Chapter 4 Exercises
Part A: Recall and Comprehension
A1. Define the following terms in your own words, then give one example of each that was not used in the chapter: - (a) Structured data - (b) Unstructured data - (c) Labeled data - (d) Ground truth
A2. List the six stages of the data supply chain described in Section 4.1. For each stage, identify one decision that a human makes and explain how that decision could introduce bias.
A3. Explain the difference between selection bias and measurement bias. Why is it important to distinguish between them when auditing an AI system?
A4. What is data provenance, and why do AI researchers advocate for "datasheets for datasets"? Describe at least three pieces of information a datasheet should include.
A5. In your own words, explain the threshold concept for this chapter: "Data is never neutral — it encodes the world that created it." Give a concrete example that illustrates this idea.
Part B: Analysis and Application
B1. ContentGuard Data Audit ContentGuard moderates content across 47 countries in 12 languages. Analyze the data challenges this creates: - (a) What types of bias are most likely when content moderators from one culture label content created in another? - (b) Sarcasm, slang, and coded language vary enormously across cultures. How might this affect label quality? - (c) Propose two concrete steps ContentGuard could take to improve the representativeness of its training data across cultures.
B2. MedAssist AI — Missing Data Analysis MedAssist AI is trained on medical records from three large urban hospitals. Identify at least four groups or populations whose health data might be underrepresented in this training set. For each group, explain: - (a) Why they might be missing (the mechanism of absence) - (b) One potential consequence of their absence for the AI's diagnostic performance
B3. Proxy Variable Problem A bank trains a loan approval AI on ten years of lending data. They remove race and ethnicity from the training data to prevent discrimination. - (a) List three variables that might serve as proxies for race in lending data. - (b) Explain why removing the explicit variable is insufficient. - (c) Propose an alternative approach the bank could take to address this problem.
B4. Feedback Loop Mapping Draw (or describe in words) a feedback loop diagram for one of the following scenarios. Identify the point at which bias enters and explain how the loop amplifies it over time: - (a) A music recommendation algorithm that tracks what songs users play to completion - (b) A hiring AI that learns from which candidates a company has historically offered jobs to - (c) A healthcare triage system that prioritizes patients based on historical data about who was admitted to the ICU
Part C: Practical Skills
C1. Dataset Audit Checklist Using the concepts from this chapter, create a 10-item checklist you could use to evaluate any dataset before using it to train an AI system. Each item should be phrased as a question, and you should indicate what a "red flag" answer would look like. Your checklist should cover: provenance, representativeness, labeling quality, consent, and known limitations.
C2. Label Disagreement Exercise Below are five social media posts. Without consulting anyone else, classify each as: (a) Hate speech, (b) Offensive but not hate speech, (c) Acceptable, or (d) Cannot determine without more context.
- "People from [country X] are ruining this city. Go back where you came from."
- "I swear my roommate is the laziest human being on the planet. #CantEven"
- "This group is a cancer on society and should be eliminated." [posted in a gaming forum about an in-game faction]
- "All [political party] voters are idiots. I'm not even sorry."
- "My grandmother used to call this recipe [ethnic slur] stew. Making it tonight! #ThrowbackThursday"
After classifying, reflect: - (a) Which posts were hardest to classify? Why? - (b) If you compared your classifications with a classmate's, where do you think you would disagree? - (c) What does this exercise tell you about the challenge of creating "objective" labels for content moderation AI?
C3. Data Source Investigation Choose one AI product you use regularly (e.g., a voice assistant, recommendation system, search engine, translation tool). Research and document: - (a) What training data the system uses (if publicly documented) - (b) What you can infer about its data sources based on its behavior (strengths and weaknesses) - (c) What you cannot find out about its data, and what this opacity means for accountability
Part D: Critical Thinking and Perspective
D1. The Consent Dilemma Consider two positions:
Position A: "By posting content on a public social media platform, users implicitly consent to that content being used for any purpose, including AI training. Public data is public."
Position B: "Consent must be specific and informed. Posting a vacation photo for friends is not consent to have your face included in a facial recognition training set."
Write a 300-word argument for the position you disagree with more. Then write a 100-word reflection on what you learned from arguing against your own view.
D2. The Ghost Data Problem Identify a population or situation that is likely underrepresented in each of the following data types. Explain why the absence matters and propose one approach to address it: - (a) Online text data used to train language models - (b) Medical imaging datasets - (c) Autonomous vehicle training data - (d) Voice assistant training data
D3. Labor Ethics Debate Some argue that data labeling work provides valuable employment opportunities in countries with limited economic options, and that criticizing these jobs risks taking away income from people who need it. Others argue that the low wages and harmful working conditions constitute exploitation. - (a) Identify the strongest argument on each side. - (b) Propose a set of minimum standards for data labeling work that balances economic opportunity with worker protection. - (c) Who should be responsible for enforcing these standards — the AI companies, the outsourcing firms, governments, or some combination?
D4. Perspective Swap You are a product manager at a healthcare AI company. Your engineers report that your diagnostic model performs 95% accuracy overall, but only 78% accuracy for patients over 80 years old, because elderly patients are underrepresented in your training data. Collecting more data from this population would cost $2 million and delay launch by six months. What do you do? Write your reasoning in 200-300 words, considering: - Business pressures - Ethical obligations - Legal requirements - Who is affected by the accuracy gap
Part M: Metacognitive Reflection
M1. Before reading this chapter, how did you think about "data"? Has your understanding changed? Write 3-4 sentences describing the most significant shift in your thinking.
M2. Which type of data bias (selection, historical, measurement, aggregation, ghost data) do you think you would be most likely to overlook if you were building an AI system? Why? What could you do to guard against this blind spot?
M3. The chapter argues that "data is never neutral." Do you find this claim persuasive? Is there a type of data that you think is genuinely neutral? Explain your reasoning.
Part E: Extension and Research
E1. Research Deep Dive Research the ImageNet dataset in more detail. Write a 500-word analysis covering: - (a) How ImageNet was created and why it became so influential - (b) The controversies surrounding its person categories - (c) What changes have been made in response to criticism - (d) What lessons ImageNet's story offers for future dataset creation
E2. Data Regulation Comparison Compare how data collection for AI is regulated in three different jurisdictions (e.g., EU under GDPR, United States, China, Brazil under LGPD). Create a comparison table and write a 300-word analysis of which approach you find most effective and why.
E3. Python Extension (Optional) Extend the Python example from the chapter. Load the UCI Adult Income dataset and answer the following questions with code: - (a) What percentage of the dataset is female? How does income distribution differ by sex? - (b) Create a cross-tabulation of race and income. What patterns do you observe? - (c) What is the average age for each education level? Are any education levels underrepresented? - (d) Write a short paragraph interpreting your findings through the lens of data bias.