AI Associate Ethical Considerations of AI Practice Test 9 — 15 Questions

Question 1

Which TWO actions are most effective in promoting transparency in AI systems? (Choose two.)

Accepted Answer

Provide a model card that describes the purpose, accuracy, and limitations of the AI model.. Option B is correct because a model card is a standardized transparency artifact that documents a model's intended purpose, performance metrics such as accuracy, and known limitations, giving stakeholders the context needed to understand and appropriately use the system. Option E is correct because an explanation feature (e.g., local interpretable explanations such as SHAP or LIME outputs) reveals the factors behind an individual prediction, which is a core requirement of transparency and the right to explanation in frameworks like the EU AI Act and GDPR Article 22. Option A does not belong because dumping raw source code is not effective transparency—it exposes intellectual property, may reveal security weaknesses, and does not communicate model behavior or limitations to non-expert stakeholders. Option C is incorrect because withholding training data sources actively reduces transparency and undermines accountability and bias auditing. Option D is incorrect because encrypting customer data is a security and privacy control, not a transparency mechanism—it protects data confidentiality but does not make the AI system's behavior or decisions more understandable.

Answer

Publish the entire source code of the model online.

Answer

Withhold the data sources used for training to protect the company's competitive advantage.

Answer

Encrypt all customer data used in the model.

Question 2

Which TWO are best practices for mitigating bias in AI models when using Salesforce Einstein? (Choose two.)

Accepted Answer

Use diverse and representative training data that reflects the target population.. Option A is correct because using diverse and representative training data that reflects the target population is a foundational best practice for reducing bias in Einstein models — if the data underrepresents certain groups, the model's predictions will systematically disadvantage them. Option E is correct because conducting regular bias audits with Einstein's fairness evaluation tools allows teams to detect disparate impact across protected groups and remediate issues before deployment, supporting ongoing governance. Options B, C, and D are not best practices: excluding demographic features with small sample sizes hides rather than fixes underrepresentation, prioritizing majority-group accuracy sacrifices fairness for aggregate metrics, and convenience sampling produces skewed, non-representative datasets that amplify bias.

Answer

Exclude demographic features that have small sample sizes to avoid statistical noise.

Answer

Prioritize accuracy for the majority group to maximize overall performance.

Answer

Rely on convenience sampling to quickly gather a large dataset.

Question 3

Which THREE are key principles of Salesforce's AI Ethics framework? (Choose three.)

Accepted Answer

Be Inclusive. Options A, C, and D are correct. Option B is wrong because 'speed over accuracy' contradicts responsible AI. Option E is wrong because 'maximum profit' is not an ethical principle.

Answer

Be Fast Over Accurate

Answer

Be Profit-Maximizing

Question 4

A company is developing an AI model to screen job applications. The training data is heavily skewed toward candidates from a specific demographic. What is the most important step the team should take to address potential ethical concerns?

Accepted Answer

Ensure the training data is diverse and representative of all demographics.. Using diverse and representative training data helps mitigate bias by ensuring the model learns from a balanced distribution of candidates. Option A is wrong because simply removing sensitive attributes does not guarantee fairness; the model can still learn proxies for those attributes, perpetuating bias. Option C is wrong because deploying the model quickly and monitoring complaints is reactive and may cause harm before issues are detected. Option D is wrong because increasing dataset size without addressing class imbalance can amplify existing biases rather than reduce them.

Answer

Remove all sensitive attributes from the data before training.

Answer

Deploy the model quickly and monitor for complaints.

Answer

Increase the overall size of the training dataset without regard to distribution.

Question 5

A retail company uses AI to personalize marketing emails. A customer complains that their data was used without explicit permission. What ethical principle was most likely violated?

Accepted Answer

Consent and privacy. Consent and privacy, because the scenario describes a customer whose data was used for AI-driven email personalization without explicit permission, which directly violates the ethical requirement to obtain informed consent and protect personal data. Consent and privacy principles require organizations to be transparent about data collection and to give individuals control over how their information is used, especially for marketing purposes. Transparency (A) relates to disclosing how data is used, but the core violation here is the absence of permission, not merely a lack of disclosure. Fairness (B) concerns avoiding bias and equitable treatment, which is not the issue described, and accountability (D) concerns taking responsibility for outcomes, which is secondary to the missing consent.

Answer

Transparency

Answer

Fairness

Answer

Accountability

Question 6

A financial services firm uses a deep learning model to approve loans. The model is highly accurate but cannot explain its decisions. Regulators now require the firm to provide reasons for loan denials. What is the best approach to address this ethical concern?

Accepted Answer

Replace the deep learning model with a simpler, interpretable model like logistic regression.. The correct option is C: replacing the deep learning model with a simpler, interpretable model like logistic regression. In this regulated financial scenario, the firm must provide actual reasons for loan denials, and logistic regression offers inherently interpretable coefficients that directly show how each feature contributes to the decision, satisfying the regulatory requirement for explainability. Options A and B do not address explainability: removing influential features can degrade accuracy without making the model transparent, and retraining with more data only improves performance, not interpretability. Option D, post-hoc tools like LIME, only approximate the model's behavior locally and may not provide the faithful, legally defensible reasons regulators require.

Answer

Remove the most influential features from the model.

Answer

Retrain the model with more data to improve accuracy further.

Answer

Use a post-hoc explanation tool like LIME to approximate decisions.

Question 7

An AI system used for medical diagnosis occasionally produces incorrect results. A doctor notices the errors but continues using the system without reporting them. Which ethical principle is primarily at risk?

Accepted Answer

Accountability. Accountability is at risk because the doctor observes incorrect AI outputs and fails to report them, breaking the chain of responsibility for the system's outcomes. In AI ethics, accountability requires that humans remain responsible for monitoring, reporting, and correcting AI errors. Ignoring known errors without escalation undermines the governance structure that ensures the system is safe and reliable.

Answer

Fairness

Answer

Transparency

Answer

Privacy

Question 8

A company launches a chatbot that interacts with customers. The chatbot does not disclose that it is an AI. Which ethical principle is most directly violated?

Accepted Answer

Transparency. Transparency, because the scenario describes a chatbot that conceals its AI nature from customers, which directly violates the ethical principle of transparency—the obligation to disclose automated/AI involvement so users can make informed decisions. Transparency in AI ethics specifically requires that systems not deceive users about their identity or capabilities. Accountability (A) concerns who is responsible for the system's actions, not disclosure of AI identity. Fairness (B) addresses bias and equitable treatment, which are not at issue here. Privacy (D) concerns handling of personal data, which the scenario does not mention.

Answer

Accountability

Answer

Fairness

Answer

Privacy

Question 9

An autonomous vehicle AI is trained in simulation but performs poorly in rain and snow. The development team decides to deploy anyway, arguing that bad weather is rare. What ethical concern is most critical?

Accepted Answer

Insufficient robustness and safety. The correct option is B (Insufficient robustness and safety): the model was trained only in simulation and fails in rain and snow, so deploying it means the system lacks the robustness to handle real-world conditions, directly endangering lives—an unacceptable safety risk regardless of how rare bad weather is. Deploying a known-failing system violates the core ethical duty to avoid foreseeable harm. Option A is not the best fit because 'bias' typically refers to unfair treatment of people or groups, not environmental conditions. Option C does not apply since the team is clearly identifiable and accountable, and Option D is not the issue because the failure is a performance/robustness gap, not a lack of explainability.

Answer

Bias against certain weather conditions

Answer

Lack of human accountability

Answer

Lack of transparency

Question 10

A social media platform's AI recommends content that inadvertently amplifies misinformation. An ethical review board is considering changes. Which of the following actions best addresses the unintended harm?

Accepted Answer

Conduct an ethical impact assessment and adjust algorithms accordingly. Conducting an ethical impact assessment directly identifies how the recommendation algorithm amplifies misinformation and then adjusts the algorithm to mitigate that specific harm, which is the appropriate governance response for unintended ethical side effects. Increasing training data (A) does not address the underlying objective or feedback loop that promotes misinformation and may even worsen it. Removing all AI recommendation engines (B) is a disproportionate, overbroad response that discards legitimate functionality rather than fixing the harm. Reducing overall user engagement metrics (D) is a vague business change that does not systematically detect or correct the algorithmic cause of misinformation amplification.

Answer

Increase the amount of training data

Answer

Remove all AI recommendation engines

Answer

Reduce overall user engagement metrics

Question 11

A health app collects users' location data for AI-driven recommendations, but users are not informed about this data collection. Which ethical principle is most compromised?

Accepted Answer

Data minimization and consent. Data minimization and consent, because the scenario describes users being unaware that their location data is collected and used for AI recommendations, which directly violates the requirement to obtain informed consent and to limit collection to what is necessary and disclosed. Transparency (A) is related but is a broader principle about openness and disclosure, whereas the specific ethical breach here centers on collecting and using personal data without user permission. Fairness (C) concerns equitable treatment and avoiding bias in outcomes, which is not the primary issue described. Accountability (D) concerns being answerable for data practices, but the core violation is the lack of consent and unnecessary data collection, not the absence of a responsible party.

Answer

Transparency

Answer

Fairness

Answer

Accountability

Question 12

An AI system for hiring is deployed. After six months, the HR team notices that the model's recommendations closely mimic past human hires, which were biased. The team wants to correct this. What should be their first step?

Accepted Answer

Implement continuous monitoring and a feedback loop to detect and mitigate bias. The model is replicating historical human bias, so the first step is to establish continuous monitoring and a feedback loop that detects and mitigates bias in the model's recommendations. This creates the mechanism to measure disparate impact, catch drift, and apply corrections iteratively. It is the responsible, actionable first step before retraining or redesigning.

Answer

Shut down the AI system entirely

Answer

Retrain the model with the same historical data but with more features

Answer

Make the model's decision process fully transparent to all candidates

Question 13

Which TWO of the following are considered core ethical principles in AI according to Salesforce’s AI Ethics?

Accepted Answer

Accountability. According to Salesforce's AI Ethics guidelines, Accountability (A) is a core principle because organizations must take responsibility for the outcomes of their AI systems, ensuring human oversight and the ability to explain and remedy decisions. Transparency (C) is also a core principle, as AI systems should be designed so that their decision-making processes are open, understandable, and auditable by stakeholders. These two principles are explicitly part of Salesforce's ethical framework, alongside others such as fairness, reliability, and human-centered design. Popularity (B), Speed (D), and Profitability (E) are not ethical principles in this context; they are business or performance metrics that do not address the moral obligations of AI development and deployment.

Answer

Popularity

Answer

Speed

Answer

Profitability

Question 14

Which THREE of the following are effective strategies to mitigate bias in AI models?

Accepted Answer

Ensuring training data is diverse and representative. Option A is correct because ensuring training data is diverse and representative directly addresses the root cause of many biases: skewed or unrepresentative datasets that cause the model to learn and reproduce discriminatory patterns. Option B is correct because conducting regular bias audits on model outcomes allows teams to detect disparate impact across demographic groups after deployment and take corrective action, which is a key part of responsible AI governance. Option E is correct because involving a diverse team in model development brings varied perspectives that help identify blind spots, assumptions, and edge cases that a homogeneous team might overlook. Option C is not correct because reducing model complexity affects performance, interpretability, and overfitting, but it does not inherently mitigate bias and can even worsen it by limiting the model's ability to capture fair representations. Option D is not correct because using only historical data without modification preserves and can amplify existing societal biases embedded in that data, which is the opposite of a bias-mitigation strategy.

Answer

Reducing the complexity of the model

Answer

Using only historical data without modification

Question 15

Which TWO of the following are required for GDPR compliance when using AI with personal data?

Accepted Answer

Obtaining explicit consent from users before processing their data. Option B is correct because GDPR requires a lawful basis for processing personal data, and explicit consent (Article 6(1)(a) and Article 9(2)(a) for special categories) is one such valid basis that must be obtained before processing, including for AI training or inference. Option D is correct because GDPR grants data subjects the right to erasure ("right to be forgotten") under Article 17, so organizations using personal data in AI systems must be able to honor deletion requests. Option A is incorrect because GDPR's storage limitation principle (Article 5(1)(e)) prohibits keeping personal data longer than necessary, so indefinite retention is non-compliant. Option C is incorrect because selling user data requires a valid lawful basis and transparency, and typically falls outside legitimate AI improvement without explicit consent, violating GDPR principles. Option E is incorrect because GDPR's transparency obligations (Articles 12–14) require informing users about how their data is processed, so silent processing is non-compliant.

Answer

Storing data indefinitely for future AI training

Answer

Selling user data to third parties for AI model improvement

Answer

Processing data without informing users