AI Associate Ethical Considerations of AI Practice Test 14 — 15 Questions

Question 1

A customer service department uses an AI chatbot to handle common inquiries. Recently, customers have reported that the chatbot sometimes responds with offensive or inappropriate language. The company wants to uphold ethical standards. Which approach is the best practice?

Choose an answer to begin — your selection is scored in the full session.

Accepted Answer

Implement human review of all chatbot responses that are flagged as potentially offensive.. Implementing human review of flagged chatbot responses combines automated detection with human judgment, which is the most reliable way to catch context-dependent offensive language and uphold ethical standards while still allowing the chatbot to operate at scale. This human-in-the-loop approach addresses the root problem of inappropriate outputs rather than merely masking symptoms. Option A is insufficient because simple word filters miss context, sarcasm, and implicitly offensive phrasing, and can also over-block legitimate content. Option B would eliminate offensive responses but destroys the chatbot's usefulness by removing its ability to handle varied inquiries. Option D is flawed because no pre-trained model from any vendor can guarantee ethical behavior; model outputs still require monitoring and governance.

Answer

Implement a filter to automatically block any offensive words in the chatbot's responses.

Answer

Limit the chatbot to only respond from a fixed set of predefined answers.

Answer

Use a pre-trained language model from a trusted vendor to guarantee ethical behavior.

Question 2

Which TWO actions are most effective for ensuring fairness in an AI model used for loan approvals?

Accepted Answer

Regularly audit model outcomes for disparate impact across demographic groups.. Option A is correct because regularly auditing model outcomes for disparate impact across demographic groups is the core mechanism for detecting and quantifying bias in loan-approval decisions, using metrics such as the disparate impact ratio (e.g., the 80% rule) to verify that protected classes are not unfairly disadvantaged. Option D is correct because using diverse and representative training data ensures the model learns patterns that generalize fairly across the population it serves; biased or unrepresentative data is a primary root cause of unfair lending outcomes, so representative sampling directly reduces that risk. The remaining options do not belong: relying solely on historical data (B) perpetuates existing biases rather than correcting them, letting the model self-correct without human intervention (C) provides no guarantee of fairness and can drift further, and using a single simple algorithm (E) does not by itself ensure fairness since a simple model can still encode biased data or features.

Answer

Rely solely on historical data without any adjustments.

Answer

Allow the model to self-correct over time without human intervention.

Answer

Use a single, simple algorithm to avoid complexity.

Question 3

Which TWO actions promote transparency in AI decision-making?

Accepted Answer

Document the model's limitations and potential biases.. Option B is correct because documenting a model's limitations and potential biases is a core transparency practice: it discloses where the system may fail or produce skewed outcomes, enabling stakeholders to interpret results responsibly and satisfy governance and audit requirements. Option E is correct because providing clear explanations for significant AI decisions makes the reasoning behind outcomes understandable to users, auditors, and regulators, which is the essence of transparent decision-making (e.g., via explainability techniques such as feature attribution or model cards). The unmarked options do not belong: A (black-box models for accuracy) reduces interpretability and hides how decisions are reached; C (keeping algorithms proprietary) restricts disclosure and thus transparency; and D (sharing only final results) withholds the rationale and process, preventing meaningful scrutiny.

Answer

Use black-box models that maximize accuracy.

Answer

Keep the AI algorithms proprietary to protect intellectual property.

Answer

Only share the final results of the AI system.

Question 4

Which THREE components are essential for an ethical AI governance framework within a large enterprise?

Accepted Answer

Establish a cross-functional AI ethics board.. Option A is correct because a cross-functional AI ethics board is essential for an ethical AI governance framework in a large enterprise, as it brings together legal, compliance, data science, business, and risk stakeholders to set policy, review high-risk use cases, and resolve ethical conflicts across the organization. Option B is correct because regular ethical impact assessments are a core governance control that systematically evaluates AI systems for bias, fairness, privacy, transparency, and societal harm before and after deployment, enabling risks to be identified and mitigated. Option C is correct because defining clear accountability for AI outcomes ensures that specific roles and owners are responsible for decisions, monitoring, and remediation, which is fundamental to enforceable governance and regulatory compliance. Option D is not correct because minimizing human oversight to cut costs directly undermines ethical governance, which requires appropriate human-in-the-loop review and control for high-risk AI decisions. Option E is not correct because optimizing for accuracy as the primary goal is a narrow technical objective that can conflict with ethical requirements such as fairness, explainability, and safety, so it is not an essential component of an ethical governance framework.

Answer

Minimize human oversight to reduce operational costs.

Answer

Optimize for accuracy as the primary goal.

Question 5

A financial services firm deployed an AI model to automate loan approvals. The model was trained on historical loan data from the past 10 years, which shows that applicants from certain zip codes have higher default rates. After six months, the company's compliance team receives complaints that applicants from predominantly low-income neighborhoods are being rejected at a much higher rate than applicants from affluent areas, even when their financial profiles are similar. The model's overall accuracy remains high (95%), and the loan default rate has decreased by 15% since deployment. The company wants to address the ethical concerns without sacrificing performance. Which course of action should the company take?

Accepted Answer

Retrain the model with a balanced dataset that includes more examples from underrepresented neighborhoods and enforce fairness constraints.. The model is exhibiting disparate impact via proxy variables (zip code correlates with socioeconomic status), so the fix is to retrain with a more balanced dataset and enforce fairness constraints, which addresses the root cause while preserving performance. Simply removing zip code doesn't eliminate correlated proxies, and threshold manipulation is itself discriminatory. A balanced retraining approach targets both fairness and accuracy.

Answer

Remove the zip code feature from the model inputs.

Answer

Adjust the approval threshold lower only for applicants from low-income neighborhoods.

Answer

Continue using the existing model since it has high accuracy and reduces defaults.

Question 6

A hospital uses an AI triage system to prioritize patients in the emergency department. The AI was trained on historical patient data and assigns priority scores based on vital signs and symptoms. Recently, a study finds that the system consistently assigns lower priority to elderly patients compared to younger patients with similar clinical presentations. The hospital's ethics committee is concerned about age discrimination. The current model achieves high accuracy in predicting outcomes, and doctors have come to rely on it for efficiency. What should the hospital do to address the ethical concern while maintaining clinical effectiveness?

Accepted Answer

Retrain the model with a modified objective that penalizes age-based disparities.. Retraining the model with a fairness-aware objective that penalizes age-based disparities directly targets the observed bias while preserving the model's predictive utility and clinical workflow. This approach allows the hospital to mitigate discriminatory prioritization without discarding the system's demonstrated accuracy and efficiency. Option A is overly disruptive and unnecessary, as the existing model can be adjusted rather than rebuilt from scratch. Option C ignores the ethical and potentially legal problem of age discrimination, and Option D is insufficient because removing age as an explicit feature does not eliminate bias from correlated proxies such as vital signs, symptoms, or historical treatment patterns.

Answer

Replace the AI triage system with a completely new model built from scratch.

Answer

Continue using the current model since it has high accuracy and efficiency.

Answer

Remove age as an input feature from the model.

Question 7

A retail company uses an AI recommendation engine to suggest products to online shoppers. The engine uses past purchase history and browsing behavior. Recently, a customer advocacy group publishes a report showing that the engine recommends higher-priced products to customers in affluent zip codes and lower-priced products to customers in lower-income areas, even when both groups have similar browsing histories. The company's revenue has increased since implementing the engine, and marketing teams are pleased. However, the company wants to maintain a reputation for fairness. Which action should the company take?

Accepted Answer

Implement fairness constraints to ensure similar recommendation distributions across demographic groups.. The correct action is D: implement fairness constraints to ensure similar recommendation distributions across demographic groups. This directly addresses the documented disparity where the engine recommends higher-priced products to affluent zip codes and lower-priced products to lower-income areas despite similar browsing histories, by enforcing comparable recommendation outcomes across groups while preserving the AI system's personalization and revenue benefits. Option A does not fit because showing only the most popular products to everyone eliminates personalization and does not guarantee fair treatment across demographic groups. Option B is overly drastic, as discontinuing the system entirely discards its business value instead of mitigating the bias. Option C is inappropriate because keeping the system solely for revenue ignores the fairness problem and the company's stated reputational goal.

Answer

Show only the most popular products to all customers regardless of browsing history.

Answer

Discontinue the AI recommendation system entirely.

Answer

Keep the current system as it maximizes revenue.

Question 8

An HR department uses an AI tool to screen resumes for a software engineering position. The tool was trained on resumes of past successful hires, who were predominantly male. The tool has been in use for three months, during which only 10% of candidates shortlisted for interviews are female, even though 40% of applicants are female. The hiring managers are satisfied with the quality of candidates shortlisted, as most perform well in interviews. However, the company's diversity and inclusion officer raises an ethical concern. What should the company do to address this bias?

Accepted Answer

Retrain the model with a balanced dataset that includes more female candidates and remove gender-correlated features.. The bias stems from training data reflecting predominantly male past hires, so retraining with a balanced dataset and removing gender-correlated features directly addresses the root cause of the disparate shortlisting rate (10% female vs. 40% female applicants). This approach aligns with standard fairness remediation in AI hiring: mitigate biased training data and proxy variables rather than only patching outcomes. Option B does not fix the biased model and merely adds an appeal process, which still disadvantages female candidates at scale. Option C ignores the ethical and legal disparate-impact concern because interview performance does not justify a biased screening stage. Option D may reduce bias but is a blunt, non-scalable workaround that discards the AI tool rather than correcting it.

Answer

Use the AI tool for initial screening but allow candidates to appeal the decision.

Answer

Continue using the current tool since it selects high-quality candidates.

Answer

Manually review all resumes without using the AI tool.

Question 9

A social media platform uses an AI model to automatically detect and remove hate speech. The model uses natural language processing and was trained on public posts. Recently, an internal audit reveals that the model removes posts from minority ethnic groups at a rate 3 times higher than from majority groups, even when the content is similar. The model achieves high precision and recall on the test set. The platform's content moderation team is overwhelmed with appeals. The company wants to maintain a safe environment while being fair. Which approach best addresses both goals?

Accepted Answer

Conduct an audit of the training data to identify gaps, then retrain with more representative data including diverse examples of hate speech and non-hate speech.. The disparity stems from the training data: if the model was trained on public posts that under-represent minority dialects, slang, or reclaimed language, it will misclassify similar content differently across groups. Auditing the training data for representation gaps and retraining with balanced, diverse examples of both hate speech and non-hate speech directly addresses the root cause while preserving the AI moderation capability. This is the standard responsible-AI remediation for biased models.

Answer

Disable the AI moderation and rely solely on user reports.

Answer

Add more human moderators to review all flagged content from minority groups.

Answer

Adjust the detection threshold only for minority group posts to reduce flags.

Question 10

A university uses an AI system to predict first-year student retention. The system uses factors such as high school GPA, SAT scores, and socioeconomic indicators. After two years, administrators notice that the model consistently predicts lower retention probabilities for students from low-income families, even when their academic profiles are strong. The university's mission emphasizes equity and inclusion. The admissions office is considering using the predictions to allocate support resources. The model's accuracy on historical data is 85%. What should the university do to align with ethical AI principles?

Accepted Answer

Retrain the model with a fairness constraint such as equalized odds to reduce income-based disparities.. The scenario describes a model whose predictions are systematically biased against low-income students despite strong academic profiles, so retraining with a fairness constraint such as equalized odds directly targets the income-based disparity while preserving the AI system's utility for allocating support resources. Equalized odds requires similar true positive and false positive rates across groups, which is an appropriate technical remedy when the university's mission emphasizes equity and inclusion. Option A is insufficient because simply removing socioeconomic indicators does not eliminate proxy discrimination, since correlated features like high school GPA or neighborhood data can still encode income status. Option B is overly drastic and discards a system that can be made fairer, while Option D ignores the ethical harm by prioritizing aggregate 85% accuracy over disparate impact on a protected or vulnerable group.

Answer

Remove socioeconomic indicators from the model inputs.

Answer

Abandon the AI system and rely on human advisors for resource allocation.

Answer

Use the predictions as-is because they are accurate for the majority of students.

Question 11

An insurance company uses an AI model to set auto insurance premiums. The model uses factors including driving history, age, and ZIP code. A regulator finds that premiums in certain low-income neighborhoods are significantly higher than in affluent neighborhoods with similar risk profiles. The company's actuaries argue that the model is actuarially sound because it accurately predicts claims based on historical data. The company wants to comply with ethical guidelines and avoid legal action. Which action should they take?

Accepted Answer

Incorporate a fairness constraint that requires similar premiums for similar risk profiles regardless of ZIP code.. The model produces disparate impact by using ZIP code as a proxy for race or socioeconomic status, resulting in higher premiums for similar-risk individuals in low-income neighborhoods. Incorporating a fairness constraint that requires similar premiums for similar risk profiles regardless of ZIP code directly addresses the discriminatory outcome while preserving actuarial risk differentiation. This aligns with ethical AI principles and reduces legal exposure under fair-lending and insurance regulations.

Answer

Defend the model based on its actuarial accuracy and historical claims data.

Answer

Cap premium increases in low-income neighborhoods at a fixed percentage.

Answer

Remove ZIP code from the model inputs entirely.

Question 12

A news aggregator app uses an AI algorithm to personalize the news feed for each user. The algorithm selects articles based on past clicks and reading time. Recently, a study reveals that the algorithm disproportionately shows sensational and polarizing news to users from certain political orientations, while showing more neutral content to others. The company's user engagement metrics have increased, but journalists express concern about reinforcing echo chambers and misinformation. The company wants to uphold ethical standards while keeping users engaged. What should they do?

Accepted Answer

Modify the algorithm to include diversity and reliability scores for news sources, promoting a balanced feed.. Adding diversity and reliability scores directly addresses the algorithmic bias identified in the study while preserving personalization and engagement; it mitigates echo chambers and misinformation by weighting sources for balance and trustworthiness rather than optimizing solely for clicks and reading time. Options B, C, and D do not fit: B shifts the burden to users and does not fix the underlying biased ranking, C removes personalization entirely and would likely reduce engagement, and D ignores the ethical problem by prioritizing engagement metrics over fairness and accuracy.

Answer

Allow users to manually select the types of news they want to see.

Answer

Show the same generic news feed to all users.

Answer

Continue with the current algorithm since it increases engagement.

Question 13

A credit scoring company develops an AI model that includes social media activity as a factor. The model awards higher scores to individuals with many online connections and consistent posting. Consumer advocates argue that this penalizes individuals with limited internet access or those who value privacy. The company defends the model, stating that it predicts creditworthiness better than traditional models. However, a regulatory body is investigating potential discrimination. The company wants to address ethical concerns without completely abandoning the model. Which approach is most appropriate?

Accepted Answer

Conduct a thorough analysis to determine whether social media activity is a legitimate, non-discriminatory predictor of creditworthiness.. The scenario calls for investigating whether social media activity is a legitimate, non-discriminatory predictor of creditworthiness before deciding how to proceed, which directly addresses the regulatory discrimination concern while preserving the model if it proves valid. A thorough analysis would assess disparate impact on protected groups, such as individuals with limited internet access, and evaluate whether the model's predictive benefit justifies its use. Option A is too abrupt because it abandons the model without evidence that social media data is unlawful or unethical. Option B may reduce reliance on social media but does not determine whether the model itself is discriminatory, and Option D still uses the potentially discriminatory model for some applicants while offering only an alternative, which does not resolve the regulatory investigation.

Answer

Remove social media data from the model immediately.

Answer

Increase the weight of traditional factors like income and payment history.

Answer

Continue using the current model but offer an alternative traditional scoring option.

Question 14

A government agency uses an AI system to allocate resources for public services such as healthcare and education. The system is designed to optimize overall efficiency based on historical usage data. After deployment, it becomes clear that underserved regions with less historical data receive significantly less funding than well-served regions. The agency's mission is to promote equity. The system's performance metrics show high efficiency, but community leaders protest the unfair distribution. What should the agency do?

Accepted Answer

Redesign the system to include fairness constraints that ensure minimum resource levels for underserved regions.. The scenario describes an AI system that optimizes efficiency but produces inequitable outcomes for underserved regions, conflicting with the agency's equity mission. The correct response is to redesign the system to incorporate fairness constraints that guarantee minimum resource levels for underserved regions, aligning the AI's objective with the agency's values. This balances efficiency with equity and addresses the root cause of the bias.

Answer

Abandon the AI system and use a manual, rule-based allocation system.

Answer

Collect more historical data from underserved regions before making adjustments.

Answer

Continue using the system as is, since it maximizes efficiency.

Question 15

A retail company uses Salesforce Einstein Vision to analyze customer images for product recommendations. The AI team notices that the model performs poorly on images of customers with darker skin tones, leading to fewer recommendations for that demographic. The team has access to a dataset of diverse skin tones but the company's data privacy policy prohibits using demographic data in training. What should the team do?

Accepted Answer

Document the bias, escalate to the ethics board, and seek guidance on using diverse data while maintaining privacy.. Option C is correct because when a model shows demographic performance disparity and privacy policy blocks the obvious fix, the proper response is to document the bias, escalate to the ethics board, and obtain governance guidance on how to use diverse data compliantly. This preserves accountability and ensures any mitigation respects the data privacy policy rather than unilaterally overriding it. Option A is wrong because retraining on the diverse dataset without considering privacy regulations violates the stated policy. Option B is wrong because ignoring the disparity leaves the discriminatory performance unaddressed. Option D is wrong because simply removing skin tone labels does not resolve the privacy issue or guarantee the bias is mitigated, since the underlying demographic data would still be used without approval.

Answer

Retrain the model on the diverse dataset without considering privacy regulations.

Answer

Ignore the performance disparity as it is a result of natural data distribution.

Answer

Use the diverse dataset but remove skin tone labels to avoid privacy issues.