Chapter 2: Assessment Quiz

Chapter 2 | AI Ethics for Business Professionals

Total Questions: 20 | Estimated Time: 45–60 minutes Formats: 8 Multiple Choice | 5 True/False | 4 Short Answer | 3 Applied Scenario

Instructions: Complete all questions. For short answer and applied scenario questions, aim for the word count range indicated. The answer key is provided at the end of this document.

Part I: Multiple Choice

Select the single best answer for each question.

Question 1 Norbert Wiener's The Human Use of Human Beings (1950) is considered a foundational AI ethics text primarily because it:

A) Proposed the first algorithmic test for machine intelligence B) Argued that feedback systems could replace human judgment and that the social context of deployment determined whether that was beneficial or harmful C) Demonstrated mathematically that AI systems could not achieve genuine understanding D) Was the first text to document racial bias in automated decision-making systems

Question 2 The 1956 Dartmouth Summer Research Project is historically significant in AI ethics primarily because:

A) It produced the first AI ethics principles document B) It demonstrated the dangers of autonomous weapons systems for the first time C) It established overconfident assumptions about AI timelines that shaped funding, policy, and governance in ways that created accountability gaps D) It was the first occasion on which AI researchers acknowledged the social risks of their work

Question 3 ProPublica's 2016 "Machine Bias" investigation of COMPAS found that:

A) The system had been deliberately programmed to produce racially biased predictions B) Black defendants were wrongly flagged as high risk for future crime at nearly twice the rate of white defendants C) The system was more accurate than human judges for all demographic groups D) The company (Northpointe) refused to release its algorithm even to courts using the system

Question 4 Anna Jobin and colleagues' 2019 analysis of 84 AI ethics principles documents found that:

A) There was genuine consensus across all documents on principles and their implementation B) The documents showed superficial agreement on vocabulary like "fairness" while masking deep disagreement about meaning and implementation C) Most documents came from civil society organizations rather than corporations or governments D) Only a minority of documents addressed concerns about bias and discrimination

Question 5 The primary ethical problem with "ethics washing" as described in this chapter is that:

A) It involves lying about AI system capabilities B) It deploys ethics language to reduce reputational and regulatory pressure without substantively changing practices that cause harm C) It allows organizations to avoid complying with GDPR requirements D) It results in ethics teams having excessive authority to block product deployments

Question 6 Joy Buolamwini's Gender Shades research (2018) found that commercial facial analysis systems from IBM, Microsoft, and Face++ performed worst on:

A) Older adults and children B) Individuals photographed in low-light conditions C) Darker-skinned women, with error rates up to 34.7 percentage points higher than for lighter-skinned men D) Individuals with unconventional facial expressions

Question 7 The "filter bubble" concept introduced by Eli Pariser (2011) identified what mechanism as its primary cause?

A) Social media users deliberately seeking out confirmation of their existing beliefs B) Personalization algorithms that optimize for engagement, showing users content consistent with their prior behavior C) Technology companies deliberately hiding information from users for political reasons D) Internet infrastructure designed to route users to geographically proximate servers with different content

Question 8 The EU AI Act (2024) takes what overall regulatory approach to AI systems?

A) A blanket prohibition on all AI systems in high-stakes domains pending regulatory approval B) A principles-based approach that articulates values without specific requirements C) A risk-based approach that classifies AI systems by potential for harm and imposes requirements proportional to that risk D) A self-regulatory approach that requires industry to establish compliance mechanisms without government oversight

Part II: True/False

Mark each statement True or False. If False, briefly explain why (one sentence).

Question 9 Alan Turing's 1950 paper "Computing Machinery and Intelligence" was exclusively focused on demonstrating that machines could think, without raising any ethical or social concerns about AI development.

Question 10 The first AI winter (1974–1980) was precipitated in part by the Lighthill Report, which assessed that AI had failed to deliver on its promises and recommended significant reductions in government funding.

Question 11 Amazon's hiring algorithm was found to penalize female applicants because engineers had deliberately programmed it to favor male candidates.

Question 12 GDPR is significant for AI governance primarily because it was the first comprehensive regulation to address AI systems specifically, with requirements designed exclusively for machine learning.

Question 13 Content moderators who review AI outputs and user-generated content for policy violations are routinely exposed to psychologically harmful material and have documented high rates of PTSD, depression, and anxiety.

Part III: Short Answer

Answer each question in the word count range provided. Demonstrate understanding of the chapter concepts; do not simply quote the text.

Question 14 (150–200 words) Explain what the "blank sheets" metaphor from Turing's writing means for how organizations should think about their responsibilities when developing AI systems. Who in an organization holds the responsibilities that the metaphor implies, and what specifically would fulfilling those responsibilities require?

Question 15 (200–250 words) What is the "genie problem" that Norbert Wiener identified, and why is it still relevant to organizations deploying AI systems today? Give one concrete example from a current AI deployment (not from the chapter) that illustrates the problem.

Question 16 (150–200 words) The chapter describes a five-stage pattern in AI ethics failures: (1) capability deployment, (2) harm documentation, (3) organizational denial or minimization, (4) public pressure, and (5) partial, delayed response. Identify a case from the chapter that illustrates this pattern and briefly explain which evidence maps to each stage.

Question 17 (200–250 words) Mary Gray and Siddharth Suri coined the term "ghost work" to describe the on-demand, platform-mediated labor that powers AI systems. Explain why this labor is typically invisible in AI products, what organizational and economic functions that invisibility serves, and what ethical responsibilities it creates for organizations that buy and deploy AI systems.

Part IV: Applied Scenario

Read each scenario carefully and respond as directed. There are no single correct answers; you will be evaluated on the quality of your reasoning, your application of chapter concepts, and the specificity of your response.

Question 18 (300–350 words) Scenario: You are the Chief Product Officer at a financial technology startup that is preparing to launch an AI-powered lending product. The product uses a machine learning model trained on historical lending data from partner banks over the past 15 years. The model has strong performance metrics on a test set held out from the training data. Your CTO says the model is ready to ship.

Before you approve the launch, what specific questions do you need answered about the training data and the model's behavior? Draw on at least three specific lessons from this chapter's history. What would you require to be demonstrated before approving deployment?

Question 19 (300–350 words) Scenario: You are advising a nonprofit organization that wants to use a generative AI tool to assist with producing advocacy communications — grant applications, public reports, and social media content. The organization serves communities that have historically been harmed by biased AI systems. Leadership is enthusiastic about the efficiency gains. Several staff members have raised concerns about whether using this technology is ethically consistent with the organization's mission.

Drawing on the history in this chapter, structure a framework for the organization to think through this decision. What historical lessons are most relevant? What specific risks should the organization evaluate before proceeding? What conditions, if any, would make deployment ethically defensible?

Question 20 (350–400 words) Scenario: A major hospital system has announced that it will deploy an AI triage tool in its emergency departments. The tool will assess incoming patients and predict which patients need immediate attention and which can safely wait. The hospital has published an AI ethics statement and established an AI ethics committee. Local civil rights organizations have raised concerns about whether the tool has been adequately tested on patients of color, elderly patients, and patients with disabilities.

Apply the concept of "ethics washing" from this chapter to evaluate the hospital's ethics statement and committee. What questions would you ask to determine whether the hospital's ethics infrastructure represents genuine commitment or performative compliance? What would genuine accountability for this deployment look like? Who should have the power to delay or halt deployment if concerns are not adequately addressed?

Answer Key

Note: For short answer and scenario questions, the answer key provides key concepts and reasoning frameworks rather than a single correct answer. Assess student responses on the quality of reasoning and application of chapter concepts.

Part I: Multiple Choice — Answers

Q1: B Wiener argued that feedback systems could replace human judgment in complex environments and that whether this was beneficial or harmful depended entirely on the social context of deployment — a framing that places ethics at the center of technology governance. Option A describes Turing's contribution. Option C describes Searle's Chinese Room argument. Option D describes subsequent empirical work, not Wiener's 1950 text.

Q2: C The Dartmouth project's primary ethical significance is its contribution to the hype cycle: overconfident assumptions about AI timelines shaped funding, policy, and public expectation in ways that created governance gaps when those timelines proved incorrect. Option A is incorrect — the Dartmouth project did not produce an ethics document. Option B is incorrect — autonomous weapons were not discussed. Option D is incorrect — the Dartmouth researchers were enthusiastic rather than cautious.

Q3: B ProPublica's central finding was that Black defendants were wrongly flagged as high risk at nearly twice the rate of white defendants, while white defendants were more likely to be incorrectly flagged as low risk. Option A is incorrect — the bias emerged from training data patterns, not deliberate programming. Option C is incorrect — the system's accuracy relative to human judges was not a finding of the ProPublica study. Option D is incorrect — Northpointe did provide limited access to the algorithm to courts; the concern was about public transparency.

Q4: B Jobin et al.'s key finding was superficial vocabulary consensus masking substantive disagreement about meaning, implementation, and trade-offs. Option A is incorrect — the paper specifically found that apparent consensus concealed disagreement. Option C is incorrect — many documents came from corporations and governments. Option D is incorrect — bias and discrimination were among the most commonly addressed topics.

Q5: B Ethics washing specifically refers to deploying ethics language as a reputational and regulatory strategy without substantive accompanying change. Option A is incorrect — ethics washing does not necessarily involve lying about capabilities. Option C is incorrect — ethics washing is not specifically about GDPR. Option D is incorrect — the problem with ethics washing is that ethics functions have too little authority, not too much.

Q6: C The Gender Shades paper specifically found that commercial facial analysis systems performed worst on darker-skinned women, with error rates up to 34.7 percentage points higher than for lighter-skinned men. The other options describe different populations not identified as the primary locus of error in the Gender Shades research.

Q7: B Pariser's filter bubble analysis located the primary mechanism in personalization algorithms optimizing for engagement — showing users content consistent with their prior behavior because that behavior generates more engagement. Option A describes a user behavior that the algorithm reinforces but does not cause. Options C and D describe mechanisms that Pariser did not identify as the primary cause.

Q8: C The EU AI Act takes a risk-based approach: classifying systems by potential for harm and imposing requirements proportional to that classification. Options A and B are incorrect characterizations of the Act's approach. Option D describes self-regulation, which the Act explicitly rejects in favor of regulatory oversight.

Part II: True/False — Answers

Q9: FALSE Turing's paper devoted substantial space to anticipating and responding to objections to machine intelligence, and it included intuitions about the social implications of learning machines — including the observation that a learning system's values and capabilities are substantially determined by its training inputs. The paper was not exclusively focused on demonstrating AI possibility; it was also concerned with the conceptual and social questions that AI raises.

Q10: TRUE The Lighthill Report (1973) assessed that AI had failed to deliver on its promises in machine translation and general problem-solving, and its recommendations contributed to significant funding cuts in both the UK and US. This is accurately described in the chapter.

Q11: FALSE Amazon's algorithm was not deliberately programmed to favor male candidates. The bias emerged from the pattern recognition the system applied to historical data — resumes submitted over a 10-year period during which Amazon's workforce was predominantly male. The system learned from patterns in the data rather than being explicitly programmed with discriminatory rules. This distinction matters because it illustrates how statistical learning systems can produce discriminatory outputs without discriminatory intent.

Q12: FALSE GDPR is a comprehensive data protection regulation, not an AI-specific regulation. Its significance for AI governance comes from provisions relevant to automated decision-making, transparency requirements, and data minimization that constrain how AI training data can be collected and used — but GDPR was not designed exclusively for AI and does not address many AI-specific concerns. AI-specific regulation in the EU came later, with the EU AI Act (2024).

Q13: TRUE This is accurately documented in the chapter, drawing on multiple sources including Karen Hao's MIT Technology Review reporting on content moderation work for OpenAI in Kenya, and reporting on Facebook moderator lawsuits. Studies have documented high rates of PTSD, depression, anxiety, and substance abuse among content moderation workers.

Part III: Short Answer — Key Concepts

Q14: Key concepts The blank sheets metaphor implies that learning systems' values and capabilities are substantially determined by their training inputs — what they are exposed to during development. Strong responses will identify that this creates responsibilities for: (a) the curation of training data (who selects it, what it represents, what it excludes), (b) the design of training objectives (what the system is optimized for), and (c) the testing of trained systems before deployment. Responsibility should be assigned to specific organizational roles — data scientists, product managers, ethics reviewers — rather than left diffuse. Weaker responses will be vague about what "responsibility" specifically requires.

Q15: Key concepts The genie problem refers to machines optimizing for specified objectives in ways that achieve those objectives while violating human values the specification did not capture. Strong responses will articulate this clearly and connect it to the alignment problem (specifying objectives that fully capture human values is very difficult). A good contemporary example might include: recommendation algorithms optimizing for engagement while producing radicalization; language models optimizing for human approval ratings while producing flattering but inaccurate outputs; or advertising systems optimizing for click-through rates while exploiting psychological vulnerabilities.

Q16: Key concepts Any of the major cases in the chapter can be used. For COMPAS: (1) deployment in criminal sentencing; (2) ProPublica's 2016 investigation; (3) Northpointe disputed ProPublica's methodology; (4) academic debate and advocacy attention to algorithmic criminal justice; (5) some jurisdictions limited COMPAS use but no comprehensive regulatory response. Strong responses will be specific about the evidence for each stage.

Q17: Key concepts Invisibility is maintained through: contractual intermediaries between AI companies and workers; non-disclosure agreements; product communications that describe AI capabilities without disclosing human labor inputs; classification of workers as independent contractors rather than employees. Organizational functions served include: reducing labor relations exposure, simplifying product narratives, and limiting accountability for working conditions. Ethical responsibilities for AI buyers include: supply chain due diligence, vendor qualification on labor standards, and advocacy for industry standards.

Part IV: Applied Scenario — Key Concepts

Q18: Key concepts Strong responses will include at minimum: (a) questions about the demographic composition of the historical lending data and what inequities it reflects (lesson from FICO and disparate impact), (b) testing of the model's performance across demographic groups, not just overall accuracy (lesson from Gender Shades, COMPAS), and (c) pre-deployment adversarial testing for unexpected failure modes (lesson from Tay, expert systems failures). Strong responses will require specific documentation — disaggregated performance metrics, data provenance documentation — rather than assurances. They will also recognize that "strong performance on test set" does not guarantee absence of disparate impact if the test set has the same demographic composition issues as the training set.

Q19: Key concepts Strong responses will apply: the pattern of harm documentation by communities (is this organization positioned to receive and act on such documentation?); the distinction between ethics washing and genuine commitment (is the organization merely announcing values or building processes?); the specific risks of generative AI identified in Section 2.7 (hallucination, bias in outputs, content attribution). Conditions for defensible deployment might include: explicit testing for bias in outputs on the organization's specific use cases, transparency with constituents about AI use, human review of AI-generated content before publication, and a defined process for withdrawing from AI use if problems emerge.

Q20: Key concepts Strong responses will apply the ethics washing framework by asking specific questions that distinguish appearance from substance: Does the ethics committee have authority to delay deployment? What criteria would trigger that authority? How are affected community members (patients of color, elderly patients, patients with disabilities) represented in ethics review? Were the civil rights organizations' concerns specifically addressed in testing? What are the performance metrics on different patient populations? "Genuine accountability" requires: public disclosure of disaggregated performance metrics, mechanisms for affected patients to raise concerns, and independent review with authority to halt deployment. Power to delay or halt deployment should reside with a body that includes representatives of affected communities, not exclusively with hospital administration or the AI vendor.