AI Associate Practice Exam 56 — 20 Questions

Question 1

During model development, the data scientist realizes the training data is not representative of the intended population. What should they do?

Accepted Answer

Augment data with synthetic samples for underrepresented groups.. The correct answer is D: augment data with synthetic samples for underrepresented groups. When training data is not representative of the intended population, the model will learn biased patterns and perform poorly on underrepresented groups, so adding synthetic samples (e.g., via SMOTE, data augmentation, or generative methods) helps balance the dataset and improve generalization to the target population. Option A is wrong because removing underrepresented groups worsens the representativeness problem and introduces exclusion bias. Option B is wrong because regularization addresses overfitting, not dataset imbalance or sampling bias. Option C is wrong because using unrepresentative data as-is will not magically generalize; the model will inherit the same biases present in the training set.

Answer

Remove the underrepresented groups from the scope.

Answer

Increase model regularization.

Answer

Use the data as is, as the model will generalize.

Question 2

A data integration specialist is using Data Pipelines to combine Salesforce data with an external CSV file. The CSV has a header row but some rows have extra commas, causing parsing errors. What should the specialist do?

Accepted Answer

Use a Data Transform recipe to clean the data before ingestion. The correct option is A: Use a Data Transform recipe to clean the data before ingestion. In Data Pipelines, a Data Transform recipe can apply transformations such as parsing, filtering, and field-level cleanup to handle malformed CSV rows with extra commas before the data is loaded, which directly resolves the parsing errors without manual intervention. Option B is not scalable or automatable and does not address the pipeline's ingestion logic. Option C is irrelevant because timeout settings do not fix CSV parsing errors. Option D is overly disruptive and unnecessary since the data can be cleaned programmatically.

Answer

Edit the CSV manually

Answer

Increase the pipeline timeout

Answer

Reject the entire file and request a corrected version

Question 3

A nonprofit uses Salesforce AI to prioritize outreach to donors. The model recommends contacting only high-income individuals. Which ethical principle is most compromised?

Accepted Answer

Fairness. Fairness, because the model's practice of prioritizing only high-income individuals creates discriminatory, biased outreach that excludes lower-income donors, directly violating equitable treatment. Fairness in AI requires that automated decisions avoid unjust bias and do not systematically disadvantage protected or vulnerable groups. Accountability (A) concerns who is responsible for outcomes, transparency (C) concerns explainability and openness, and privacy (D) concerns handling personal data, none of which is the primary principle compromised by biased targeting.

Answer

Accountability

Answer

Transparency

Answer

Privacy

Question 4

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 5

A Salesforce admin is setting up an AI-powered lead scoring system. To ensure ethical use, what should they prioritize?

Accepted Answer

Regularly audit the model for bias and adjust scoring to ensure fairness across customer segments. The correct answer is C: regularly audit the model for bias and adjust scoring to ensure fairness across customer segments. Ethical AI use in lead scoring requires ongoing monitoring for disparate impact, because models trained on historical sales data can encode and amplify existing biases against protected or underrepresented groups; periodic bias audits and recalibration keep outcomes fair and compliant with responsible-AI principles. Option A is wrong because simply maximizing record volume does not address data quality, representation, or fairness. Option B is wrong because automatic prioritization of top scores can operationalize biased outputs without human oversight. Option D is wrong because using an unmodified default model does not guarantee ethical or context-appropriate behavior.

Answer

Ensure the training data includes as many records as possible

Answer

Set the system to automatically prioritize the highest-scoring leads

Answer

Use the default model provided by Salesforce without customization

Question 6

A Salesforce consultant is advising a client on the ethical use of AI for lead scoring. The client wants to ensure their AI system respects user privacy and complies with regulations. Which action should the consultant recommend as a best practice?

Accepted Answer

Implement data minimization and obtain explicit consent for data usage.. Implementing data minimization and obtaining explicit consent are core ethical AI practices. They protect user privacy, ensure compliance with regulations, and build trust. Salesforce's ethical guidelines strongly advocate for these principles. The other options involve overcollection, non-disclosure, or ignoring laws, all of which are unethical and risky. Thus, data minimization and explicit consent are the best recommendations.

Answer

Collect as much personal data as possible to improve lead scoring accuracy.

Answer

Ignore privacy regulations as they are not enforceable for AI systems.

Answer

Use third-party data without informing users to gain a competitive edge.

Question 7

An AI system for hiring is found to have a disparate impact on a protected class. The company is legally required to...

Accepted Answer

Conduct a bias audit and take corrective action.. The correct answer is A: Conduct a bias audit and take corrective action. When an AI hiring system produces a disparate impact on a protected class, the legally required response is to investigate the source of the bias through an audit and then remediate it, which aligns with anti-discrimination enforcement frameworks such as Title VII and emerging AI regulations. Option B is incorrect because obtaining applicant consent does not cure a discriminatory outcome or satisfy disparate-impact liability. Option C is incorrect because publishing the algorithm is a transparency measure, not a legal remedy for bias. Option D is incorrect because discontinuing the system is not the mandated first step; remediation through audit and corrective action is required.

Answer

Obtain consent from applicants.

Answer

Publish the algorithm.

Answer

Discontinue use of the system.

Question 8

A healthcare provider uses AI to predict patient outcomes. Which THREE measures should be implemented to ensure ethical AI use?

Accepted Answer

Establish a human-in-the-loop for critical decisions. Option A is correct because a human-in-the-loop ensures that critical healthcare decisions informed by AI predictions are reviewed and validated by qualified clinicians, preserving accountability and preventing harmful automated outcomes. Option B is correct because strict data anonymization (e.g., de-identification per HIPAA) and explicit patient consent protocols protect privacy and respect patient autonomy, which are foundational ethical requirements when using sensitive health data. Option E is correct because regularly testing for bias across demographic groups (e.g., measuring disparate impact and performance parity by age, sex, race, and socioeconomic status) detects and mitigates algorithmic discrimination that could worsen health inequities. Option C does not belong because black-box models reduce interpretability and explainability, which undermines transparency and clinical trust rather than supporting ethical use. Option D does not belong because optimizing for maximum profit prioritizes financial gain over patient welfare and equity, directly conflicting with ethical AI principles in healthcare.

Answer

Use a black-box model for better accuracy

Answer

Optimize model for maximum profit

Question 9

A company deploys an Einstein AI model that recommends products to customers. To ensure transparency, what should the company include in the customer-facing interface?

Accepted Answer

A brief note that recommendations are generated by AI, along with a way to get more information about why a product was suggested.. Option D is correct because transparency in AI-driven recommendations requires disclosing that the recommendations are AI-generated and providing an accessible path for customers to understand the reasoning behind a specific suggestion, such as an explanation link or 'Why am I seeing this?' feature. This aligns with responsible AI principles like explainability and user trust, ensuring customers are informed without overwhelming them with technical detail. Option A is incorrect because falsely claiming human selection misrepresents the AI system and undermines transparency. Option B is insufficient because a bare 'AI-generated' label lacks the explanatory context customers need to understand recommendations. Option C is impractical and not user-friendly, as exposing the complete feature list would overwhelm customers and may reveal proprietary or sensitive model details rather than a meaningful explanation.

Answer

A statement that each recommendation is personally selected by a human expert.

Answer

A simple label 'AI-generated' without further explanation.

Answer

A complete list of all features used by the model.

Question 10

A global company uses Salesforce Einstein Discovery to predict customer churn. They have a dataset with fields: Customer_Since__c (date), Last_Interaction_Date__c (date), Support_Cases__c (number), Product_Usage__c (percentage), Region__c (picklist), and Churned__c (boolean target). The model was trained and deployed, but predictions show bias against customers in the "EMEA" region. The data scientist notices that in the training data, 80% of EMEA customers are labeled as churned, while only 20% of other regions. Additionally, the Product_Usage__c field has many missing values for EMEA customers. The company wants to retrain the model to reduce bias. What is the best course of action?

Accepted Answer

Preprocess the data to impute missing Product_Usage__c values using region-specific averages, and then rebalance the dataset using stratified sampling. Option D is correct because the bias stems from two data-quality issues: missing Product_Usage__c values concentrated in EMEA and severe class imbalance across regions. Imputing missing values with region-specific averages preserves regional patterns instead of distorting them with a global mean, and stratified sampling ensures each region's churned/non-churned proportions are represented fairly during retraining. This directly addresses the root causes of the biased predictions rather than masking them. Option A is wrong because oversampling/undersampling without fixing the missing values still leaves EMEA records with incomplete features, and it can amplify noise or discard useful data. Option B is wrong because adding synthetic EMEA data without addressing the missing values and imbalance may reinforce the existing bias. Option C is wrong because removing Region__c does not fix the missing Product_Usage__c values and may hide legitimate regional effects rather than reduce bias.

Answer

Oversample EMEA churned customers and undersample non-churned from other regions

Answer

Increase the sample size of EMEA customers by adding synthetic data

Answer

Remove the Region__c field from the model and retrain

Question 11

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. Audit training data for bias. When a model shows discriminatory outcomes tied to a protected or proxy attribute like zip code, the first step is to investigate the root cause by auditing the training data for historical bias, underrepresentation, or skewed labels, since removing a feature or changing model architecture without understanding the source can mask rather than fix the problem. Auditing also establishes the evidence needed for remediation such as rebalancing data, applying fairness constraints, or using bias-detection tooling. Option A (increasing model complexity) would likely amplify the biased pattern rather than resolve it. Option B (adding more features) risks introducing additional proxy variables and worsening disparate impact. Option C (removing the zip code feature) may seem intuitive but is insufficient because correlated features can still encode the same bias, so it does not address the underlying data issue.

Answer

Increase model complexity

Answer

Add more features to the model

Answer

Remove zip code feature

Question 12

A hospital uses an AI model to predict patient deterioration. The model was trained on data from a single hospital with a predominantly white patient population. When deployed at a hospital serving a diverse population, the model underperforms for minority groups. What is the most effective way to address this ethical issue?

Accepted Answer

Retrain the model using a more diverse dataset that represents the target population.. Retrain the model using a more diverse dataset that represents the target population. The root cause of the underperformance is that the training data came from a single hospital with a predominantly white patient population, so the model learned patterns that do not generalize to minority groups; retraining on data that reflects the target population directly addresses this representativeness problem and is the most effective ethical remedy. Option A is problematic because separate models per demographic group can be discriminatory and may not be feasible for many groups, and it does not fix the underlying data gap. Option B is useful for detecting and reporting disparities but is only a monitoring/reporting measure, not a fix for the biased model. Option D merely adjusts a threshold to change sensitivity, which can mask rather than correct the underlying bias and may introduce other harms.

Answer

Create separate models for each demographic group to ensure accuracy.

Answer

Continuously monitor model performance across demographic groups and report disparities.

Answer

Adjust the decision threshold for minority groups to improve sensitivity.

Question 13

A company wants to use Einstein OCR to extract text from uploaded documents. To protect customer privacy, what should they ensure before processing documents containing personal data?

Accepted Answer

Enable the Einstein Trust Layer to mask sensitive data and comply with data protection policies.. The correct answer is D: Enable the Einstein Trust Layer to mask sensitive data and comply with data protection policies. Einstein Trust Layer provides built-in data masking, zero-retention, and audit capabilities that prevent personal data from being exposed to or retained by the LLM during Einstein OCR processing, which directly addresses the privacy requirement. Option A is insufficient because it shifts the burden to customers and does not enforce protection at the platform level. Option B only secures data in transit via TLS/HTTPS and does nothing about masking or handling sensitive data once received. Option C is incorrect because relying on the model to ignore sensitive data is not a reliable safeguard and consent alone does not satisfy data protection obligations.

Answer

Ask customers to manually blur sensitive information before uploading.

Answer

Use HTTPS to securely upload documents.

Answer

Obtain consent from customers and rely on the model to ignore sensitive data.

Question 14

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.. Option A is correct because the bias originates in the training data, which overrepresents male successful hires, so re-sampling to include more female candidates combined with fairness-aware algorithms addresses the root cause and enforces equitable outcomes during training. Fairness-aware methods (e.g., reweighting, adversarial debiasing, or constrained optimization) explicitly optimize for group fairness metrics rather than relying on the skewed historical distribution. Option B is weaker because post-processing score adjustments only mask the symptom and can be brittle or legally questionable without fixing the underlying data imbalance. Option C is wrong because accepting bias perpetuates discriminatory outcomes and likely violates anti-discrimination law. Option D is insufficient because removing the gender feature does not eliminate bias, since correlated proxy features (e.g., names, schools, career gaps) still encode gender information.

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 15

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 16

A Salesforce admin is configuring Einstein Bots. Which TWO actions are essential to maintain ethical AI practices?

Accepted Answer

Provide a clear option to escalate to a human agent. Option A is correct because ethical AI in Einstein Bots requires transparency and user agency: offering a clear escalation path to a human agent ensures users are not trapped in automated interactions and can obtain human help for sensitive or unresolved issues. Option D is correct because ongoing monitoring of bot conversations for biased or offensive responses is essential to detect and remediate harmful outputs, aligning with responsible AI governance and continuous quality control. Option B is not appropriate because retaining all conversation logs indefinitely raises data-privacy and data-minimization concerns under regulations like GDPR; retention should be limited and purposeful. Option C is incorrect because disabling feedback loops does not improve ethics and would prevent the bot from learning and improving; feedback loops should be governed, not eliminated. Option E is incorrect because operating without human oversight removes accountability and the ability to intervene, which is contrary to ethical AI practices.

Answer

Log all conversations indefinitely for analysis

Answer

Disable all feedback loops to avoid data contamination

Answer

Let the bot operate without human oversight

Question 17

Which TWO are best practices for mitigating bias in AI models?

Accepted Answer

Using balanced training datasets.. Option B is correct because using balanced training datasets ensures that the data used to train the AI model proportionally represents all relevant groups, which directly reduces the risk of the model learning and amplifying skewed patterns or underrepresenting certain populations. Option E is correct because applying disparity analysis involves measuring outcomes across different demographic groups (e.g., via metrics like demographic parity, equalized odds, or disparate impact ratios) to detect and quantify bias, which is a fundamental step in identifying and mitigating unfair model behavior. Option A is not a best practice for bias mitigation because complex deep learning models can increase opacity and make bias harder to detect, and model complexity itself does not address data or algorithmic fairness. Option C is not correct because simply removing sensitive attributes does not eliminate bias, since proxy variables and correlated features can still encode protected information. Option D is not correct because training only on the most recent data can introduce temporal bias, exclude historically underrepresented groups, and does not by itself ensure fairness across populations.

Answer

Using complex deep learning models.

Answer

Removing all sensitive attributes.

Answer

Training models on the most recent data only.

Question 18

Refer to the exhibit. An administrator runs an audit on a sentiment analysis model. What is the primary ethical concern?

Accepted Answer

The model has low accuracy on negative text, indicating potential bias.. The correct answer is C: the model has low accuracy on negative text, indicating potential bias. In a sentiment analysis audit, a marked disparity in performance across classes—here, degraded accuracy on negative samples—is the primary ethical concern because it means the model systematically misclassifies or under-serves a specific group of inputs, producing unfair or harmful outcomes. Option A is wrong because a bias score below threshold does not by itself establish that no action is needed, and it ignores the observed class-specific accuracy gap. Option B is not the primary ethical concern here; imbalanced training data is a possible root cause, but the exhibit's ethical issue is the resulting biased performance. Option D is incorrect because overfitting to positive text would typically show high training accuracy and poor generalization, not specifically low accuracy on negative text as the ethical red flag.

Answer

The bias score is below threshold so no action needed.

Answer

The training data is imbalanced.

Answer

The model is overfitting to positive text.

Question 19

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 20

A company wants to use customer data to train an AI model. Which ethical consideration is paramount?

Accepted Answer

Data minimization. Data minimization (option C) is the paramount ethical consideration here because training an AI model on customer data requires collecting and using only the personal data strictly necessary for the stated purpose, which reduces privacy risk and aligns with principles such as GDPR's purpose limitation and data minimization requirements. Model accuracy (A) is a technical performance goal, not an ethical safeguard, and pursuing it by ingesting more customer data than needed can worsen privacy harms. Cost efficiency (B) and speed of deployment (D) are operational and business concerns that do not address the ethical duty to protect customer data, so they are not the primary consideration.

Answer

Model accuracy

Answer

Cost efficiency

Answer

Speed of deployment