Ethical AI and Data Privacy practice questions

Empathy

Why wrong: Empathy is about designing for human benefit, but the core issue here is erroneous scoring.

Safety

Why wrong: Safety focuses on preventing harmful outputs, which is not directly the issue here.

Transparency

Why wrong: Transparency is about explaining AI decisions, not about correctness of predictions across segments.

To replace all third-party AI services with Salesforce-owned models

Why wrong: The Trust Layer is agnostic and can work with various AI providers.

To automatically generate AI models without any human oversight

Why wrong: The Trust Layer facilitates human oversight and governance.

To train large language models on customer data for better predictions

Why wrong: The Trust Layer prevents customer data from being used to train underlying models.

Enabling zero data retention in the Trust Layer

Why wrong: Zero data retention prevents customer data from being used to train base models, but does not prevent PII from being used in custom models.

Data anonymization via the Einstein Trust Layer

Why wrong: The Trust Layer masks PII in prompts and responses but does not automatically exclude PII from training datasets.

Regularly auditing the model for bias

Why wrong: Auditing for bias checks fairness, not the presence of PII in training data.

Allow the AI to send emails automatically if confidence is high

Why wrong: Automated sending without human review may violate GDPR's right to human intervention.

Disable the AI composer entirely to avoid GDPR risk

Why wrong: Overly restrictive; human oversight can mitigate risk without disabling the feature.

Anonymize customer data before drafting

Why wrong: Anonymization is a privacy measure but does not address the right to human intervention for automated decisions.

Einstein Activity Capture

Why wrong: Activity Capture logs emails and events, not prediction explanations.

Einstein Copilot prompt template

Why wrong: Prompt templates are for conversational AI, not lead scoring explanations.

Einstein Trust Layer audit trail

Why wrong: The audit trail logs AI actions but does not provide per-prediction explanations.

To increase the speed of AI responses

Why wrong: Toxicity detection adds a processing step and may slightly reduce speed.

To monitor user interactions for compliance with data protection laws

Why wrong: Compliance monitoring is broader; toxicity detection is specifically about harmful language.

To identify and mask PII in user prompts

Why wrong: PII masking is a separate component of the Trust Layer.

Keep both features to maximize information

Why wrong: Keeping both can inflate variance and make the model less interpretable.

Apply principal component analysis to combine them

Why wrong: PCA creates new components that are harder to interpret in the business context.

Increase the regularization strength

Why wrong: Regularization helps but does not directly address interpretability; removing one feature is simpler.

Customer data may be used to improve Salesforce's AI models unless the customer opts out

Why wrong: The DPA commits to zero data retention; customer data is not used to train base models.

Customer data is anonymized and then used to train public AI models

Why wrong: Salesforce does not use customer data for public model training.

Customer data is only used to train models for that specific customer

Why wrong: Even for custom models, customer data is not used for base model improvement.

Einstein Discovery

Why wrong: Discovery provides insights but not detailed audit logs of actions.

Einstein Copilot

Why wrong: Copilot is an assistant, not an audit logging tool.

Einstein Prediction Builder

Why wrong: Prediction Builder creates models, but audit logging is a Trust Layer feature.

Customer purchase history from the past year

Why wrong: Purchase history is commonly used for personalization and is generally acceptable with notice.

Customer email address

Why wrong: Email address is basic contact information and is usually collected with consent.

Customer browsing behavior on the company website

Why wrong: Browsing behavior is also typical, though disclosure is required.

Disable the AI model and make all decisions manually

Why wrong: Overly drastic; bias can be mitigated with proper controls.

Remove demographic data from the model entirely and continue using it

Why wrong: Removing demographic data does not guarantee fairness; bias can still exist through proxy variables.

Retrain the model with more data from the affected group and deploy automatically

Why wrong: Simply retraining may not solve bias; oversight and transparency are also needed.

PII masking

Why wrong: PII masking protects sensitive data but does not prevent data retention.

Grounding

Why wrong: Grounding connects AI to CRM data but does not address model training.

Toxicity detection

Why wrong: Toxicity detection filters harmful content; it is unrelated to data retention.

Store all recommendation logs for at least 10 years

Why wrong: Storage duration is a data retention practice, not directly an explanation right.

Allow customers to opt out of all AI recommendations

Why wrong: Opt-out is a right but not specifically 'right to explanation'.

Use only anonymized data for recommendations

Why wrong: Anonymization reduces privacy risk but does not fulfill the right to explanation.

Use only historical data from the past six months

Why wrong: Short time windows may not capture enough data for robust model training and could introduce seasonality bias.

Include as many features as possible to maximize predictive power

Why wrong: Including too many features, especially sensitive ones, can introduce bias and reduce interpretability.

Einstein Copilot

Why wrong: Copilot is a conversational assistant, not a lead scoring explanation tool.

Einstein Activity Capture

Why wrong: Activity Capture logs emails and events, not scoring explanations.

Einstein Trust Layer audit trail

Why wrong: Audit trail logs actions but does not provide per-lead explanations.

Delete the lead from the CRM to avoid influencing future predictions

Why wrong: Deleting data removes evidence and does not address the underlying model bias.

Manually override the score to 50 and move on

Why wrong: Manual override without investigation does not fix the root cause and may reintroduce bias.

Contact Salesforce support to disable the AI model

Why wrong: Disabling the model is drastic; the rep should first investigate and escalate.

Automatic model retraining

Why wrong: Automatic retraining would use customer data, violating privacy commitments.

Data masking for all PII

Why wrong: Data masking is a feature, but zero data retention is the key privacy protection that prevents model training on customer data.

Open-source model deployment

Why wrong: Open-source deployment does not directly address data privacy.

Share the full training dataset for transparency

Why wrong: Sharing the dataset raises privacy concerns and is not an explanation.

Allow stakeholders to retrain the model themselves

Why wrong: Retraining does not explain existing predictions.

Provide the raw model weights and coefficients

Why wrong: Raw weights are too technical for non-technical audiences.

Safety and Honesty

Why wrong: Safety is important, but empathy is more directly about patient benefit; honesty is secondary.

Honesty and Transparency

Why wrong: Honesty and transparency are about communication, not the core outcome of correct recommendations.

Transparency and Safety

Why wrong: Transparency and safety are important but less critical than accuracy and empathy in this direct patient-care scenario.

The model's algorithm is inherently biased against certain regions

Why wrong: Algorithms are not inherently biased; bias comes from data or feature selection.

The model overfits to the training data

Why wrong: Overfitting would cause poor generalization but not necessarily bias against specific postal codes.

The model was not trained long enough

Why wrong: Training duration does not cause systematic bias against specific groups.