AI Associate Ethical Considerations of AI — 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?

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

Accepted Answer

Retain the features but monitor for disparate impact and ensure compliance with regulations.. Retaining protected attributes like race and age allows the model to leverage potentially legitimate medical correlations, but it requires ongoing monitoring for disparate impact to ensure fairness and compliance with regulations like HIPAA. This aligns with Salesforce's ethical AI principles of transparency, accountability, and fairness. Removing features entirely (Option B) may lose important medical context, while bucketing age (Option C) still risks bias and loss of granularity. Using the model as-is (Option D) ignores ethical responsibilities.

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

An AI Associate reviews the bot configuration and test results. Which action best addresses the ethical issue?

Accepted Answer

Add Spanish language support with separate sentiment model and intents.. The ethical issue is that the bot performs poorly for Spanish-speaking customers, leading to unfair treatment. Adding Spanish language support with a separate sentiment model and intents directly addresses this disparity by ensuring the bot can understand and respond appropriately to Spanish speakers. Option A (increasing fallback threshold) does not resolve the language-specific performance gap. Option B (collecting more Spanish data but retraining the English model) is ineffective because the English model is not suited for Spanish. Option C (disabling sentiment analysis for non-English conversations) avoids the problem rather than solving it. Therefore, Option D is the best ethical action.

Answer

Increase the fallback threshold for all languages.

Answer

Collect more data from Spanish-speaking customers and retrain the English model.

Answer

Disable sentiment analysis for non-English conversations.

Question 3

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?

Accepted Answer

Re-sample the training data to include more female candidates and use fairness-aware algorithms.. The most appropriate mitigation because it addresses the root cause of bias by ensuring the training data is representative and using algorithms designed to detect and correct unfairness. Re-sampling provides a more balanced dataset, and fairness-aware algorithms actively work to prevent discriminatory outcomes. Option B (post-processing) can help but may not be sufficient alone and does not fix the underlying data bias. Option C (accepting bias) ignores the ethical responsibility. Option D (removing gender) is ineffective because proxy variables like job history or education can still correlate with gender and perpetuate bias.

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 4

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 5

A company uses an AI model to screen job candidates. They discover the model is rejecting candidates from certain zip codes. What should they do first?

Accepted Answer

Audit training data for bias. Auditing the training data for bias is the first step to identify and mitigate unfairness. Option A is wrong because increasing model complexity may exacerbate bias. Option B is wrong because adding more features without bias analysis could introduce more bias. Option C is wrong because removing the zip code feature alone does not address underlying bias in other correlated features.

Answer

Increase model complexity

Answer

Add more features to the model

Answer

Remove zip code feature

Question 6

A Salesforce admin wants to use Einstein Prediction Builder to predict customer churn. What ethical consideration is most important?

Accepted Answer

Data privacy. Data privacy is the most important ethical consideration when using customer data for predictions like churn. Option A is wrong because cost of implementation is a business concern, not an ethical one. Option B is wrong because while model accuracy is important, it is a technical metric and not the primary ethical concern; privacy takes precedence. Option D is wrong because transparency of predictions is important for trust, but data privacy is more foundational when handling sensitive customer data.

Answer

Cost of implementation

Answer

Model accuracy

Answer

Transparency of predictions

Question 7

A company deploys an AI chatbot for customer service. After training on historical chats, the chatbot frequently gives incorrect answers to minority language queries. What is the likely cause?

Accepted Answer

Data imbalance. Data imbalance. When training data has an overrepresentation of majority language queries and very few minority language examples, the AI model fails to learn patterns for minority languages, leading to incorrect answers on those queries. Option A (Underfitting) would cause poor performance across all data, not just minority languages. Option C (Lack of compute resources) might affect model training speed or scale but does not directly cause bias toward specific languages. Option D (Overfitting) would result in memorizing training data, causing poor generalization on new data overall, not selectively on minority languages.

Answer

Underfitting

Answer

Lack of compute resources

Answer

Overfitting

Question 8

An organization uses Einstein Recommendation Builder to suggest products. They want to ensure recommendations are fair across demographics. Which action should they take?

Accepted Answer

Regular bias audits. Regular bias audits help detect and mitigate unfairness over time, ensuring recommendations are fair across demographics. Option A is wrong because using only demographic features could lead to stereotype-based recommendations. Option B is wrong because removing all user data would eliminate personalization. Option D is wrong because increasing model depth does not address fairness.

Answer

Use only demographic features

Answer

Remove all user data

Answer

Increase model depth

Question 9

A healthcare company uses AI to predict patient readmission rates. What is a critical ethical requirement?

Accepted Answer

Explanation of predictions to doctors. Explainability is a critical ethical requirement in healthcare AI. Doctors need to understand the reasoning behind predictions to ensure patient safety, maintain trust, and fulfill accountability obligations. Options B and C are performance or technical requirements, not ethical ones, and Option D raises privacy concerns rather than addressing ethical transparency.

Answer

Low latency

Answer

High precision

Answer

Use of external data sources

Question 10

A financial institution uses AI for loan approvals. They notice the model is denying loans to women more often. After retraining with balanced data, the disparity persists. What is the next best step?

Accepted Answer

Use adversarial debiasing technique. Adversarial debiasing is a technique that explicitly reduces bias by training a model to prevent discrimination while maintaining accuracy. Option B (removing the gender feature) is not sufficient because other features may correlate with gender and perpetuate bias. Option A (increasing regularization) helps with overfitting, not bias. Option C (reducing complexity) may not address the root cause of bias.

Answer

Increase model regularization

Answer

Remove the gender feature

Answer

Reduce model complexity

Question 11

A company uses AI to monitor employee productivity. Employees feel surveilled. What ethical principle is being violated?

Accepted Answer

Transparency. The scenario describes employees feeling surveilled without awareness, which directly violates the ethical principle of transparency. Transparency requires that organizations clearly inform individuals about the collection, use, and purpose of AI monitoring. Option A (Inclusion) is about ensuring diverse representation and fair access, not about notification. Option B (Accountability) refers to taking responsibility for outcomes, not about disclosure. Option C (Safety) focuses on preventing harm, not on informing employees. Therefore, D is the correct answer.

Answer

Inclusion

Answer

Accountability

Answer

Safety

Question 12

An AI system for medical diagnosis is trained on data from one region. When deployed globally, it performs poorly. This is an issue of?

Accepted Answer

Generalizability. Generalizability refers to a model's ability to perform well on new, unseen data from different distributions. In this case, the model trained on data from one region fails to generalize globally, indicating poor generalizability. Option A (Accountability) is incorrect because accountability deals with responsibility for outcomes, not performance issues. Option B (Overfitting) is incorrect because overfitting would cause poor performance on new data from the same region, not specifically across regions. Option C (Privacy violation) is incorrect because privacy concerns data protection, not model performance across regions.

Answer

Accountability

Answer

Overfitting

Answer

Privacy violation

Question 13

Which TWO actions align with ethical AI practices in Salesforce?

Accepted Answer

Monitoring models for bias. The correct answers are C and E. Monitoring models for bias (C) is a core ethical practice to ensure fairness and prevent discrimination. Providing explanations for AI predictions (E) ensures transparency and helps build trust with users. Option A is incorrect because using only internal data does not inherently violate ethics; the issue is using data without consent. Option B is wrong because automating all decisions without human review removes necessary oversight, which can lead to unethical outcomes. Option D is incorrect because using customer data without explicit consent violates privacy principles and regulations like GDPR.

Answer

Using only internal data for training

Answer

Automating all decisions without human review

Answer

Using customer data without explicit consent

Question 14

A company wants to ensure their AI is fair. Which TWO steps are appropriate?

Accepted Answer

Test model performance on different demographic groups. Options D and E are correct. Testing the model on different demographic groups helps identify disparities, and involving diverse stakeholders brings multiple perspectives to address bias. Option A is wrong because a single fairness metric is insufficient to capture all aspects of fairness. Option B is wrong because deploying quickly without proper evaluation can lead to unfair outcomes. Option C is wrong because removing sensitive attributes may not eliminate bias due to proxy features.

Answer

Use a single fairness metric to evaluate the model

Answer

Deploy the model quickly to gather real-world data

Answer

Remove all sensitive attributes from the data

Question 15

Refer to the exhibit. A company uses this policy for a customer-facing AI model. What is the most critical ethical risk?

Accepted Answer

Lack of transparency. Lack of transparency. The policy explicitly states 'explainability: none', which means the AI model's decisions cannot be explained or understood by users. This lack of transparency prevents users from knowing why a decision was made, challenging the decision, or trusting the system, making it the most critical ethical risk. Option A (Performance) is not indicated in the policy and is less critical than transparency. Option B (Data privacy) is important but the policy does not mention any privacy violations. Option C (Model accuracy) is not addressed in the policy and while important, transparency is more fundamental for ethical AI.

Answer

Performance

Answer

Data privacy

Answer

Model accuracy

Question 16

Refer to the exhibit. What action should be taken?

Accepted Answer

Retrain the model with balanced training data. Retraining the model with balanced training data helps reduce disparate impact by ensuring the model learns from a representative sample of all groups. Option A is wrong: increasing the accuracy threshold does not address the underlying bias in the data or model. Option B is wrong: removing the model is too drastic without first attempting mitigation strategies like retraining. Option C is wrong: deploying the model as is with a disparate impact ratio of 0.6 is likely unethical and may violate fairness principles.

Answer

Increase the accuracy threshold

Answer

Remove the model

Answer

Deploy as is

Question 17

Refer to the exhibit. Which ethical principle is violated?

Accepted Answer

Transparency. The AI fails to provide an explanation for the decision, violating transparency. Option A is incorrect because safety is not directly relevant to the lack of explanation. Option B is incorrect because fairness is not the primary issue here. Option C is incorrect because accountability relates to responsibility, but the core issue is the absence of transparency.

Answer

Safety

Answer

Fairness

Answer

Accountability

Question 18

A company uses an AI model to screen job applications. They discover the model is less likely to recommend female candidates. What should the company prioritize first?

Accepted Answer

Implement a fairness metric to evaluate bias.. The first step when bias is suspected is to measure and quantify the bias using fairness metrics before taking any corrective action. Option A is wrong because simply removing gender-related features may not eliminate proxy bias from other correlated features. Option B is wrong because training only on male candidates would exacerbate bias and is unethical. Option D is wrong because increasing training data size does not necessarily address the underlying biased patterns and could even reinforce them.

Answer

Remove gender-related features from the model.

Answer

Re-train the model with only male candidates.

Answer

Increase the model's training data size.

Question 19

A company deploys an AI system that recommends loan amounts. They want to ensure explainability. Which approach best aligns with ethical AI?

Accepted Answer

Provide loan recommendations with a detailed rationale.. Providing loan recommendations with a detailed rationale supports explainability and transparency, which are key ethical AI principles. Option A is wrong because allowing the system to update features dynamically without documentation undermines traceability and accountability. Option B is wrong because restricting use to internal employees does not address the need for explainability to stakeholders or regulators. Option D is wrong because black-box neural networks lack interpretability, making it difficult to explain decisions.

Answer

Allow the system to update its features dynamically without documentation.

Answer

Restrict the system's use to only internal employees.

Answer

Use a black-box neural network for highest accuracy.

Question 20

An organization uses Einstein to predict sales opportunities. They notice the model performs poorly for small businesses. What is the most ethical approach?

Accepted Answer

Retrain the model with more data from small businesses.. Retraining the model with more data from small businesses addresses the data imbalance and improves fairness, which is the most ethical approach. Option A is wrong because reducing model complexity does not address the root cause of poor performance for small businesses—lack of representative data. Option B is wrong because using a completely different model for small businesses could lead to inconsistent user experiences and increased complexity, and it may not be necessary if the existing model can be improved with better data. Option D is wrong because ignoring small businesses is unethical, as it perpetuates bias and may violate principles of fairness and inclusivity.

Answer

Reduce the model's complexity to improve performance.

Answer

Use a completely different model for small businesses.

Answer

Ignore small businesses as they contribute little revenue.