CCNA Sfai Ethical Ai Privacy Questions

The right to explanation requires that individuals can obtain meaningful information about the logic of automated decisions, including the factors that led to the outcome.

Zero Data Retention is the Einstein Trust Layer policy that ensures customer data is not used to train or improve Salesforce's base AI models. It is a core part of Salesforce's responsible AI framework.

Einstein Trust Layer includes audit trails and bias detection capabilities. Additionally, Einstein Discovery can be used to audit models for bias. The specific tool for bias detection is the Einstein Trust Layer's bias detection feature.

Using only relevant fields reduces unnecessary data collection, obtaining consent ensures compliance with privacy laws, and documenting data sources supports transparency and auditability.

To address bias, the team should audit the model for disparate impact, investigate and correct the historical data, and consider adding features that reduce reliance on biased proxies. Removing the model entirely is unnecessary, and manually adjusting scores is not a systemic fix.

Providing explanations and indicating AI-generated predictions help sales reps understand and trust the AI outputs, aligning with transparency.

Einstein Discovery provides prediction explanations with score factors, which can be used to explain automated decisions. This aligns with the right to explanation under GDPR.

Data minimisation is a core principle: only use relevant features for the prediction. Sensitive attributes that could lead to discriminatory decisions should be excluded unless legally required and properly managed.

CCPA gives consumers the right to know what personal information is collected and used. Using purchase history and browsing behavior are typical for personalization, but using location data without explicit consent may be problematic.

Removing one of the highly correlated features (D) is the best action because multicollinearity between 'number of support tickets' and 'average response time' can inflate the variance of coefficient estimates, making the model unstable and harder to interpret. By dropping one feature, you reduce redundancy without significant information loss, preserving both accuracy and interpretability in a linear or tree-based model.

The Einstein Trust Layer is designed to provide security, privacy, and governance controls for AI features within the Salesforce platform.

GDPR requires a lawful basis for processing personal data, such as consent or legitimate interest. The right to explanation is also relevant for automated decisions, but the primary requirement is obtaining a lawful basis (e.g., consent) before using the data.

Einstein Discovery provides model explanations, including score factors (most influential fields) for individual predictions. This is the standard way to explain why a particular lead scored as it did.

Einstein Trust Layer provides transparency and explainability features. The audit trail logs the factors that influenced the recommendation, and model explainability surfaces those factors. The combination of audit trail and explainability directly enables the right to explanation.

Right to explanation requires that customers can get reasons for automated decisions. Logging the factors and providing a way for customers to request explanations are key. Using consent management and data retention are important but not specifically for explanation.

Removing the prohibited field and re-evaluating feature importance aligns with data minimisation and fairness. Auditing for bias is good but not sufficient if the field is still used. Retraining without removal does not address the legal issue.

The Honesty and Transparency principles require that users are informed that scores are AI-generated, understand the factors influencing them, and are aware of limitations.

PII masking protects customer privacy by removing sensitive data from requests. Human oversight ensures that automated decisions are reviewed when necessary, especially for high-stakes interactions.

The AI-driven lead scoring model is producing outputs that do not match the ground truth (historical conversion rates), which is a direct failure of the Accuracy principle. Accuracy requires that AI systems perform as intended and produce reliable, correct predictions; here, the model's scores are systematically wrong for a specific region, violating that requirement.

When 'Score Factors' is enabled for an AI prediction in Salesforce, end users see a breakdown of the most influential fields that contributed to the prediction, along with their relative contribution (e.g., positive or negative impact). This provides transparency into why a specific prediction was made, helping users trust and act on the AI's output without exposing the underlying model logic.

Zero data retention ensures that customer data used in prompts or predictions is not stored by Salesforce to train or improve base models, protecting privacy.

The Data Processing Addendum (DPA) is a legally binding contract between Salesforce and its customers that defines how Salesforce will process customer data, including data used by AI features like Einstein. It ensures Salesforce adheres to data protection laws (e.g., GDPR, CCPA) by specifying data handling, security measures, and sub-processor obligations, making it essential for compliance when using AI services on the platform.

Option D is correct because the question specifically asks how to ensure no PII is used in model training. Data minimization by selecting only non-PII fields as predictors directly prevents PII from entering the training dataset at the source. This is a proactive governance practice that avoids reliance on post-processing or masking, which may still expose PII during intermediate steps.

Under GDPR, Article 22 grants individuals the right not to be subject to a decision based solely on automated processing, including profiling, which produces legal effects or similarly significant effects. For an AI email composer that drafts responses to customer inquiries, requiring human review before any AI-generated email is sent ensures that the final decision to communicate is not fully automated, thereby complying with GDPR's requirement for meaningful human intervention in automated decision-making.

The Einstein Trust Layer provides toxicity detection to prevent harmful outputs, PII masking to protect sensitive data, and audit trails for governance. Direct human review of promotional offers is also important for empathy and oversight.

Zero data retention is the specific commitment that customer data used for predictions is not stored or used for training base models. This is a key component of the Einstein Trust Layer.

Option B is correct because it aligns with Salesforce's Trusted AI principles, which emphasize accountability, transparency, and human oversight. Auditing the model for bias identifies disparities, providing transparency on decision factors ensures stakeholders understand how outcomes are determined, and requiring human review for denied applications introduces a safeguard against automated discrimination. This combination addresses bias without abandoning AI's benefits.

Option A is correct because reporting suspicious leads and requesting a model audit is the proper governance procedure when an AI model produces outputs that contradict observable data. This triggers a review of the model's feature weights and training data to identify bias or overvaluation of specific fields, ensuring the model remains reliable and compliant with ethical AI standards.

Option A is correct because the core issue is data sparsity and unrepresentative historical conversion data for that region. By collecting more data from that region, the model can be retrained on a balanced dataset, which directly addresses the root cause of the biased lead scoring. This aligns with the principle of fairness in AI, where models must be trained on representative data to avoid systematic discrimination.

The Salesforce Trusted AI principle of Empathy emphasizes that AI systems should be designed with a human-centered approach, ensuring they benefit people and avoid causing harm. This principle guides the ethical development of AI by prioritizing user well-being and societal impact over purely technical or business objectives.

Option A is correct because the Einstein Trust Layer includes an audit trail that logs when AI recommendations are generated and when users act on them. This audit trail is essential for compliance and governance, as it records the AI's decision-making process and user interactions, ensuring transparency and accountability in AI-driven recommendations.

Einstein Discovery provides 'score factors' that show the most influential fields and their impact on the prediction, enabling transparency.

Auditing for bias, ensuring diverse training data, and providing transparency are key to mitigating bias and following Trusted AI principles.

The Einstein Trust Layer is a security architecture within Salesforce that intercepts prompts sent to large language models (LLMs) and automatically masks personally identifiable information (PII) before the data reaches the external AI service. After the LLM returns a response, the Trust Layer reinserts the masked data, ensuring that sensitive information never leaves the Salesforce ecosystem in plain text. This feature is specifically designed to address data privacy and compliance requirements when using generative AI features like Einstein GPT.

Accuracy requires that the model should be accurate and tested. If the training data is missing entire categories, the model cannot be accurate for those categories, violating accuracy.

Toxicity detection in the Einstein Trust Layer is designed to identify and filter harmful or abusive language in AI-generated outputs before they reach the user. It uses natural language processing (NLP) models to score content against categories such as hate speech, profanity, and harassment, ensuring outputs remain safe and appropriate. This prevents the AI from inadvertently disseminating offensive material, which is critical for maintaining trust and compliance in enterprise environments.

PII masking directly protects health information, while audit trail provides accountability and traceability for compliance.

Option A is correct because Score Factors in Einstein Lead Scoring provides transparency by listing the specific data points (e.g., lead source, industry, engagement history) that contributed to a lead's score. This feature allows admins to see exactly why a lead received a high score, enabling them to validate or adjust the scoring model. It directly addresses the need for explainability in AI-driven lead scoring.

Option C is correct because the first step in addressing potential bias is to audit the training data for representation and bias. This involves examining whether certain categories or demographics are underrepresented or overrepresented in the dataset, which can lead to skewed model recommendations. Without understanding the data composition, other interventions may fail to address the root cause of bias.

Transparency requires that AI systems are open about their capabilities and limitations. Informing users that a recommendation is AI-generated is a direct application of transparency.

Data minimisation means using only relevant fields. Removing unnecessary PII features and performing feature selection reduces risk. Auditing for bias and enabling toxicity detection are good practices but not directly about minimisation.

The model was trained on historical data that may have reflected past biases (e.g., underinvestment in that region). The biased training data leads to unfair predictions. The correct action is to audit the model for bias.

An audit trail logs when AI recommendations are made and whether they were followed, and can be used to track outcomes over time. This enables accountability and measurement of AI impact, supporting transparency and governance.

Human oversight requires that AI-generated content be reviewed before sending. Einstein Service Replies can be configured to require agent approval before the reply is sent, which is a direct implementation of human oversight.

Toxicity detection identifies harmful or offensive content. In this context, harmful financial advice can be flagged as toxic. Human oversight (agent review) would also be applicable, but the question specifically asks about blocking before reaching the customer via a capability of the Trust Layer.

Option C is correct because in healthcare AI, accuracy ensures that treatment recommendations are clinically valid and based on reliable patient data, while empathy ensures the system respects patient well-being and avoids harmful or insensitive suggestions. Without accuracy, the model could recommend incorrect treatments; without empathy, it might ignore patient context or ethical considerations. These two principles together directly address the dual need for clinical correctness and human-centered care in treatment planning.

Option B is correct because validating model accuracy separately for different patient subgroups is a key fairness practice. It ensures the model performs consistently across diverse populations, preventing hidden disparities that could lead to biased clinical decisions. This aligns with responsible AI principles that require subgroup performance analysis to detect and mitigate algorithmic bias.

Einstein Discovery provides prediction explanations and score factors. For recommendations within Sales Cloud, the 'Why this recommendation' feature (part of Einstein Discovery) shows the most influential fields.

Bias in historical data can lead to unfair predictions. The best approach is to audit the model for bias and retrain with corrected labels or additional features that capture the true drivers of conversion, rather than simply omitting the feature or adjusting scores manually.

HIPAA requires protection of Protected Health Information (PHI). PII masking in the Einstein Trust Layer ensures that PHI is not exposed or retained in logs or training.

Option B is correct because Salesforce's Trusted AI principles emphasize human oversight to ensure AI-generated content is accurate, appropriate, and aligned with brand values. The manager must have a human review the content before sending to mitigate risks like bias, inaccuracy, or inappropriate messaging, which aligns with the principle of accountability in AI deployment.

Option B is correct because GDPR's right to explanation requires that individuals can understand the logic behind automated decisions. Providing a list of the most influential factors for each recommendation directly addresses this by offering transparency into the model's decision-making process, such as feature importance scores from a tree-based or linear model.

Transparency in AI means that decisions should be explainable and understandable to users. Providing an explanation for a loan rejection directly supports the principle of Transparency, which requires AI systems to be open about their reasoning.

Empathy means designing for human benefit, including user-centric design, fallback to humans, and considering emotional impact. Automating all responses and using a formal tone are less empathetic.

Option D is correct because it directly implements human oversight by requiring the sales rep to click 'Approve' before the email is sent, aligning with Salesforce's Trusted AI Principle of Empathy and Human Oversight. This principle mandates that AI systems should include mechanisms for human review and control, especially in high-stakes communications, ensuring that AI-generated content is vetted for accuracy, tone, and compliance before reaching customers. The approval step creates a clear human-in-the-loop checkpoint, preventing automated sending without human judgment.

Under GDPR Article 22, customers have the right to request an explanation of automated decisions and to contest them. Transparency about AI limitations is also part of Salesforce's Honesty principle.

Option A is correct because Einstein Copilot employs a zero data retention policy specifically for the underlying large language model (LLM). This means that any customer data processed during email generation is not stored, logged, or used for model training or fine-tuning, ensuring compliance with data privacy standards. The mechanism explicitly prevents the LLM from learning from or being improved by customer interactions, isolating the AI's behavior from proprietary data.

Human oversight can be achieved by requiring agents to confirm before an action is taken, and by clearly presenting AI recommendations as suggestions. Disabling the AI or automating actions removes human involvement. Training agents is helpful but does not directly ensure oversight in the workflow.

Honesty means being transparent about AI limitations and capabilities. Providing documentation, clearly communicating that predictions are AI-based, and explaining the influencing factors align with honesty. Automating without review and overpromising accuracy are contrary.

The model assigns lower scores to leads from certain postal codes because the training data reflects historical human biases, such as sales representatives prioritizing leads from affluent areas. Machine learning models learn patterns from the data they are trained on; if past sales decisions were biased against certain regions, the model will replicate those biases. This is a classic case of bias in training data leading to biased model outcomes, not an inherent flaw in the algorithm itself.

Data minimisation means using only the data necessary for the prediction. Purchase history is relevant to churn prediction, while names and email addresses are personally identifiable information (PII) that are not needed for the model and should be excluded to minimize privacy risks.

Option B aligns with the data minimization principle by restricting data collection to only fields directly relevant to the prediction task (purchase history and engagement) and explicitly excluding personally identifiable information (PII) like social security numbers. This reduces privacy risk and complies with regulations such as GDPR and CCPA, which require that data collected be adequate, relevant, and limited to what is necessary for the processing purpose.

Option C is correct because the Salesforce Trusted AI Principle of Honesty requires transparency about AI-generated content. By explicitly stating 'Recommended by Einstein AI' on each recommendation, the company clearly discloses that the recommendations are AI-generated, building trust with customers. This aligns with the principle's focus on avoiding deception and ensuring users understand the nature of the content they are seeing.

Safety involves preventing harmful outputs. Human review, toxicity detection, and using grounding to ensure relevance all mitigate harmful recommendations. Hiding the model's confidence score reduces transparency, and logging without review does not prevent harm.

Grounding in the Einstein Trust Layer connects the AI model to relevant CRM records (e.g., Accounts, Opportunities, Cases) so that generated responses are based on accurate, up-to-date customer data rather than the model's general training data. This ensures context-aware and factual outputs while maintaining data privacy by not exposing raw records to the model.

The best practice is to exclude protected attributes from the model features. Data minimisation supports this. Auditing for bias is also important but does not prevent the model from using the attribute in the first place.

Score Factors show why a lead scored as it did, and labeling AI-generated scores helps reps understand that the score is AI-driven.

Data minimisation means using only the data necessary for the task. Excluding sensitive attributes from the feature set is the most direct way to prevent them from being used, even if they are correlated with other features. Correlation does not imply causation, and if those features are not essential, they should be removed.

Salesforce's DPA states that customer data will not be used to train or improve Salesforce's base AI models, ensuring customer data privacy.

Data minimisation is a key principle: only use relevant features for the model and avoid including sensitive PII unnecessarily. The Einstein Trust Layer's PII masking also helps, but the principle that directly addresses using only relevant fields is data minimisation.

Consent management involves checking the customer's communication preferences and honoring opt-outs. This can be done by configuring the recommendation logic to check consent fields or using Data Cloud's consent data.

Option A is correct because Einstein Discovery's score factors feature explicitly lists the top predictors and their contributions to each prediction, making the model's reasoning transparent and accessible to non-technical stakeholders. This aligns with the need for explainable AI (XAI) without requiring deep data science knowledge.

Toxicity detection is a feature of the Einstein Trust Layer that identifies and filters harmful or offensive language before it is sent to customers. This directly addresses the concern about offensive responses, aligning with the Safety principle.