How tech companies, journalists, and policymakers can prevent AI decision-making from going wrong.
AI Reading Assistant
Whole-book reading guide from stratified index samples; jump to passages in the text
Tip the Site
Support this siteYour recognition and a small knowledge-service contribution help keep this technical work open source.Scan the WeChat Pay or Alipay code below. Logged-in and guest visitors can both tip.
WeChat Pay
Alipay
Open WeChat or Alipay and scan. No login required.
AI guide
【One-Line Pitch】
A concise, non-technical field guide to how independent researchers, journalists, companies, and regulators can investigate opaque AI systems—what audits can prove, what they cannot, and how to commission or read one without being misled. Best for decision-makers, civil-society advocates, and curious professionals confronting an AI product for the first time.
【Book Arc】
- **Opening (~0%–10%)**: Frames the core question—how do we know whether an AI system works and is fair?—through historical cases (kidney-allocation software, the SABRE reservation system, Facebook ad targeting) and positions audits as the tool society uses to see behind the screen.
- **Early (~10%–32%)**: Traces auditing's lineage from 1940s fair-housing committees and Civil Rights–era paired testing through correspondence and algorithm audits, then distinguishes independent third-party audits from company-led ones (with Airbnb's civil rights audit as a case).
- **Early–Middle (~32%–48%)**: Walks through the five-step anatomy of a prototypical audit using the "Gender Shades" face-analysis study, covering input design, data collection, and the iterative, trade-off-heavy nature of analysis.
- **Middle (~48%–60%)**: Surveys audit variants beyond the classic input/output test—volunteer and crowdsourced audits, scraping versus APIs, and the practical and ethical costs of automated data collection.
- **Late (excerpts do not cover specifics)**: The book signals later chapters on what happens when audits prompt abandonment of a system, legal protections for auditors' right to scrape, and the costs of auditing, but the excerpts do not detail these.
- **Ending (excerpts do not cover)**: Presumably synthesizes guidance for companies, journalists, and policymakers; the sampled excerpts stop before this.
【Key Takeaways】
- **Audits are a legacy of civil-rights enforcement, not a tech invention** (Early): paired-testing studies of landlords and employers established the template—vary inputs, compare outputs against a legal or ethical standard—that algorithm audits now inherit.
- **Audits are tripwires, not verdicts** (Early): like the HHI in antitrust, an audit flags that a problem may exist and justifies deeper scrutiny; it rarely settles a legal question on its own.
- **"Boring" audits have real value** (Early): audits that find nothing wrong build confidence in safety-critical systems, much like routine egg inspections—passing results are a feature, not a failure.
- **Audits examine behavior, not mechanism** (Early): most focus on what a system does rather than how it works, and full explainability may be impossible—so don't expect audits to answer every technical question.
- **Independent and company-led audits trade off differently** (Early): outsiders offer transparency (including damaging findings); insiders get fuller data access but risk "audit-washing" if details stay private.
- **The five-step audit process is iterative, not push-button** (Middle): good analysis involves repeated statistical work and consultation with affected groups—the kidney-allocation committee only concluded once multiple communities were satisfied.
- **Data collection is legally and ethically fraught** (Middle): scraping, APIs, sock puppets, and personal-data privacy laws (18 US states by 2024) all constrain what auditors may do, and careless automation can degrade or crash the target system.
- **Audit methods vary widely** (Middle): volunteer and crowdsourced audits fill gaps where test inputs are illegal or inaccessible, extending the umbrella well beyond the prototypical design.
【Reading Tips】
- Read the opening cases and the five-step chapter closely—they give you the vocabulary and mental model everything else builds on.
- Skim the historical lineage if you already know the civil-rights audit literature; the payoff is in how it maps onto algorithms.
- Treat the independent-vs-company-led comparison as a decision checklist if you may commission an audit yourself.
- Watch for the "what audits can't do" passages—they are the best defense against overclaiming in meetings or reporting.
- The later legal and cost chapters are flagged but not sampled here; read them in the book if scraping rights or audit budgets matter to your work.
【Coverage Limits】
This guide is based on stratified excerpts covering roughly the first half of the book; later chapters on system abandonment, auditor legal protections, and audit costs are referenced but not detailed, so the Ending is not summarized.
Page 10
dentified disparities across patient populations, measuring differences based on age, gender, geography and race. This work informed decisions that would hel...
of AI led to a shift in terminology, and today, we refer to algorithm audits as AI audits (although the terms are used interchangeably). We use AI audits to...
nal tests of landlords and employers, noting the kinds of 3 The Steps of an AI Audit As part of a graduate course at MIT, Joy Buolamwini and her classmates w...
ames that had strong associations with race and gender (e.g., Monique, Kenji, or Rodrigo). The group then analyzed the outputs (step 4) to check for any diff...
tineraries because it makes the airline more money, as long as the displayed sort order is labeled “recommended,” this technically does not exhibit display b...
anuary 1 and December 31, 2022. An astute reader will note, then, that this audit only gives us insight into the system’s behavior in 2022, not before or sin...
actically, outweigh the benefits. The decision to abandon a product enables leaders to focus their efforts and resources on more promising innovations. If an...
violate their rights—like someone’s race, for instance, in decisions about who is arrested or incarcerated. These are just some examples of how audits are al...
Support this siteYour recognition and a small knowledge-service contribution help keep this technical work open source.
Scan the WeChat Pay or Alipay code below. Logged-in and guest visitors can both tip.
WeChat PayAlipay
Open WeChat or Alipay and scan. No login required.
Add Tag
Enter tag name (max 50 characters)
Share E-Book
Auditing AI ( etc.)(Z-Library)
Scan QR code with your phone to access
Copy the link or scan the QR code to access this e-book on your phone
Share to WeChat Moments
Auditing AI ( etc.)(Z-Library)
Scan the QR code with WeChat to open this book
Tap the ··· menu in the top-right corner
Choose Share to Moments
You are already in WeChat. Tap ··· in the top-right corner and choose Share to Moments.
Share E-Book via Email
Please enter email address
Donation Statistics
¥.00
Total Donations
0
Donation Count
Auditing AI ( etc.)(Z-Library)
Find Your Favorite Books
Only registered users can comment after logging in. Comments need to be reviewed by administrators before being displayed
这本书围绕一个紧迫的问题展开:当人工智能参与决策时,科技公司、记者和政策制定者如何防止它出错。书名中的“审计”提示了核心思路,即把 AI 决策当作需要被检验、被问责的对象,而不是默认其正确。作者群体来自多个领域,讨论因此不局限于技术本身,而是把算法系统放在公司治理、新闻报道与公共政策的交叉点上考察。
读者可以抓住几条线索。其一是“出错”的具体含义:AI 决策可能产生偏差、造成不透明或难以追责的后果,而防止这些后果需要事前的审视与事后的核查,而非事后补救。其二是行动主体的分工:科技公司作为系统的开发与部署者,记者作为外部监督与信息传播者,政策制定者作为规则设定者,三者各自承担不同角色,缺一不可。其三是审计作为一种方法的位置——它既是技术性的检查手段,也是问责机制的一部分,把分散的监督力量连接起来。
全书不提供单一方案,而是呈现多方视角下的共同关切:如何让 AI 决策过程可被理解、可被质疑、可被纠正。对于关心算法问责、技术治理与公共利益的读者,这本书提供的是一个讨论框架,帮助理解不同角色在防止 AI 决策失误中能做什么、应当做什么。
Digital LibraryAI Assistant
Hi — I can help you pick a book. Tell me a topic, language, or level, e.g. “a light English read”. If that's too broad, I'll ask one follow-up.
Loading comments...
Reply to Comment
Edit Comment