AI Associate Ethical Considerations of AI — Timed 100 Questions

Question 1

A healthcare organization uses Einstein Discovery to predict patient readmission risk. The model uses protected attributes like race and age as features. Which action best aligns with Salesforce's ethical AI principles?

100 minute time limit — choose an answer to begin.

Accepted Answer

Retain the features but monitor for disparate impact and ensure compliance with regulations.. Option A is correct because Salesforce's ethical AI principles emphasize responsible use of protected attributes with ongoing monitoring for disparate impact and regulatory compliance, rather than assuming that simply removing such fields guarantees fairness. In healthcare, race and age can be clinically relevant predictors of readmission risk, so retaining them while auditing outcomes and bias is the most defensible approach. Option B does not fit because removing race and age entirely can reduce model accuracy and may not eliminate bias, since proxy variables can still encode protected information. Option C is insufficient because bucketing age only coarsens the feature and does not address disparate impact or compliance. Option D is wrong because high accuracy alone does not satisfy ethical AI requirements when protected attributes are involved.

Answer

Remove race and age features entirely to ensure fairness.

Answer

Replace age with an age group bucket to reduce granularity.

Answer

Use the model as is because predictions are accurate.

Question 2

Which TWO practices are recommended when using AI for automated decision-making in hiring?

Accepted Answer

Regularly audit the model for adverse impact.. Option B is correct because regularly auditing the model for adverse impact is a core fairness practice in AI-assisted hiring: it detects disparate impact across protected groups (e.g., using metrics like the 4/5ths rule or selection-rate ratios) so bias can be identified and mitigated before it harms candidates. Option D is correct because incorporating human review for high-stakes decisions ensures meaningful human oversight of consequential outcomes like hiring, allowing a reviewer to catch erroneous or biased AI recommendations and preserving accountability. Option A is not recommended because making the AI model the sole decision-maker removes human oversight and accountability, which is inappropriate for high-stakes employment decisions. Option C is not recommended because using protected attributes such as race, gender, or age can introduce or amplify discriminatory bias and may violate anti-discrimination laws. Option E is not recommended because accuracy alone does not guarantee fairness; a model can be accurate overall yet still produce unlawful adverse impact on a protected group.

Answer

Use the AI model as the sole decision-maker.

Answer

Use all available data including protected attributes.

Answer

Ignore adverse impact if the model is accurate.

Question 3

An AI Associate reviews the Lead Scoring model exhibit. What is the primary ethical concern with this model?

Accepted Answer

The model has low recall, potentially missing minority class leads.. The primary ethical concern is that the model has low recall, meaning it fails to identify a significant portion of actual positive leads (the minority class). In a lead scoring context, this can result in missed business opportunities and potential bias against certain customer segments, as the model systematically overlooks valuable leads that do not fit the majority pattern.

Answer

The model uses too many features.

Answer

The model is not explainable.

Answer

The training data is imbalanced.

Question 4

A healthcare organization is deploying an AI model to predict patient readmission risk. The model was trained on historical data that underrepresented minority populations. During testing, the model shows lower accuracy for those groups. What should the data scientist do first?

Accepted Answer

Re-evaluate the training data to ensure balanced representation and consider re-sampling techniques.. The first step in addressing model bias is to audit the training data for representational imbalance. Re-evaluating the data and applying re-sampling techniques (e.g., oversampling minority groups or undersampling the majority) directly targets the root cause of the disparity—skewed class distributions—before modifying the model or its features.

Answer

Remove sensitive attributes like race and gender from the training data.

Answer

Ignore the disparity because the model's overall accuracy is acceptable.

Answer

Retrain the model with more complex algorithms to improve accuracy.

Question 5

A company is developing a chatbot for customer service. They want to ensure the bot does not generate offensive responses. Which practice should they implement?

Accepted Answer

Deploy a content filter and human-in-the-loop moderation for sensitive interactions.. Deploying a content filter combined with human-in-the-loop moderation provides both automated detection of offensive language and human oversight for ambiguous or sensitive interactions. This layered approach ensures that the chatbot can block clearly toxic outputs in real time while allowing human reviewers to handle edge cases, reducing the risk of generating offensive responses.

Answer

Use unsupervised learning to allow the bot to learn naturally from conversations.

Answer

Remove all toxicity detection to avoid false positives.

Answer

Train the model only on customer service transcripts without review.

Question 6

A credit scoring AI uses 50 features including zip code, age, and income. The model has high accuracy but denies credit disproportionately to a protected group. An audit reveals that zip code is a proxy for race. What is the best course of action?

Accepted Answer

Replace zip code with more relevant non-discriminatory features and retrain with fairness constraints.. Replacing biased proxy with more relevant features can maintain accuracy while reducing discrimination. Option A is wrong because simply removing zip code may not eliminate all proxies. Option C is wrong because retraining with same data yields same bias. Option D is wrong because transparency alone doesn't fix bias.

Answer

Remove zip code from the feature set and retrain.

Answer

Keep zip code but add a fairness penalty to the loss function.

Answer

Increase transparency by publishing the model's decision criteria.

Question 7

An AI system is used to approve loan applications. The model uses income, zip code, and credit score as features. What is a potential ethical concern?

Accepted Answer

Zip code may act as a proxy for race, leading to discrimination. Using zip code as a feature can introduce proxy discrimination. Zip codes are strongly correlated with race and socioeconomic status due to historical redlining and residential segregation. When the model learns patterns from zip code, it may inadvertently deny loans to applicants from certain racial or ethnic groups, violating fair lending laws and ethical AI principles.

Answer

The model should be a black box to avoid bias

Answer

Credit scores are rarely accurate

Answer

Income is not a reliable predictor of repayment

Question 8

A company deploys an AI recommender system that personalizes content. The system is trained on user click data. After deployment, the company notices that the system increasingly recommends sensationalist content, leading to user polarization. Which principle is being violated?

Accepted Answer

Beneficence. The recommender system's shift toward sensationalist content, which polarizes users, violates the principle of beneficence because it causes harm (user polarization) rather than promoting well-being. Beneficence requires AI systems to act in the best interests of users and society, not to optimize for engagement metrics at the expense of ethical outcomes.

Answer

Accuracy

Answer

Privacy

Answer

Transparency

Question 9

Refer to the exhibit. A team is deploying an AI model for credit scoring. The model uses a complex neural network with high accuracy. The team has performed bias testing and used a representative dataset. According to the policy, what is the MOST significant ethical gap?

Accepted Answer

The model lacks explainability, which is required by the policy. The policy requires explainability for high-stakes AI models, such as credit scoring. The model's lack of explainability is the most significant ethical gap, as it prevents stakeholders from understanding how decisions are made, which is a core ethical principle.

Answer

The training data may not be representative

Answer

Customer consent was not obtained

Answer

Bias testing was not performed

Question 10

A financial institution deploys an AI system to recommend investment portfolios to retail clients. The system uses reinforcement learning to maximize returns based on client risk profiles. After six months, an internal audit reveals that the system has been consistently recommending high-risk, high-commission products to elderly clients with low risk tolerance, resulting in significant financial losses for those clients. The system's training data included historical transactions, which showed that elderly clients were less likely to complain or switch advisors. The institution's AI ethics policy mandates fairness, transparency, and accountability. The system currently provides no explanations for its recommendations, and there is no human oversight process. The compliance team needs to remediate the situation. Which course of action BEST addresses the ethical violations?

Accepted Answer

Retrain the model on a balanced dataset, implement explainability features, and require human approval for high-risk recommendations to elderly clients.. It directly addresses the root cause of the ethical violations: biased training data (historical transactions where elderly clients were less likely to complain) and lack of transparency. Retraining on a balanced dataset mitigates the reinforcement learning model's exploitation of that bias, while explainability features (e.g., SHAP values or LIME) and human-in-the-loop approval for high-risk recommendations ensure accountability and fairness as mandated by the AI ethics policy.

Answer

Disable the AI system and revert to manual portfolio management.

Answer

Add a disclaimer to all recommendations stating that past performance does not guarantee future results.

Answer

Adjust the model to lower the risk threshold for all clients.

Question 11

A healthcare provider uses an AI system to predict patient readmission risk. The system was trained on historical data from the past five years, during which the hospital served a predominantly urban population. Recently, the hospital expanded to rural areas with different demographic and socioeconomic profiles. The AI predictions have been less accurate for rural patients, leading to misallocation of care resources. The AI Ethics committee is reviewing the system for potential bias. The model outputs a risk score from 0 to 100. The data science team has identified that the model uses features such as income, distance from hospital, and insurance type, which may correlate with race and socioeconomic status. The team wants to make the model fairer without retraining from scratch. Which approach best balances fairness and predictive accuracy?

Accepted Answer

Apply a post-processing calibration that adjusts the risk score thresholds separately for urban and rural populations to achieve equal false positive rates.. Post-processing calibration adjusts the decision thresholds for each subgroup (urban vs. rural) to equalize a fairness metric (e.g., false positive rate) without modifying the underlying model. This approach preserves the predictive signal from the original features while directly addressing the bias caused by distribution shift, making it the most practical solution when retraining from scratch is not feasible.

Answer

Remove the features income, distance, and insurance type from the model and retrain.

Answer

Continue using the current model but add a disclaimer that predictions may be less accurate for rural patients.

Answer

Retrain the model using only the latest year of data that includes rural patients.

Question 12

A financial services company deploys an AI system to approve small business loans. The system uses a deep neural network trained on historical loan data. After deployment, an internal audit reveals that the approval rate for minority-owned businesses is 15% lower than for non-minority-owned businesses with similar financial profiles. The company's AI Ethics policy requires that AI systems be fair and transparent. The data science team has access to the training data, model architecture, and feature importance scores. The company wants to understand why the disparity exists and take corrective action. Which approach should the team take first?

Accepted Answer

Analyze the training data to determine if there is sampling bias or labeling bias that caused the model to associate minority ownership with higher risk.. The first step in diagnosing an AI fairness issue is to audit the training data for biases such as sampling bias (e.g., underrepresentation of minority-owned businesses) or labeling bias (e.g., historical loan officers unfairly labeling minority applicants as higher risk). Since the team has access to the training data, analyzing it directly addresses the root cause of the disparity before making model-level changes. This aligns with the AI Ethics policy requirement for transparency, as data bias is a common source of unfair outcomes in deep neural networks trained on historical data.

Answer

Apply a disparate impact analysis to quantify the adverse impact and then adjust the decision threshold.

Answer

Examine the model's weights and activations to identify which features contribute to the disparity.

Answer

Retrain the model with a fairness constraint that penalizes disparities in approval rates.

Question 13

An AI model predicts employee performance. The HR team uses it to identify high-potential employees. What is a potential ethical risk?

Accepted Answer

All of the above. All three listed risks—over-reliance on the model, privacy violation, and underutilization of human judgment—are potential ethical risks when an AI model predicts employee performance. Over-reliance can lead to automated decisions without human oversight, privacy violation may occur if sensitive employee data is mishandled, and underutilization of human judgment ignores contextual factors that the model cannot capture. Together, these represent a comprehensive set of ethical concerns in AI-driven HR practices.

Answer

Over-reliance on the model

Answer

Privacy violation

Answer

Underutilization of human judgment

Question 14

Refer to the exhibit. Based on the JSON policy for AI fairness checks, which fairness metric is NOT enabled?

Accepted Answer

Equal opportunity. (Equal opportunity) is correct because the JSON policy shown in the exhibit configures fairness checks for demographic parity, disparate impact, and equalized odds, but does NOT include the equal opportunity metric. Equal opportunity requires equal true positive rates across groups, which is a separate metric from equalized odds and must be explicitly enabled in the policy definition.

Answer

Demographic parity

Answer

All are enabled

Answer

Disparate impact

Question 15

Refer to the exhibit. An AI model's accuracy is shown for four demographic groups. Which group should be investigated for potential bias?

Accepted Answer

Group Beta. Group Beta is correct because it shows a significant deviation in accuracy compared to the other groups, indicating potential bias. Bias in AI models often manifests as disparate performance across demographic groups, so the group with the most divergent accuracy should be investigated first.

Answer

Group Alpha

Answer

Group Gamma

Answer

Group Delta

Question 16

Refer to the exhibit. This JSON snippet is from the Einstein Trust Layer configuration. What is the purpose of this configuration?

Accepted Answer

To detect biased predictions based on gender and race. The Einstein Trust Layer includes bias detection and audit capabilities that surface when model outputs or training data reflect protected attributes such as gender or race. The configuration shown enables monitoring and flagging of biased predictions so they can be reviewed, not blocking or removing the attributes. This supports responsible AI governance and regulatory compliance.

Answer

To block all predictions involving gender or race

Answer

To anonymize gender and race data

Answer

To remove gender and race from the model

Question 17

Refer to the exhibit. A Salesforce developer configures the Einstein Trust Layer as shown. What is the primary purpose of enabling data masking?

Accepted Answer

To anonymize personally identifiable information (PII) in the model output.. Enabling data masking in the Einstein Trust Layer ensures that personally identifiable information (PII) is anonymized before the model output is returned to the user. This protects sensitive data from exposure in AI-generated responses, which is a core requirement for privacy compliance and responsible AI use.

Answer

To improve the accuracy of sentiment analysis.

Answer

To reduce latency of the AI response.

Answer

To comply with Salesforce's service-level agreement.

Question 18

According to Salesforce's AI ethics principles, which three pillars should guide the development of AI applications?

Accepted Answer

Accountability.. Salesforce's AI ethics principles are built around five pillars, and among the options listed, Accountability (A), Transparency (B), and Fairness (E) are three of them. Accountability (A) is correct because Salesforce requires that organizations designate responsible parties who can explain and answer for AI-driven decisions and outcomes. Transparency (B) is correct because Salesforce mandates that AI systems be open about how they work, what data they use, and how they reach decisions, so users can understand and trust them. Fairness (E) is correct because Salesforce commits to designing AI that avoids bias and treats all users equitably, promoting inclusive and unbiased outcomes. Profitability (C) and Scalability (D) are not among Salesforce's AI ethics pillars; while commercially relevant, they are business goals rather than ethical principles guiding responsible AI development.

Answer

Profitability.

Answer

Scalability.

Question 19

Refer to the exhibit. An admin sees this error in the Einstein activity log. What is the most likely cause?

Accepted Answer

Data masking is configured but not applied to the sentiment analysis model.. The error indicates that sentiment analysis is not enabled, but PII is leaking, likely because data masking is not applied to the model generating the output.

Answer

Sentiment analysis is generating PII that data masking cannot hide.

Answer

The model output should not contain any text.

Answer

The Einstein Trust Layer is completely disabled.

Question 20

To comply with Salesforce's AI ethics principles when using Einstein Bots, which two practices should be implemented?

Accepted Answer

Allow users to escalate to a human agent.. Option A is correct because Salesforce's AI ethics principles emphasize transparency and human oversight, so Einstein Bots must provide a clear path for users to escalate to a live human agent rather than trapping them in automated interactions. Option D is correct because users have the right to know they are interacting with an AI system, and disclosing bot identity directly satisfies the transparency and honesty principles. Option B is incorrect because fully autonomous decision-making without human review violates the accountability and human-centered values that Salesforce's ethical AI guidelines require. Option C is incorrect because retaining conversation transcripts indefinitely conflicts with data minimization and privacy-by-design principles, which favor limited retention periods. Option E, while a sound privacy practice, is not one of the two practices specifically called out by Salesforce's AI ethics principles for Einstein Bots in this context.

Answer

Use the bot to make all customer decisions autonomously.

Answer

Store all conversation transcripts indefinitely.

Answer

Minimize data collection to only what is necessary.