AI Associate Ethical Considerations of AI Practice Test 13 — 15 Questions

Question 1

A healthcare company uses an AI model built on Salesforce to predict patient readmission risk. The model is trained on historical data that underrepresents certain ethnic groups. During testing, the model shows significantly higher false negative rates for those groups, meaning it fails to flag high-risk patients. The ethical concern is most directly related to which AI principle?

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

Accepted Answer

Fairness. Fairness, because the model's significantly higher false negative rates for underrepresented ethnic groups mean it delivers unequal outcomes across demographic groups, which is the definition of bias and a fairness violation in AI ethics. Fairness specifically addresses disparate performance or treatment of protected groups, and here the harm is that high-risk patients in those groups are not flagged. Privacy (A) concerns data protection and consent, not unequal model errors. Accountability (B) concerns who is responsible for AI outcomes, and Transparency (C) concerns explainability and disclosure, neither of which directly describes the biased error-rate disparity.

Answer

Privacy

Answer

Accountability

Answer

Transparency

Question 2

A financial services firm deploys an Einstein AI chatbot that provides investment advice. A customer asks why a particular recommendation was made. The chatbot is unable to provide any reasoning. Which ethical principle is most directly violated?

Accepted Answer

Transparency. Transparency, because the scenario centers on a customer's inability to understand why an AI system made a specific investment recommendation. Transparency in AI ethics requires that systems provide understandable explanations of their decisions and reasoning, especially in high-stakes domains like financial advice. Since the chatbot cannot explain its recommendation, it directly violates this principle. Privacy (A) concerns data protection and consent, which are not at issue here, while Reliability (B) concerns consistent and accurate performance, and Human oversight (D) concerns human review or control of AI decisions, neither of which is the primary failure described.

Answer

Privacy

Answer

Reliability

Answer

Human oversight

Question 3

A Salesforce admin is configuring an AI model to automatically approve customer refunds under $50. To ensure ethical use, what is the most important action?

Accepted Answer

Implement a human review process for all AI decisions. Option B is correct because implementing a human review process for all AI decisions ensures meaningful human oversight of automated refund approvals, which is the core ethical safeguard for consequential decisions affecting customers. This aligns with ethical AI principles of accountability and transparency, particularly since refund approvals have financial and customer-relationship impacts. Option A does not fit because training on all historical data may perpetuate bias and does not itself provide ethical oversight. Option C is wrong because raising the threshold to $100 expands automated authority without any ethical control. Option D is wrong because disabling logging reduces auditability and accountability, undermining rather than supporting ethical use.

Answer

Train the model on all historical refund data to maximize accuracy

Answer

Set the approval threshold to $100 to cover more cases

Answer

Disable logging to protect customer privacy

Question 4

A retail company uses Einstein to personalize product recommendations. The AI model is trained on customer purchase data that includes sensitive attributes like race and gender. The company wants to ensure ethical use. Which action would best address fairness concerns?

Accepted Answer

Remove race and gender fields from the training dataset. Removing sensitive attributes like race and gender from the training dataset helps prevent the model from directly using these factors to discriminate, thereby reducing fairness concerns. Option B is incorrect because obtaining consent does not address bias; it only addresses privacy. Option C is incorrect because adding more demographic data could amplify bias if that data is correlated with sensitive attributes. Option D is incorrect because randomizing recommendations sacrifices personalization and does not correct underlying bias in the model.

Answer

Obtain explicit consent from customers for data use

Answer

Add more demographic data to improve model accuracy

Answer

Randomize recommendations to ensure equal treatment

Question 5

A company deploys an AI system to screen job applications. The system is found to consistently reject candidates from a particular university, even though those candidates are qualified. What is the most ethical first step?

Accepted Answer

Investigate the training data and model for bias. Investigate the training data and model for bias. When an AI screening system systematically rejects qualified candidates from a specific university, the most ethical first step is to determine the root cause by auditing the training data and model for bias, since historical hiring data or skewed features can encode discriminatory patterns. This diagnostic step is necessary before any remediation, ensuring fixes are evidence-based rather than arbitrary. Option A is unethical because it deliberately introduces more bias to balance outcomes. Option B dismisses a clear pattern of disparate impact without evidence, which is irresponsible. Option C changes criteria without understanding the cause, potentially masking or worsening the underlying bias.

Answer

Increase rejections from other universities to balance

Answer

Ignore the finding as correlation, not causation

Answer

Change the screening criteria to include more universities

Question 6

A Salesforce developer is building an AI model to predict customer churn. What is the most important ethical consideration when collecting training data?

Accepted Answer

Collect only data that is necessary for the prediction. Option C is correct because data minimization is a core ethical and privacy principle: collecting only data necessary for the churn prediction reduces privacy risk, limits potential bias, and supports compliance with regulations like GDPR. In a Salesforce context, this means restricting fields to those genuinely relevant to churn rather than harvesting every available attribute. Option A is wrong because adding as many features as possible increases privacy exposure and can introduce noise or bias without improving the model. Option B is wrong because limiting data to the last month may omit meaningful historical patterns needed for accurate churn prediction. Option D is wrong because using as much historical data as possible conflicts with data minimization and may retain outdated or unnecessary personal information.

Answer

Include as many features as possible

Answer

Focus on data from the last month for relevance

Answer

Use as much historical data as possible

Question 7

A company uses an AI model to automate customer service responses. A customer receives an incorrect response that results in a financial loss. Who is primarily accountable for this error?

Accepted Answer

The organization that deployed the AI. The organization that deployed the AI, because accountability for an AI system's outcomes rests with the entity that chose to put it into production and use it in customer-facing operations. Under governance frameworks such as the NIST AI RMF and OECD AI Principles, the deploying organization owns the risk, must ensure human oversight, and bears responsibility for harms like financial loss caused by automated responses. The AI model (A) is a tool and cannot hold legal or ethical accountability, and the customer (D) is the harmed party, not the responsible one. The developer (B) may share some technical responsibility for design or defects, but primary accountability for how the model is used and its business impact lies with the deployer.

Answer

The AI model

Answer

The developer who built the model

Answer

The customer who received the response

Question 8

A bank uses Einstein to approve loan applications. The model is trained on data that includes zip codes. Analysis shows that applicants from low-income zip codes are disproportionately rejected, even when their credit profiles are similar. What is the most likely ethical issue?

Accepted Answer

Use of a proxy for protected attributes. Use of a proxy for protected attributes, because zip codes correlate strongly with race and socioeconomic status, so using them as a model feature causes disparate impact even though the protected attribute itself is not explicitly included. This is a classic redlining-style bias, where the model learns to reject applicants from low-income zip codes despite similar credit profiles. Option A does not fit because using zip codes is not inherently a privacy violation in this scenario. Option C does not fit because the issue is biased outcomes, not explainability. Option D does not fit because the model is performing as trained; the problem is fairness, not accuracy.

Answer

Privacy violation due to zip code use

Answer

Lack of transparency in decision-making

Answer

Unreliable model performance

Question 9

A company is implementing an AI system to recommend marketing campaigns. To align with Salesforce's ethical AI principles, which practice is most important?

Accepted Answer

Ensure a human reviews AI-generated campaigns before sending. Option A is correct because Salesforce's ethical AI principles emphasize human oversight and accountability, so having a human review AI-generated campaigns before sending ensures responsible use and prevents unintended or biased outputs. This aligns with the principle of keeping humans in control of consequential decisions. Option B is incorrect because targeting only high-value customers is a business segmentation choice, not an ethical AI safeguard, and could even raise fairness concerns. Option C is incorrect because optimizing subject lines for open rates is a performance tactic, not an ethical governance practice. Option D is incorrect because removing human involvement directly contradicts the need for human oversight and accountability in ethical AI.

Answer

Target only high-value customers

Answer

Maximize open rates by optimizing subject lines

Answer

Remove human involvement to speed up campaigns

Question 10

Which TWO actions are essential for ensuring transparency in an AI system? (Choose two.)

Accepted Answer

Log all AI decisions and allow audit. Option B is correct because transparency requires that an AI system's decisions be traceable and reviewable; logging all AI decisions and enabling audit creates an immutable record that lets stakeholders verify how and why outputs were produced. Option D is correct because providing clear explanations for AI decisions makes the system's reasoning intelligible to users, auditors, and regulators, which is the core of explainability and a key pillar of transparent AI. Option A is incorrect because hiding the model's internal logic directly undermines transparency, even if it protects intellectual property. Option C is incorrect because training on the largest dataset improves performance or coverage, not transparency. Option E is incorrect because obtaining consent from data subjects addresses privacy and lawful processing, not transparency of AI decision-making.

Answer

Hide the model's internal logic to protect intellectual property

Answer

Train the model on the largest dataset available

Answer

Obtain consent from all data subjects

Question 11

Which TWO are key principles of Salesforce's AI ethics? (Choose two.)

Accepted Answer

Transparency. Salesforce's AI ethics framework is built on core principles that include Transparency (C), meaning AI systems should be explainable so users can understand how decisions are made and what data drives them, and Accountability (D), meaning organizations and developers must take responsibility for their AI systems' outcomes and be answerable for their impacts. These two principles align with Salesforce's published guidance on responsible AI, which emphasizes trust, explainability, and human oversight rather than raw performance metrics. Speed of deployment (A) is not an ethics principle; it is an operational goal that can conflict with responsible AI practices. Profitability (B) is a business objective, not an ethical principle, and Full automation (E) is a technical capability that Salesforce's ethics guidance explicitly cautions against when it removes meaningful human judgment.

Answer

Speed of deployment

Answer

Profitability

Answer

Full automation

Question 12

A large e-commerce company uses Salesforce Einstein to recommend products to customers. The AI model is trained on purchase history, browsing behavior, and demographic data including age and gender. Recently, the company received complaints that the model seems to recommend lower-priced items to female customers and higher-priced items to male customers for the same product categories. The data science team confirms the model has a statistically significant difference in recommendation value by gender. The company's ethical AI policy requires fairness, transparency, and human oversight. The compliance team is considering several actions. Which action should the company take first?

Accepted Answer

Conduct a thorough bias audit to identify all sources of disparate impact. The correct action is C: conduct a thorough bias audit to identify all sources of disparate impact. Because the model shows a statistically significant difference in recommendation value by gender, the company must first diagnose where the bias originates—training data, feature engineering, label choice, or proxy variables—before choosing a remediation, consistent with its ethical AI policy of fairness, transparency, and human oversight. Option A is wrong because deliberately raising prices for female customers would worsen the disparity and is discriminatory. Option B is overly disruptive and premature, since disabling the system is not required before understanding the root cause. Option D is insufficient because simply removing gender does not eliminate bias, as correlated proxy features such as browsing behavior or purchase history can still encode gender-based disparate impact.

Answer

Adjust the model to increase recommendation prices for female customers

Answer

Disable the recommendation system until the issue is resolved

Answer

Immediately remove gender from the training data and retrain the model

Question 13

A non-profit organization uses Salesforce AI to help prioritize grant applications. The AI scores applications based on historical funding decisions, project impact, and community need indicators. After deployment, staff notices that applications from rural areas consistently receive lower scores than those from urban areas, even when project quality is similar. The organization's mission is to serve underserved communities, including rural areas. The AI model is trained on historical data that favored larger, urban projects. The ethics committee is meeting to decide next steps. What is the most appropriate action to align ethical AI with the organization's mission?

Accepted Answer

Retrain the model with new criteria that prioritize rural impact. Retraining the model with new criteria that prioritize rural impact directly addresses the root cause: the historical training data biased toward larger, urban projects, which produces systematically lower scores for rural applicants despite similar quality. This aligns the AI's objective function with the organization's mission to serve underserved communities, including rural areas, rather than merely compensating for the bias downstream. Option B does not fix the biased model and leaves rural applicants disadvantaged. Option C is a manual workaround that is inconsistent, unscalable, and does not correct the underlying model. Option D adds transparency but does not change the discriminatory scoring behavior.

Answer

Accept the AI scores but provide staff training on bias

Answer

Manually overrule the AI scores for rural applicants

Answer

Continue using the model but add explainability reports

Question 14

A small business uses a pre-built Salesforce AI model to predict inventory needs. The model recommends ordering extra stock based on seasonal trends. One month, the model fails to predict a sudden demand spike, resulting in stockouts and lost sales. The business owner is frustrated and considers disabling the AI. The owner wants to know if this is an ethical issue and what to do next. As an AI ethics advisor, what is the best response?

Accepted Answer

Explain that this is a model performance issue, not an ethics violation, and suggest reviewing the model's accuracy and retraining with recent data. Option C is correct because a missed demand spike from a pre-built Salesforce AI model is a model performance and accuracy problem, not an ethics violation; the appropriate next step is to evaluate the model's accuracy metrics and retrain or update it with recent, more representative data. Ethical issues would involve concerns such as bias, privacy, transparency, or misuse, none of which are described here. Option A is wrong because disabling the AI immediately is an overreaction that ignores remediation and business value. Option B is wrong because no AI model can be guaranteed 100% accurate. Option D is wrong because not all AI failures are ethical issues; many are simply performance limitations.

Answer

Agree that the AI is unreliable and recommend disabling it immediately

Answer

Recommend using a different AI model that is guaranteed to be 100% accurate

Answer

State that all AI failures are ethical issues and the company should stop using AI

Question 15

A company deployed an AI model for lead scoring. After several months, they notice that leads from certain geographic regions consistently receive higher scores than leads from other regions with similar demographic profiles. The company wants to ensure ethical AI usage. What should they do first?

Accepted Answer

Retrain the model using a more diverse and balanced training dataset.. Retraining with more diverse data addresses potential bias at the source. Option A adjusts thresholds without fixing root cause. Option C ignores the discrepancy, which does not address the bias. Option D removes a feature that may be relevant but could still leak bias through correlated features.

Answer

Adjust the scoring thresholds for each region to equalize scores.

Answer

Ignore the discrepancy since the model overall accuracy is high.

Answer

Remove the geographic region feature from the model completely.