AI Associate Ethical Considerations of AI Practice Test 7 — 15 Questions

Question 1

A financial institution uses an AI model to approve credit. The model shows disparate impact against a protected group. Under Salesforce's ethical AI principles, what is the most appropriate action?

Accepted Answer

Apply a bias mitigation technique such as reweighing.. Option B is correct because bias mitigation techniques like reweighing address disparate impact while maintaining functionality. Option A is wrong because discontinuing the model may cause operational disruption and is not necessary if bias can be mitigated. Option C is wrong because adjusting the threshold may not address the root cause. Option D is wrong because publicizing predictions may violate privacy.

Answer

Discontinue the model immediately.

Answer

Increase the model's decision threshold for all applicants.

Answer

Publish the model's predictions publicly for scrutiny.

Question 2

A company uses Einstein's predictive lead scoring. The model inadvertently overweights leads from certain geographic regions. Which action aligns with Salesforce's Responsible AI principles?

Accepted Answer

Implement a feedback loop to continuously monitor and adjust.. Option B is correct because continuous monitoring and adjustment is a key component of responsible AI. Option A is wrong because simply removing the geographic feature may not eliminate proxy variables. Option C is wrong because ignoring bias is unethical. Option D is wrong because restricting usage may cause inequity.

Answer

Remove the geographic feature from the model.

Answer

Use the model as-is because it improves overall accuracy.

Answer

Only use the model for regions where it performs well.

Question 3

An organization is deploying an AI system for loan decisions. They want to ensure human oversight. Which is the best implementation?

Accepted Answer

The system only flags edge cases for human review.. Option D is correct because flagging edge cases for human review efficiently focuses oversight on the most uncertain or risky decisions. Option A is wrong because post-hoc review for high-value loans may miss issues in other cases. Option B is wrong while thorough, it may be too slow for low-risk decisions. Option C is wrong because full autonomy reduces human involvement.

Answer

The system operates fully autonomously but logs decisions for audit.

Answer

The system makes decisions automatically, with post-hoc review only for high-value loans.

Answer

The system provides recommendations, and a human must approve all decisions.

Question 4

Which TWO practices contribute to ethical AI transparency?

Accepted Answer

Providing user-facing explanations.. Options A and C are correct. Documenting model decisions provides an audit trail, and providing user-facing explanations helps users understand AI behavior. Option B is wrong because open-source algorithms do not guarantee transparency about specific model decisions. Option D is wrong because collecting sensitive data may violate privacy principles. Option E is wrong because opt-out is about user control, not transparency.

Answer

Using open-source algorithms.

Answer

Collecting sensitive demographic data.

Answer

Allowing users to opt out of AI processing.

Question 5

Which THREE actions align with Salesforce's responsible AI principles?

Accepted Answer

Involving diverse stakeholders in AI development.. Options A, C, and E are correct. Monitoring for bias (A) ensures ongoing fairness, involving diverse stakeholders (C) reduces blind spots, and human oversight (E) ensures accountability. Option B is wrong because restricting data sources may limit representativeness. Option D is wrong because accuracy should not override fairness.

Answer

Using only internal data sources.

Answer

Prioritizing model accuracy over fairness.

Question 6

Which TWO are best practices for mitigating bias in AI models?

Accepted Answer

Using balanced training datasets.. Options A and C are correct. Using balanced training datasets helps reduce representation bias, and applying disparity analysis identifies where bias exists. Option B is wrong because removing sensitive attributes alone may not eliminate proxy variables. Option D is wrong because using only recent data may still contain bias. Option E is wrong because complex deep learning models can be harder to audit and may amplify bias.

Answer

Using complex deep learning models.

Answer

Removing all sensitive attributes.

Answer

Training models on the most recent data only.

Question 7

Refer to the exhibit. A Salesforce admin is reviewing an AI model's fairness report. Which action should the admin take?

Accepted Answer

Retrain the model because the equal opportunity score is below threshold.. Option B is correct because the equal opportunity score (0.72) is below the bias threshold (0.8), indicating potential unfairness in true positive rates across groups. Option A is wrong because demographic parity is above threshold but equal opportunity is not, so not all metrics exceed threshold. Option C is wrong because removing features may not address the root cause. Option D is wrong because increasing the threshold would mask the problem.

Answer

Remove the email_engagement feature to improve fairness.

Answer

Increase the bias threshold to 0.9.

Answer

Deploy the model because all metrics exceed the threshold.

Question 8

Refer to the exhibit. An administrator runs an audit on a sentiment analysis model. What is the primary ethical concern?

Accepted Answer

The model has low accuracy on negative text, indicating potential bias.. Option B is correct because the large accuracy gap (98% vs 80%) indicates the model performs poorly on negative text, which could lead to unfair treatment of users expressing negative sentiment. Option A is wrong because overfitting is not evident from this data. Option C is wrong even though bias score is below threshold, the significant accuracy disparity is a concern. Option D is wrong while data imbalance is present, the disparate performance is the more direct ethical issue.

Answer

The bias score is below threshold so no action needed.

Answer

The training data is imbalanced.

Answer

The model is overfitting to positive text.

Question 9

Refer to the exhibit. A developer receives this fairness check error. What is the most likely cause?

Accepted Answer

The model has higher false positive and false negative rates for Group B.. Option B is correct because the exhibit clearly shows significantly higher false positive and false negative rates for Group B, indicating the model treats groups differently. Option A is wrong because overfitting is not indicated by these metrics. Option C is wrong because the recommendation is to fix the issue, not that the model is already fair. Option D is wrong because Group A has lower rates, suggesting Group B's data may be insufficient, but the cause is the disparity itself.

Answer

The error is due to insufficient training data for Group A.

Answer

The model is overfitting.

Answer

The recommendation suggests reweighting, so the model is already fair.

Question 10

A company uses Salesforce Einstein to build an AI model that predicts customer churn. The model is trained on historical data from the past two years. During testing, the model shows significantly higher accuracy for male customers compared to female customers. What is the most ethical course of action?

Accepted Answer

Investigate the cause of the disparity, retrain the model with more representative data, and re-evaluate fairness.. Option B is correct because investigating and retraining the model to reduce bias is the ethical approach. Option A is wrong because ignoring the disparity could lead to unfair treatment. Option C is wrong because immediately deploying without correction is irresponsible. Option D is wrong because manually adjusting predictions introduces new biases.

Answer

Deploy the model but add a disclaimer that it may be less accurate for female customers.

Answer

Deploy the model as is, because it still meets the overall accuracy threshold.

Answer

Manually adjust the model's output to ensure equal churn predictions across genders.

Question 11

A retail company wants to use Einstein AI to personalize marketing offers. They plan to include customer purchase history and demographic data. What is the essential first step to ensure ethical use of customer data?

Accepted Answer

Obtain explicit consent from customers for using their data in AI personalization.. Option A is correct because obtaining explicit consent aligns with privacy ethics and regulations like GDPR. Option B is wrong as anonymization alone does not replace consent. Option C is wrong because it bypasses transparency. Option D is wrong because consent is required even if data is already collected.

Answer

Only use data from customers who have not opted out of marketing.

Answer

Proceed with available data since it was collected for business purposes.

Answer

Anonymize all customer data before feeding it into the model.

Question 12

A financial institution deploys an AI model to approve loan applications. The model uses features like income, credit score, and postal code. An audit reveals that the model denies loans at a higher rate for applicants in certain postal codes, which correlate with minority neighborhoods. What should the institution do to align with ethical AI principles?

Accepted Answer

Remove postal code from the model and retrain using only non-biased features.. Option D is correct because removing the biased feature and using alternative features is a direct way to mitigate bias. Option A is wrong as it does not address the root cause. Option B is wrong because it only adds transparency, not fairness. Option C is wrong because disclosure does not fix the bias.

Answer

Continue using the model but monitor the denial rates monthly.

Answer

Add an explanation to applicants in affected areas about why they were denied.

Answer

Publish a report on the model's disparate impact and accept it as a business risk.

Question 13

A healthcare provider uses Einstein's Prediction Builder to predict patient readmission risk. The model outputs a risk score, but clinicians do not understand how the score is calculated. According to ethical AI principles, what should the provider implement?

Accepted Answer

Integrate model explanations using Einstein's Explainability feature to show key factors influencing each prediction.. Option B is correct because explainability is a key ethical principle, and providing interpretable insights builds trust. Option A is wrong because ignoring the lack of understanding is risky. Option C is wrong because it only provides training to a few. Option D is wrong because simplifying the model may reduce accuracy, but the core issue is explainability, not complexity.

Answer

Proceed with the model since it provides accurate predictions.

Answer

Train only the data science team on the model's inner workings.

Answer

Replace the model with a simpler, less accurate but fully transparent model.

Question 14

A company uses an AI chatbot that automatically responds to customer service inquiries. When a customer questions the bot's response, there is no mechanism for the customer to appeal or speak to a human. What ethical principle is being violated?

Accepted Answer

Accountability. Option D is correct because human oversight and the ability to escalate to a human is essential for accountability. Option A is wrong because privacy may be fine. Option B is wrong because the bot may be accurate but lack fairness due to no recourse. Option C is wrong because the bot may be transparent about what it does.

Answer

Privacy

Answer

Transparency

Answer

Fairness

Question 15

A company is building an AI model to score sales leads. They have a dataset with historical leads, including whether they converted. The dataset contains 90% male and 10% female leads. The model will be used to prioritize leads for sales follow-ups. What is the primary ethical concern?

Accepted Answer

The training data is imbalanced, which may cause the model to be less accurate for female leads, leading to unfair prioritization.. Option A is correct because the imbalance can lead to the model performing poorly for female leads, causing gender bias. Option B is wrong because the concern is not that accuracy is already biased, but that training data is biased. Option C is wrong because the dataset size is not the main issue; the imbalance is. Option D is wrong because using historical data is not inherently unethical, but the imbalance is problematic.

Answer

The model will be biased against male leads because they are overrepresented.

Answer

The dataset is too small to build a reliable model.

Answer

Using historical data is unethical because it may not reflect current conditions.