Chapter 7 Quiz: AI Decision-Making
Test your understanding of the concepts from this chapter. Try answering each question before checking the explanation.
Question 1. An AI system at a bank examines a loan application and assigns it a label of "approve" or "deny." What decision mode is this system primarily using?
a) Recommendation b) Classification c) Prediction d) All of the above equally
Answer
**b) Classification.** The system is sorting loan applications into discrete categories (approve/deny). While the system likely uses prediction internally (estimating the probability of default), the final output is a categorical label — a classification. A recommendation system would suggest different loan products rather than making a binary decision.Question 2. A streaming music platform creates a personalized playlist for you based on what other users with similar listening histories enjoy. Which technique is this platform primarily using?
a) Content-based filtering b) Collaborative filtering c) Proxy variable analysis d) Classification
Answer
**b) Collaborative filtering.** The key signal here is "what other users with similar listening histories enjoy." Collaborative filtering identifies users with overlapping preferences and recommends items liked by similar users. Content-based filtering would focus on the attributes of the songs themselves (genre, tempo, instrumentation) rather than on other users' behavior.Question 3. An AI system correctly identifies 95% of fraudulent transactions. However, it also flags 10% of legitimate transactions as fraudulent. The 10% of legitimate transactions incorrectly flagged are examples of:
a) False negatives b) True positives c) False positives d) True negatives
Answer
**c) False positives.** A false positive occurs when the system says "yes" (this is fraud) but the reality is "no" (the transaction is legitimate). These are "false alarms." The 5% of fraudulent transactions the system misses would be false negatives — fraud that slips through undetected.Question 4. Which of the following best describes the accuracy-interpretability trade-off?
a) More accurate systems are always better, regardless of interpretability b) Interpretable systems are always preferred because they can be audited c) In general, more complex and accurate models tend to be harder for humans to understand, creating tension in high-stakes decisions d) Accuracy and interpretability always decrease together
Answer
**c)** The accuracy-interpretability trade-off describes the general pattern where more complex models achieve higher accuracy but are harder to interpret. This creates real tension in contexts where both accuracy and explainability matter — like medical diagnosis, criminal justice, or lending decisions. Options (a) and (b) are absolutist positions that ignore context; option (d) incorrectly states they decrease together rather than being in tension.Question 5. A predictive policing system directs more police patrols to Neighborhood X, leading to more arrests there. The arrest data is then used to retrain the system, which predicts even higher crime in Neighborhood X. This is an example of:
a) Collaborative filtering b) A feedback loop c) Content-based filtering d) The accuracy-interpretability trade-off
Answer
**b) A feedback loop.** The system's output (directing patrols to Neighborhood X) influences the data used to retrain it (more arrests in Neighborhood X), reinforcing the original prediction. This is a vicious feedback loop because it amplifies the pattern regardless of whether the underlying reality (actual crime rates) supports it. The additional arrests are an artifact of increased surveillance, not necessarily increased crime.Question 6. A hiring AI system doesn't use "gender" as an input variable, but it does use "years of uninterrupted work history" — which correlates strongly with gender because women are more likely to take parental leave. In this context, "years of uninterrupted work history" is best described as:
a) A training label b) A proxy variable c) A false positive d) A threshold concept
Answer
**b) A proxy variable.** A proxy variable is a data point that is technically neutral but correlates strongly with a protected characteristic (in this case, gender). Even though the system doesn't directly use gender, the proxy variable effectively smuggles gender-based discrimination into the model's decisions. This is one of the most challenging problems in AI fairness — simply removing protected variables doesn't prevent discrimination if correlated proxies remain.Question 7. An AI system predicts that students from low-income zip codes are more likely to drop out of college. A university administrator proposes cutting admission rates for students from those zip codes. What critical distinction is the administrator failing to make?
a) The distinction between recommendation and classification b) The distinction between accuracy and interpretability c) The distinction between prediction and explanation d) The distinction between false positives and false negatives
Answer
**c) The distinction between prediction and explanation.** The system predicts *that* students from certain zip codes are more likely to drop out, but it doesn't explain *why*. The causes might include financial hardship, lack of family support, underprepared K-12 schools, or institutional failures — all things the university could address. Cutting admission rates doesn't address causes; it punishes students for circumstances beyond their control. Good prediction should lead to supportive intervention, not exclusion.Question 8. Which of the following is NOT a strategy for breaking a feedback loop?
a) Introducing randomization into decisions to gather unbiased data b) Incorporating data from sources not influenced by the system's decisions c) Increasing the system's accuracy on existing data d) Having humans periodically override algorithmic decisions
Answer
**c) Increasing the system's accuracy on existing data.** If the existing data is itself a product of the feedback loop, training a more accurate model on that data will simply reinforce the loop more efficiently. Breaking a feedback loop requires disrupting the connection between the system's output and its input data — through randomization (a), external data (b), or human intervention (d). This is a crucial and counterintuitive point: a "better" model trained on distorted data can produce *worse* real-world outcomes.Question 9. ContentGuard's content moderation system must decide whether user posts violate platform policies. If the system is tuned to be very aggressive about removing potentially harmful content, which of the following consequences is most likely?
a) More false negatives (harmful content left up) b) More false positives (legitimate content removed) c) Fewer errors overall d) Better interpretability
Answer
**b) More false positives (legitimate content removed).** An aggressive system casts a wider net, catching more harmful content (reducing false negatives) but also sweeping up more legitimate content (increasing false positives). This is a direct consequence of the false positive / false negative trade-off: you can't reduce one type of error without increasing the other. This is why content moderation is so contentious — users who experience false positives feel censored, while users who experience false negatives feel unprotected.Question 10. The chapter's threshold concept states: "AI decisions are probability estimates, not truths." Which of the following scenarios best illustrates a violation of this principle?
a) A doctor uses an AI system's 75% probability estimate as one input among many in making a diagnosis b) A judge sentences a defendant to a longer prison term because an AI system assigned a "high risk" score of 0.82 c) A retail company uses purchase prediction to stock inventory, accepting that some predictions will be wrong d) A student checks an AI-generated essay for errors before submitting it