AI Associate Ethical Considerations of AI Practice Test 4 — 15 Questions

Question 1

An AI system recommends job candidates to recruiters. The system was trained on resumes of past successful hires, most of whom were male. As a result, it consistently ranks female candidates lower. What is the most appropriate mitigation?

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

Accepted Answer

Re-sample the training data to include more female candidates and use fairness-aware algorithms.. Option A is correct because the bias originates in the training data, which overrepresents male successful hires, so re-sampling to include more female candidates combined with fairness-aware algorithms addresses the root cause and enforces equitable outcomes during training. Fairness-aware methods (e.g., reweighting, adversarial debiasing, or constrained optimization) explicitly optimize for group fairness metrics rather than relying on the skewed historical distribution. Option B is weaker because post-processing score adjustments only mask the symptom and can be brittle or legally questionable without fixing the underlying data imbalance. Option C is wrong because accepting bias perpetuates discriminatory outcomes and likely violates anti-discrimination law. Option D is insufficient because removing the gender feature does not eliminate bias, since correlated proxy features (e.g., names, schools, career gaps) still encode gender information.

Answer

Add a post-processing adjustment to increase female candidates' scores.

Answer

Accept the bias as a reflection of historical data.

Answer

Remove the gender feature from the model.

Question 2

A company wants to deploy an AI system that makes hiring decisions. To comply with ethical guidelines, what should they do before deployment?

Accepted Answer

Conduct an ethics review and perform bias testing on diverse datasets.. Conducting an ethics review and performing bias testing on diverse datasets are essential steps to identify and mitigate potential discriminatory outcomes in AI-driven hiring systems. This aligns with ethical AI frameworks that require fairness, accountability, and transparency before deployment, ensuring the model does not perpetuate historical biases or violate anti-discrimination laws.

Answer

Ensure the system achieves high accuracy and ignore other metrics.

Answer

Deploy immediately and monitor for issues.

Answer

Test the system only on a small dataset to expedite launch.

Question 3

An AI system used for medical diagnosis has been shown to have lower accuracy for certain ethnic groups. The development team is considering releasing it anyway because most patients are from the majority group. Which ethical principle is most compromised?

Accepted Answer

Fairness. The scenario describes an AI system that performs worse for certain ethnic groups, yet the team plans to release it anyway because the majority group is unaffected. This directly violates the principle of fairness, which requires that AI systems do not discriminate or perpetuate bias against any group. Releasing a model with known accuracy disparities without mitigation prioritizes overall performance over equitable treatment, compromising fairness.

Answer

Transparency

Answer

Accountability

Answer

Privacy

Question 4

Which TWO actions best promote transparency in an AI system?

Accepted Answer

Publish an audit trail of model inputs and decisions.. Option B is correct because publishing an audit trail of model inputs and decisions creates a verifiable record that lets stakeholders trace how and why outputs were produced, which is a core mechanism of transparency and accountability. Option D is correct because providing clear explanations for individual predictions (e.g., via interpretability techniques such as SHAP, LIME, or counterfactual explanations) makes the model's reasoning understandable to users and auditors on a per-decision basis. Option A does not belong because restricting access to the model's logic reduces transparency by hiding how the system works, even if it protects intellectual property. Option C does not belong because using a complex deep learning model for higher accuracy addresses performance, not transparency, and can actually decrease interpretability. Option E does not belong because removing feature importance strips away information about which inputs drive predictions, directly undermining transparency rather than promoting it.

Answer

Limit access to the model's logic to protect intellectual property.

Answer

Use a complex deep learning model for higher accuracy.

Answer

Remove feature importance to simplify the model.

Question 5

Which THREE factors should be considered when evaluating the fairness of an AI model?

Accepted Answer

Disparate impact ratio across groups.. The fairness of an AI model is assessed with group-fairness metrics that compare outcomes across protected groups, so option A (disparate impact ratio across groups) is correct because it measures the ratio of favorable outcome rates between groups (typically using the 80% rule) to detect adverse impact. Option D (equal opportunity difference) is correct because it compares true positive rates across groups, ensuring qualified individuals receive favorable predictions regardless of group membership. Option E (demographic parity in predictions) is correct because it requires equal positive prediction rates across groups, a standard statistical parity fairness criterion. Option B (overall accuracy on the test set) is not a fairness metric; it aggregates performance across all groups and can mask disparate harm. Option C (model training time) is an operational efficiency measure and has no bearing on fairness.

Answer

Overall accuracy on the test set.

Answer

Model training time.

Question 6

Which TWO practices help ensure accountability in AI systems?

Accepted Answer

Assign a human owner for each AI system.. Option B is correct because assigning a named human owner for each AI system establishes a clear point of responsibility, so there is always an identifiable person accountable for the system's behavior, oversight, and remediation rather than diffuse or absent ownership. Option D is correct because implementing logging and auditing of model decisions creates an immutable, reviewable record of inputs, outputs, and model versions, which is essential for tracing what happened, detecting misuse or drift, and holding people accountable after the fact. By contrast, Option A is wrong because blaming the AI system shifts responsibility away from the humans who design, deploy, and operate it, and an AI cannot bear accountability. Option C is wrong because fully automating decisions removes human oversight and actually erodes accountability rather than strengthening it. Option E is wrong because open-sourcing code may improve transparency but does not by itself assign responsibility, and shared or diffuse responsibility can weaken accountability.

Answer

Blame the AI system for mistakes to protect employees.

Answer

Automate all decisions to eliminate human error.

Answer

Open-source the model code to share responsibility.

Question 7

Refer to the exhibit. The fairness evaluation shows a disparate impact of 0.85, equal opportunity difference of 0.12, and demographic parity difference of 0.18. Which fairness thresholds are violated?

Accepted Answer

Equal opportunity and demographic parity.. The fairness evaluation shows violations of both equal opportunity and demographic parity thresholds. The equal opportunity difference of 0.12 exceeds the commonly accepted threshold of 0.1, and the demographic parity difference of 0.18 exceeds the typical threshold of 0.1. Disparate impact of 0.85 is within the acceptable range (typically 0.8 to 1.25), so it is not violated.

Answer

Demographic parity only.

Answer

Equal opportunity only.

Answer

Disparate impact only.

Question 8

Refer to the exhibit. What is the most likely cause of the fairness issue?

Accepted Answer

The training data is imbalanced, causing the model to perform better on the majority group.. The fairness issue is most likely caused by imbalanced training data, where one demographic group (e.g., males) is overrepresented, causing the model to learn patterns that favor the majority group and perform worse on the minority group. This is the classic source of disparate impact in ML fairness — the model reflects the statistical distribution of the data it was trained on, so underrepresentation of females leads to higher error rates for that group.

Answer

The model overfits to the male group.

Answer

The overall accuracy is too low.

Answer

The model is inherently biased against females.

Question 9

A company is deploying an AI-powered chatbot to handle customer service inquiries. The bot uses historical chat data for training. Which ethical consideration is MOST important to address before deployment?

Accepted Answer

Obtaining consent from customers whose data is used for training. The most critical ethical consideration is obtaining consent from customers whose historical chat data is used to train the chatbot. Under regulations like GDPR and CCPA, personal data (including chat transcripts) requires explicit consent for processing, especially when used to train an AI model. Deploying without consent violates data privacy laws and erodes user trust, regardless of the chatbot's technical performance.

Answer

Maximizing the chatbot's response accuracy

Answer

Ensuring the chatbot can handle high traffic volumes

Answer

Designing a human handoff protocol for complex issues

Question 10

A healthcare provider uses an AI model to predict patient readmission risk. The model is trained on historical data that underrepresents minority populations. What is the MOST significant ethical risk?

Accepted Answer

The model may produce biased predictions against minorities. Training on historical data that underrepresents minority populations leads to a model that has insufficient examples to learn patterns for those groups, resulting in biased predictions that systematically disadvantage minorities. This is a direct violation of fairness in AI ethics, as the model's outputs will be less accurate or equitable for underrepresented groups, potentially causing harm in critical healthcare decisions like readmission risk assessment.

Answer

The model may overfit to the majority population

Answer

The model cannot scale to real-time predictions

Answer

The model lacks explainability

Question 11

A company is designing an AI system to screen job applicants. To ensure fairness, which practice should be implemented?

Accepted Answer

Conduct regular fairness audits on model outcomes. Regular fairness audits are essential because they systematically evaluate model outcomes for bias across demographic groups, using metrics like disparate impact or equal opportunity difference. This practice aligns with responsible AI frameworks (e.g., NIST AI Risk Management Framework) and helps detect subtle biases that may emerge from proxy variables or data drift, ensuring the screening process remains equitable over time.

Answer

Use only one data source for consistency

Answer

Maximize the model's accuracy on historical hiring decisions

Answer

Remove all demographic data from the training set

Question 12

An AI model for predicting employee performance is found to have a higher false positive rate for women than for men. What is the best course of action?

Accepted Answer

Investigate the cause and retrain the model to reduce bias. A higher false positive rate for women indicates the model has learned biased patterns from the training data, likely due to imbalanced or skewed historical data. Investigating the cause—such as examining feature correlations, data distribution, and model architecture—allows for targeted retraining (e.g., reweighting, adversarial debiasing, or fairness constraints) to reduce bias without sacrificing overall performance. This aligns with ethical AI principles and regulatory expectations, ensuring the model is fair across demographic groups.

Answer

Lower the decision threshold for women to equalize false positive rates

Answer

Proceed with deployment because the overall accuracy is acceptable

Answer

Use the model but require manual review for all female candidates

Question 13

A nonprofit uses an AI system to allocate resources to communities in need. The system uses historical data which shows that certain neighborhoods have lower service usage. What ethical risk should be considered?

Accepted Answer

The system may perpetuate historical inequities. The AI system uses historical data that reflects lower service usage in certain neighborhoods. If that historical data is biased due to past inequities (e.g., redlining, underinvestment, or systemic discrimination), the model will learn and amplify those patterns, leading to unfair resource allocation that perpetuates historical disadvantages. This is a classic case of algorithmic bias where the training data encodes societal biases, and the model's predictions reinforce them.

Answer

The system may violate data minimization principles

Answer

The system cannot be held accountable for decisions

Answer

The system lacks explainability

Question 14

A company is developing an AI system to assist with hiring. Which TWO practices are essential for ethical AI deployment?

Accepted Answer

Conduct regular bias audits on model predictions. Option C is correct because regular bias audits on model predictions are essential to detect and mitigate disparate impact across protected groups, using metrics such as demographic parity or equalized odds to verify the model does not systematically disadvantage any group. Option E is correct because obtaining informed consent from applicants when their data is used respects data-subject autonomy and complies with privacy regulations like GDPR, which require lawful bases and transparency for processing personal data in automated decision-making. Options A, B, and D do not belong: optimizing for speed (A) addresses latency, not ethics; removing all demographic features (B) does not guarantee fairness because proxy variables can still encode bias; and maximizing accuracy on historical hiring data (D) can perpetuate and amplify past discriminatory patterns rather than promote ethical outcomes.

Answer

Optimize the system for speed to reduce waiting times

Answer

Remove all demographic features to ensure fairness

Answer

Maximize accuracy on historical hiring data

Question 15

An AI system is used to detect fraud in financial transactions. Which THREE steps should be taken to address ethical concerns?

Accepted Answer

Implement a human-in-the-loop for high-stakes decisions. Option C is correct because a human-in-the-loop for high-stakes decisions ensures that consequential fraud determinations—such as freezing accounts or denying transactions—receive human review, preserving accountability and allowing context that an automated model may miss. Option D is correct because requiring the model to provide explanations for its decisions supports transparency and contestability, letting investigators, auditors, and affected customers understand the basis of a fraud flag and challenge errors. Option E is correct because testing for disparate impact across demographic groups detects and mitigates bias, helping ensure the system does not unfairly disadvantage protected classes in violation of fair-lending and anti-discrimination obligations. Option A does not belong because simply lowering the fraud detection threshold increases false positives and can amplify harm and unfair burden rather than resolve ethical concerns. Option B does not belong because automatically accepting all flagged transactions removes human oversight and accountability, directly undermining ethical safeguards and increasing fraud risk.

Answer

Lower the fraud detection threshold to catch more cases

Answer

Automatically accept all flagged transactions to improve user experience