedure, and what evidence shows that it works.

As a reading exercise, take one privacy or fairness claim and identify its unit of protection, assumptions, and limits. Then ask which remaining questions concern purpose, authority, or remedies. This keeps mathematical precision useful without asking it to make decisions that require legal and social judgment.

Resources

Foundational and Bridging Texts

Lessig and Solove explain how architectures and information practices affect governance and privacy. Dwork and Roth’s differential-privacy text defines the technical guarantee and its composition. Barocas, Hardt, and Narayanan develop fairness criteria and their relation to social questions; the Kleinberg–Mullainathan–Raghavan paper is a source for conditional incompatibility results. O’Neil offers cases for examining objectives, data, and institutional consequences.

Resource Role Type
Lessig, Code: And Other Laws of Cyberspace (v2.0) How technical architecture shapes governance Depth
Solove, Understanding Privacy Different privacy interests and information practices Depth
Dwork & Roth, The Algorithmic Foundations of Differential Privacy (free) Definitions, mechanisms, and composition guarantees Depth
Barocas, Hardt & Narayanan, Fairness and Machine Learning (free) Fairness criteria, assumptions, and social context Depth
O’Neil, Weapons of Math Destruction Cases involving data-driven institutional decisions Entry

Intellectual Property, AI, and Ongoing Developments

Use primary legal texts with official explanations for the relevant jurisdiction. The EDPB explains lawful bases; its automated-decision guidance addresses the scope and safeguards of Article 22. The U.S. Copyright Office’s 2025 copyrightability announcement and report concern human authorship and AI assistance. Carlini and colleagues’ training-data extraction study supports the technical memorization distinction. Risk frameworks and policy analyses provide additional methods and perspectives; they should not be mistaken for binding rules everywhere.

Resource Role Type
Lessig, Free Culture (free) Arguments about copyright and cultural participation Depth
GDPR and major privacy regimes (primary text) Jurisdiction-specific duties and individual rights Reference
EU AI Act official text EU rules; check the applicable provisions and dates Reference
U.S. Copyright Office, Copyright and Artificial Intelligence reports (free) U.S. analysis of AI-related copyright questions Reference
NIST AI Risk Management Framework (free) Methods for identifying and managing AI risks Practice
OECD AI Principles (free) International policy principles for AI Reference
AI-and-copyright litigation and policy (ongoing) Case-specific decisions and policy proposals Reference
Technology policy research (AI Now and others) Research perspectives on technology governance Reference

Traps

Trap Why it misleads Better response
Equating privacy with consent or encryption Purpose, access, inference, and other duties remain relevant. Identify the lawful basis, threat model, and full data lifecycle.
Treating differential privacy as no information about anyone The guarantee concerns changes between specified neighboring datasets. State the contribution bound, parameter, and composition of releases.
Using a general AI claim to decide copyright Training, human contributions, outputs, permissions, and jurisdictions differ. Analyze the particular material and activity under applicable rules.
Equating an explanation or fairness score with accountability Neither alone establishes appropriate use or a working remedy. Combine evidence, responsible review, and a usable correction process.

中文

计算系统决定谁能够行动、哪些信息可用，以及怎样纠正决定。法律与政策规定可能适用的权利、义务和程序。技术设计影响这些要求能否实际行使：如果系统无法定位记录或识别下游副本，书面规定的记录更正权就很难使用。

这种联系是双向的。法律要求可以促成技术机制，技术结果也可以揭示某项要求含糊或难以满足。数学保证或法律标签，都不能单独回答关于适当使用的全部问题。

背景知识：安全与隐私（§4.5）、数据系统（§4.4）、可解释性（§5.6）与职业实践（§8.9）。法律例子需要结合司法辖区与具体情境阅读；本节关注要求怎样转化为设计问题。

这种关系：代码、法律与计算治理

隐私：在技术观念与法律观念之间被争夺的价值

隐私不只涉及隐藏姓名。收集、关联、推断、访问、保留和使用，都可能影响一个人。没有姓名的位置历史，仍可能暴露住址或工作地点。加密可以保护传输或存储，但获得授权的接收者仍可能查看每条记录。应说明在哪些条件下，哪一方应该知道哪些信息。

数据保护规则还涉及目的与程序。在欧盟 GDPR 下，同意是多种可能合法依据之一，不是所有处理活动的统一依据。目的限制、数据最小化与个人权利，也提出额外要求。技术保护本身不能证明目的合法，一份同意表也不能消除其他义务。

差分隐私对随机机制的输出提供特定统计保证。对于纯 ε 差分隐私，在所声明的相邻数据集定义下，改变一个人的贡献，会使任意输出事件的概率至多变化一个乘法因子 exp(ε)。当 ε = 0.5 时，该因子约为 1.65。它限制这个人的参与对发布结果的影响，不表示系统完全不能揭示总体信息，也不表示 ε 是隐私失败概率。

计数可以说明机制。如果每人至多贡献一条符合条件的记录，增加或删除一个人，就使精确计数至多变化一。合适的随机发布机制，会根据这一敏感度与隐私参数添加噪声。重复发布答案需要合并计算隐私预算；如果一个人可以贡献许多行，仅限制单行影响还不够。保证属于机制、贡献上限与发布过程的整体。

其他方法处理不同威胁。安全多方计算与同态加密，可以限制指定计算过程中的输入暴露，但输出仍可能泄露信息。联邦学习让原始训练数据保持分散，同时交换更新；没有额外保护时，更新也可能泄露信息。应先定义威胁再选择机制，不能把所有与隐私有关的技术当作可以互换的保证。

知识产权：用旧框架处理新对象

软件可以同时是表达性作品与功能性产物。著作权通常保护包括代码在内的特定表达，而不是抽象思想或方法本身。专利资格、商业秘密保护和合同义务，处理的是不同问题，并且因司法辖区而异。把某个东西称为“算法”，不能解决适用哪些权利或义务的问题。

开源许可证在规定条件下授予权限。有些要求保留声明；copyleft 许可证可能在分发所覆盖软件或特定使用方式下，要求提供源代码。具体许可证与活动很重要。源代码公开可见，不自动表示可以无条件复用；“开源”也不是一种统一许可证。

对于生成式 AI，应区分训练数据的获取、训练过程中产生的复制、模型本身，以及具体输出。各环节可能提出不同问题。模型并不只是源文件目录，但可能记忆并复现部分训练材料；训练数据提取实验说明，“模型从不存储例子”过于绝对。某个案例中的记忆现象，或采用学习方法这一事实，都不能单独决定所有训练与输出的法律性质。

输出能否取得著作权，也不同于是否有权使用对其产生贡献的材料。美国版权局 2025 年报告区分了人类创作贡献与缺少足够人类作者性的生成内容；使用 AI 辅助，不会自动排除对人类贡献的保护。这是特定司法辖区的说明，不是各国通用规则。实际判断复用时，应明确材料、权限、适用法律和准备进行的活动，而不是依赖“所有 AI 输出都有人拥有”或“全都不受保护”的笼统说法。

算法问责：透明度、公平性与救济

问责需要明确决定、责任人、所用证据，以及调查和纠错的方法。特征归因图可能帮助检查预测，但不能单独证明输入准确、目标合适或决定合法。简单模型也可能被用于不透明、缺少合理治理的流程。

公平性指标使不同问题变得精确。假设两组各有 100 个案例，A 组实际正例为 20 个，B 组为 60 个。分类器在每组中的真正例率均为 80%，假正例率均为 10%。那么 A 组产生 16 个真正例预测和八个假正例，B 组则是 48 个真正例预测和四个假正例。真正例率与假正例率相同，但正预测中正确者的比例，A 组为 16/24，B 组为 48/52，约为 67% 与 92%。

例子说明，条件错误率相等与阳性预测值相等，是不同要求。形式上的不兼容结果需要明确假设，例如基准发生率不同、预测并不完美；它们不能证明一切意义的公平都不可能。选择指标需要理解决定、标签怎样取得、错误成本与受影响的人。满足一个指标，是关于该指标的证据，并不是完整的道德或法律判断。

救济需要可用流程。在适当保留期限下，保存足够的版本与决定信息，以调查争议结果。允许更正相关输入、找到负责的审查者，并告知处理结果。审查需要在适当情况下真正改变决定的权限；仅展示解释或增加一个批准按钮，未必提供有意义的纠错。

法律要求取决于决定类型与适用规则。GDPR 对产生法律效果或类似重大影响的纯自动化决定的保护，具有条件与例外，并不是全面禁止自动化。应把适用要求转化为具体流程，再验证人们确实能够使用。技术可解释性、程序问责与法律合规相互重叠，但不是同义词。

这种视角会改变什么

对于处理个人数据或重要决定的系统，把每项要求联系到机制与程序。明确机制保证什么，谁能够启动程序，以及什么证据能说明它确实有效。

可以选择一项隐私或公平性结论，找出保护单位、假设与限制，再判断剩余哪些问题涉及目的、权限或救济。这样既保留数学精确性的价值，也不要求数学替代法律与社会判断。

资源

基础与桥接文本

Lessig 与 Solove 解释架构和信息实践怎样影响治理与隐私。Dwork 和 Roth 的差分隐私教材定义技术保证及其组合。Barocas、Hardt 和 Narayanan 发展公平性标准，并联系社会问题；Kleinberg–Mullainathan–Raghavan 论文提供有条件的不兼容结果。O’Neil 的案例有助于检查目标、数据与制度后果。

资源 作用 类型
Lessig，Code: And Other Laws of Cyberspace（v2.0） 技术架构怎样影响治理 深入
Solove，Understanding Privacy 不同隐私利益与信息实践 深入
Dwork & Roth，The Algorithmic Foundations of Differential Privacy（免费） 定义、机制与组合保证 深入
Barocas、Hardt & Narayanan，Fairness and Machine Learning（免费） 公平性标准、假设与社会背景 深入
O’Neil，Weapons of Math Destruction 涉及数据驱动制度决策的案例 入门

知识产权、AI 与持续发展

应结合相关司法辖区的法律原文与官方解释阅读。EDPB 解释了合法依据，其自动化决定指南讨论第 22 条的范围与保障。美国版权局 2025 年著作权资格公告与报告涉及人类作者性与 AI 辅助。Carlini 等人的训练数据提取研究支持前文对记忆现象的技术区分。风险框架与政策分析提供其他方法和视角，不应被当作处处适用的强制规则。

资源 作用 类型
Lessig，Free Culture（免费） 有关著作权与文化参与的论证 深入
GDPR 和主要隐私制度（原始文本） 特定辖区的义务与个人权利 参考
欧盟 AI Act 官方文本 欧盟规则，需核对适用条款与日期 参考
U.S. Copyright Office，Copyright and Artificial Intelligence reports（免费） 美国关于 AI 著作权问题的分析 参考
NIST AI Risk Management Framework（免费） 识别与管理 AI 风险的方法 实践
OECD AI Principles（免费） AI 国际政策原则 参考
AI 与版权诉讼及政策（持续发展中） 具体案件决定与政策建议 参考
技术政策研究（AI Now 等） 技术治理的研究视角 参考

陷阱

误区 为什么会误导 更好的做法
把隐私等同于同意或加密 目的、访问、推断与其他义务仍然相关。 明确合法依据、威胁模型与完整数据生命周期。
把差分隐私当作不能揭示任何人的任何信息 保证涉及指定相邻数据集之间的变化。 明确贡献上限、参数与多次发布的组合。
用一般 AI 说法决定著作权 训练、人类贡献、输出、权限与辖区各不相同。 根据适用规则分析具体材料与活动。
把解释或公平性分数等同于问责 两者都不能单独证明使用适当或救济有效。 结合证据、负责的审查与可用纠错流程。