AI Associate Ethical Considerations of AI Practice Test 11 — 15 Questions

Question 1

Refer to the exhibit. An organization implements this AI fairness policy for their Einstein Prediction Builder model. What is the most significant ethical gap in this policy?

Accepted Answer

It does not specify a process for reviewing and remediating models that fail the fairness threshold. The policy states a fairness threshold and a monitoring cadence but never defines who reviews a failing model, what remediation steps follow, or how a model is retrained, replaced, or retired, which is the critical ethical gap for an enforceable AI fairness policy. A fairness policy is only meaningful if it closes the loop from detection to corrective action, including accountability, escalation, and documentation of outcomes. Options A and B are not the most significant gap here: the number of protected attributes and the choice of demographic parity are design considerations that may be appropriate depending on the exhibit's context, but they are not the missing governance mechanism. Option D is also not the key issue, since monitoring frequency alone does not ensure that failures are acted upon.

Answer

It does not include enough protected attributes

Answer

The demographic parity metric is inappropriate for this use case

Answer

The monitoring schedule is too frequent

Question 2

Refer to the exhibit. Which ethical principle is most at risk based on this AI governance configuration?

Accepted Answer

Accountability. Accountability, because an AI governance configuration that lacks clear ownership, oversight, or defined responsibility for AI decisions leaves no party answerable for the system's outcomes, which is precisely the accountability principle at risk. Accountability requires assigned roles, audit trails, and mechanisms to hold someone responsible for AI behavior, and a governance setup missing these elements directly undermines it. Transparency (A) concerns disclosure and explainability of how the AI works, Fairness (B) concerns bias and equitable treatment, and Privacy (C) concerns protection of personal data, so while each is a distinct ethical principle, none is the one most directly threatened by a governance configuration that fails to establish responsibility.

Answer

Transparency

Answer

Fairness

Answer

Privacy

Question 3

Refer to the exhibit. Which ethical principle is most at risk with this AI model configuration?

Accepted Answer

Transparency. The configuration has explainability disabled, which directly violates transparency. Therefore, Option D (Transparency) is the correct answer. Option A (Accountability) is at risk because human review is not required, but the most direct ethical risk is transparency. Option B (Fairness) is partially addressed by the fairness check, though not fully. Option C (Privacy) is not directly affected by this configuration.

Answer

Accountability

Answer

Fairness

Answer

Privacy

Question 4

A company deployed an AI chatbot to handle customer service. The chatbot sometimes generates responses that are biased against certain demographics. The company wants to mitigate this. What is the best first step?

Accepted Answer

Conduct an AI ethics audit.. The best first step is D, conduct an AI ethics audit, because the company must first systematically identify, assess, and document where and how bias is entering the chatbot's responses before choosing a remediation technique. An ethics audit evaluates training data, model behavior, deployment context, and governance controls, providing the evidence needed to select effective mitigations. Options A, B, and C do not fit as first steps: restricting the chatbot to predefined responses (A) is a blunt functional limitation rather than a bias diagnosis, increasing model complexity (B) can worsen or obscure bias, and simply removing all demographic data from training (C) is not sufficient because bias can persist through proxy variables and historical patterns in the remaining data.

Answer

Restrict chatbot to only predefined responses.

Answer

Increase model complexity.

Answer

Remove all demographic data from training.

Question 5

A sales team uses an AI tool to prioritize leads. The tool is found to give lower scores to leads from certain regions. What ethical principle is most violated?

Accepted Answer

Fairness. Fairness, because the AI tool is systematically assigning lower scores to leads from certain regions, which constitutes discriminatory or biased treatment across groups and directly violates the ethical principle of fairness (often framed as justice, equity, or non-discrimination in AI ethics). Fairness requires that automated decisions not disadvantage individuals based on protected or irrelevant attributes such as geographic origin. Accountability (A) concerns who is responsible for the system's outcomes, privacy (B) concerns handling of personal data, and transparency (C) concerns explainability and disclosure of how the system works; none of these is the principle most directly breached by biased scoring itself.

Answer

Accountability

Answer

Privacy

Answer

Transparency

Question 6

During model development, the data scientist realizes the training data is not representative of the intended population. What should they do?

Accepted Answer

Augment data with synthetic samples for underrepresented groups.. The correct answer is D: augment data with synthetic samples for underrepresented groups. When training data is not representative of the intended population, the model will learn biased patterns and perform poorly on underrepresented groups, so adding synthetic samples (e.g., via SMOTE, data augmentation, or generative methods) helps balance the dataset and improve generalization to the target population. Option A is wrong because removing underrepresented groups worsens the representativeness problem and introduces exclusion bias. Option B is wrong because regularization addresses overfitting, not dataset imbalance or sampling bias. Option C is wrong because using unrepresentative data as-is will not magically generalize; the model will inherit the same biases present in the training set.

Answer

Remove the underrepresented groups from the scope.

Answer

Increase model regularization.

Answer

Use the data as is, as the model will generalize.

Question 7

An AI system for hiring is found to have a disparate impact on a protected class. The company is legally required to...

Accepted Answer

Conduct a bias audit and take corrective action.. The correct answer is A: Conduct a bias audit and take corrective action. When an AI hiring system produces a disparate impact on a protected class, the legally required response is to investigate the source of the bias through an audit and then remediate it, which aligns with anti-discrimination enforcement frameworks such as Title VII and emerging AI regulations. Option B is incorrect because obtaining applicant consent does not cure a discriminatory outcome or satisfy disparate-impact liability. Option C is incorrect because publishing the algorithm is a transparency measure, not a legal remedy for bias. Option D is incorrect because discontinuing the system is not the mandated first step; remediation through audit and corrective action is required.

Answer

Obtain consent from applicants.

Answer

Publish the algorithm.

Answer

Discontinue use of the system.

Question 8

A company wants to use customer data to train an AI model. Which ethical consideration is paramount?

Accepted Answer

Data minimization. Data minimization (option C) is the paramount ethical consideration here because training an AI model on customer data requires collecting and using only the personal data strictly necessary for the stated purpose, which reduces privacy risk and aligns with principles such as GDPR's purpose limitation and data minimization requirements. Model accuracy (A) is a technical performance goal, not an ethical safeguard, and pursuing it by ingesting more customer data than needed can worsen privacy harms. Cost efficiency (B) and speed of deployment (D) are operational and business concerns that do not address the ethical duty to protect customer data, so they are not the primary consideration.

Answer

Model accuracy

Answer

Cost efficiency

Answer

Speed of deployment

Question 9

A developer notices that an AI model performs differently for different age groups. What should be done?

Accepted Answer

Investigate the cause and evaluate fairness metrics.. The correct option is C: Investigate the cause and evaluate fairness metrics. When a model performs differently across age groups, this indicates potential bias or disparate impact, so the proper first step is to diagnose the root cause (e.g., data imbalance, proxy features, or label noise) and quantify the disparity using fairness metrics such as demographic parity, equalized odds, or disparate impact ratio. Only after this analysis can an informed mitigation be chosen. Option A is premature because simply adding data may not address the underlying cause and could even worsen bias. Option B is not automatically valid because removing age can still leave correlated proxy features and may reduce model utility. Option D is incorrect because high overall accuracy can mask subgroup performance gaps and does not satisfy fairness or compliance requirements.

Answer

Retrain with more data from all ages.

Answer

Remove age as a feature.

Answer

Ignore it if overall accuracy is high.

Question 10

An organization wants to implement AI in a way that builds trust. Which practice is most important?

Accepted Answer

Providing explanations for AI decisions.. Providing explanations for AI decisions increases transparency and trust. Option A is wrong because secrecy reduces trust. Option C is wrong because complexity can reduce trust if not explainable. Option D is wrong because cheap data sources may compromise data quality and ethics.

Answer

Keeping model details secret.

Answer

Using complex models for better performance.

Answer

Using the cheapest data sources.

Question 11

A company is deploying an AI system that makes recommendations to users. To ensure ethical use, they should:

Accepted Answer

Allow users to opt out and understand how decisions are made.. Allowing users to opt out and understanding decisions respects autonomy and transparency. Option B is wrong because making recommendations without oversight can lead to harmful outcomes. Option C is wrong because maximizing engagement without regard to user well-being is unethical. Option D is wrong because using only internal data may not be sufficient and could raise privacy concerns.

Answer

Make recommendations without oversight.

Answer

Maximize engagement regardless of user well-being.

Answer

Use only internal data.

Question 12

A data scientist is building a model for credit scoring. They have access to a dataset with historical bias. What should they do?

Accepted Answer

Apply fairness constraints during training.. Apply fairness constraints during training. When a dataset contains historical bias, the model can learn and perpetuate that bias, so incorporating fairness constraints (e.g., demographic parity, equalized odds) into the training objective helps mitigate disparate impact while still learning from the data. Discarding all biased variables (B) is often ineffective because bias can be encoded in many correlated features and may remove useful signal. Using the data as is (C) would perpetuate discrimination, and a more complex model (D) typically increases the capacity to fit and amplify bias rather than reduce it.

Answer

Discard all biased variables.

Answer

Use the data as is because it reflects reality.

Answer

Use a more complex model to reduce bias.

Question 13

Which TWO components are essential for an AI ethics governance framework?

Accepted Answer

Conducting regular audits. Option D (Conducting regular audits) is correct because an AI ethics governance framework requires ongoing, independent review of models, data, and outcomes to detect bias, drift, and compliance gaps, ensuring accountability over time rather than a one-time check. Option E (Establishing a code of ethics) is correct because a formal code of ethics codifies the organization's principles, values, and behavioral expectations for AI development and use, giving teams a shared standard against which decisions and audits are measured. Together, the code sets the normative rules while audits verify adherence, which is why these two are essential governance components. Option A (Using the most recent algorithms) is not essential, since newer algorithms are not inherently more ethical and governance is about process and accountability, not recency. Option B (Maximizing data collection) is not essential and can even increase privacy and bias risks, so it is not a governance requirement. Option C (Assigning an ethics officer) is valuable in practice but is not marked correct here, as the question asks for the two essential framework components of a code of ethics and regular audits.

Answer

Using the most recent algorithms

Answer

Maximizing data collection

Answer

Assigning an ethics officer

Question 14

Which TWO approaches are recommended for mitigating bias in AI models?

Accepted Answer

Re-weighting training samples. Re-weighting training samples (C) is a recommended bias-mitigation technique because it assigns higher weights to underrepresented or historically disadvantaged groups during training, so the learned loss better reflects balanced group outcomes without discarding data. Adversarial debiasing (E) is also recommended: it trains a predictor alongside an adversary that tries to infer the protected attribute from the predictor's output, penalizing the predictor when the adversary succeeds and thereby reducing disparate impact. By contrast, increasing model depth (A) only raises capacity and can amplify bias, and adding regularization (D) controls overfitting but does not specifically target group fairness. Removing all sensitive attributes (B) is not sufficient because proxy variables (e.g., ZIP code or name) can still encode protected information and reintroduce bias.

Answer

Increasing model depth

Answer

Removing all sensitive attributes

Answer

Adding regularization

Question 15

Refer to the exhibit. The model is deployed and monitoring triggers an alert for a fairness violation. What does this indicate?

Accepted Answer

The demographic parity difference exceeded 0.1.. The alert indicates that the demographic parity difference exceeded 0.1, which is the configured fairness threshold that triggers a fairness violation in Amazon SageMaker Clarify model monitoring. Demographic parity difference measures the absolute difference in positive prediction rates between groups defined by a protected attribute, and a value above 0.1 signals meaningful disparity. The other options do not fit: removing protected attributes (B) is a data-preprocessing action, not a fairness metric alert; retraining the model (C) is an operational event unrelated to the violation; and an accuracy drop below threshold (D) would trigger a model quality or drift alert, not a fairness violation.

Answer

The protected attributes were removed.

Answer

The model has been retrained.

Answer

The model's accuracy has dropped below threshold.