Chapter 8 Exercises: When AI Gets It Wrong — Errors, Hallucinations, and Failures

Part A: Foundational (Understanding)

Exercise 8.1 — Failure Type Classification For each of the following AI failures, identify which type from the taxonomy in Section 8.1 it represents (Type 1: wrong answer; Type 2: hallucination; Type 3: distributional shift; Type 4: adversarial; Type 5: cascading). Briefly explain your reasoning.

a) A chatbot asked about a historical event provides a detailed, confident narrative that contains multiple factual errors, including inventing a treaty that never existed.

b) A self-driving car trained in Phoenix, Arizona, encounters heavy fog in San Francisco and mistakes a parked motorcycle for a pedestrian.

c) A facial recognition system at an airport fails to match a passenger to their passport photo because the passenger recently shaved their beard.

d) A researcher adds imperceptible pixel-level noise to an image of a panda, causing an image classifier to identify it as a gibbon with 99% confidence.

e) An AI transcription error changes "not allergic to penicillin" to "allergic to penicillin" in a medical record. A prescribing AI reads the record and avoids prescribing penicillin for an infection where it would be the best treatment. The patient receives a less effective antibiotic and has a slower recovery.

f) A sentiment analysis tool trained on product reviews from 2015-2019 performs poorly on reviews written after 2020, because pandemic-era language patterns (references to lockdowns, shipping delays, working from home) weren't in the training data.

Exercise 8.2 — Confidence vs. Accuracy Scenarios For each scenario, explain whether the confidence score is being used appropriately or inappropriately, and why.

a) A weather app says "85% chance of rain." You bring an umbrella but don't cancel your outdoor plans.

b) An AI grading tool says it is "94% confident" that a student's essay is AI-generated. The instructor, based on this score alone, reports the student for academic dishonesty.

c) A diagnostic AI reports "72% confidence: benign, 28% confidence: malignant" for a skin lesion image. The dermatologist orders a biopsy to confirm.

d) A hiring AI rates a candidate as "88% match" for a position. The recruiter decides not to interview anyone rated below 85%.

Exercise 8.3 — Vocabulary Practice Define each term in your own words and provide an original example for each:

a) Hallucination (in the AI context) b) Distributional shift c) Calibration (of confidence scores) d) Automation bias e) Cascading failure f) Graceful degradation

Part B: Intermediate (Apply / Analyze)

Exercise 8.4 — The Hallucination Hunt Use a publicly available AI chatbot or language model to complete the following exercise. (If you don't have access, discuss the exercise with a partner who does, or use the sample outputs provided by your instructor.)

Ask the AI to: a) Provide five specific citations (author, title, journal, year) on a topic you know well b) Describe a historical event from a period you've studied c) Explain a concept from a field you're knowledgeable in

For each response: 1. Fact-check every specific claim against independent sources 2. Note which claims are accurate, which are inaccurate, and which you couldn't verify 3. Rate the overall accuracy on a scale of 1-5 4. Assess whether the output sounded more confident than its accuracy warranted 5. Describe any patterns you notice in the types of errors the AI made

Write a one-page reflection on what this exercise taught you about AI reliability and your own susceptibility to automation bias.

Exercise 8.5 — Distributional Shift Analysis Choose one of the following AI systems and identify at least three ways distributional shift could cause it to fail:

a) A crop disease detection app trained on images from large commercial farms in the American Midwest, now being deployed by smallholder farmers in sub-Saharan Africa

b) A customer service chatbot trained on pre-2020 customer inquiries, now operating during a major product recall

c) A credit scoring model trained on data from 2015-2019, now evaluating applicants in 2024 after a major recession

d) A speech recognition system trained primarily on American English speakers, now deployed in a multinational call center

For each distributional shift you identify: 1. Describe specifically what differs between training and deployment 2. Predict what type of errors would result 3. Propose a mitigation strategy

Exercise 8.6 — Cascading Failure Mapping Design a plausible cascading failure scenario involving at least four AI systems in one of the following domains:

a) Criminal justice (from arrest to sentencing) b) Financial services (from application to account management) c) Education (from admissions to graduation) d) Healthcare (from intake to treatment)

For your scenario: 1. Diagram the chain of systems, showing data flows between them 2. Introduce a single initial error and trace its propagation 3. Identify the point(s) where a human could intervene to break the cascade 4. Explain why, in practice, the human intervention might not happen (consider automation bias, time pressure, information opacity) 5. Propose a system design change that would make the cascade less likely

Part C: Advanced (Analyze / Evaluate)

Exercise 8.7 — The Verification Challenge Apply the VERIFY framework (from Section 8.6) to evaluate the following AI-generated output:

"According to a 2023 study published in the Annual Review of Public Health by researchers at Johns Hopkins University, approximately 37% of diagnostic errors in U.S. emergency departments involve AI-assisted tools, with the highest error rates occurring in cardiology (42%) and radiology (39%). The study, led by Dr. Anita Ramirez, recommended mandatory human review for all AI-generated diagnoses in emergency settings."

Work through each step of VERIFY: - V: What kind of source would produce this claim? - E: What level of specificity and confidence does this output express? - R: Does anything about this claim seem suspicious or too precise? - I: Try to independently verify any element of this claim (the journal, the institution, the researcher, the statistics) - F: If this claim appeared in a policy document, how much would its accuracy matter? - Y: Who would you consult to verify a claim like this?

Write up your verification process and findings.

Exercise 8.8 — The Stakes Matrix Create a matrix analyzing AI errors across two dimensions: (1) the severity of consequences (low, medium, high, catastrophic) and (2) the frequency of errors (rare, occasional, frequent). Place at least two real-world AI applications in each severity category.

For each application you place: - Identify the most likely failure type - Assess whether current verification practices are adequate - Propose what level of human oversight is appropriate

Then answer: Is it possible that a system with frequent but low-severity errors could be more dangerous than a system with rare but high-severity errors? Under what conditions?

Exercise 8.9 — Perspective-Taking: Automation Bias Consider the following scenario:

Dr. Sarah Kim is an emergency medicine physician. She's 14 hours into a 16-hour shift. MedAssist AI flags an X-ray as "96% likely: standard rib fracture." The patient is complaining of severe chest pain, but the vitals are stable. Dr. Kim reviews the X-ray briefly, agrees with MedAssist's assessment, prescribes pain medication, and discharges the patient. Twelve hours later, the patient returns in critical condition with a pneumothorax (collapsed lung) that was visible on the original X-ray but subtle enough that MedAssist's classification missed it.

Write three responses to this scenario, each approximately 200 words: a) From Dr. Kim's perspective — what factors influenced her decision? Is she at fault? b) From the hospital administrator's perspective — what systemic factors contributed? What changes should be made? c) From the patient's perspective — who is responsible for what happened?

Then write a 200-word synthesis: Given the realities of automation bias, physician fatigue, and AI overconfidence, what is the right way to integrate AI diagnostic tools into emergency medicine?

Part D: Research Extension (for Deep Dive readers)

Exercise 8.10 — The Hallucination Reduction Challenge Research two current approaches to reducing hallucinations in large language models:

a) Retrieval-Augmented Generation (RAG) b) One other approach of your choice (e.g., Constitutional AI, chain-of-thought prompting, self-consistency checking)

For each approach: 1. Explain how it works in plain language 2. Describe its strengths and limitations 3. Provide an example of a hallucination it would catch and one it might miss 4. Assess whether it truly solves the hallucination problem or merely reduces its frequency

Write a 750-word comparative analysis concluding with your assessment of whether hallucinations can ever be fully eliminated from language models, and what this means for how we should use them.

Exercise 8.11 — Case Study Deep Dive Research one of the following real-world AI failures in depth:

a) The 2018 Uber self-driving car fatality in Tempe, Arizona b) The IBM Watson for Oncology controversy (2017-2018) c) The Amazon hiring AI that discriminated against women (2018) d) The DALL-E/Midjourney generation of nonconsensual deepfakes

For your chosen case: 1. Describe what happened, using primary sources where possible 2. Classify the failure using the taxonomy from Section 8.1 3. Identify what verification or oversight could have prevented the harm 4. Analyze the accountability question: who was responsible, and who bore the consequences? 5. Connect the case to at least two concepts from this chapter