View Full PMI CPMAI Exam Dumps and Practice Test Dumps.

 

Question 221

Which activity should be performed first when defining the scope of an AI project?

1. Select the final machine learning algorithm
2. Identify the business problem, objectives, boundaries, and expected outcomes
3. Configure the production infrastructure
4. Begin model retraining

Correct Answer: 2

Explanation

Defining project scope begins with understanding the business problem and establishing what the AI initiative is intended to accomplish. The team should clarify objectives, expected outcomes, major deliverables, boundaries, assumptions, constraints, and exclusions. This provides a foundation for determining whether AI is appropriate and what technical solution may be required. Selecting an algorithm or configuring infrastructure before understanding the problem can lead to unnecessary complexity and rework. A well-defined scope also helps stakeholders maintain consistent expectations throughout the project. As requirements evolve, scope can be reviewed through an established change-control process rather than allowing uncontrolled expansion of project responsibilities.

Question 222

Which data-quality dimension addresses whether required data values are present?

1. Consistency
2. Accuracy
3. Completeness
4. Timeliness

Correct Answer: 3

Explanation

Completeness measures whether the required information is available and whether important fields contain expected values. For example, if a customer dataset requires an email address but many records have missing email values, the dataset has a completeness issue. Completeness does not necessarily mean that the available values are correct. A dataset can be complete but contain inaccurate information. Accuracy measures correctness, consistency examines agreement between related data sources or records, and timeliness considers whether data is sufficiently current. Assessing completeness is particularly important before AI development because missing information can affect feature quality, model training, evaluation, and ultimately the reliability of predictions or generated outputs.

Question 223

What is the primary purpose of a risk register in an AI project?

1. To document identified risks, their characteristics, owners, responses, and status
2. To store training datasets
3. To replace the project schedule
4. To record only technical defects

Correct Answer: 1

Explanation

A risk register provides a structured record of identified uncertainties that could affect an AI project’s objectives. It may include descriptions of risks, probability, potential impact, risk owners, response strategies, triggers, and current status. AI projects can have data, security, privacy, technical, operational, compliance, financial, and organizational risks, so maintaining a centralized record helps the team monitor them consistently. A risk register is not a substitute for the project schedule and should not be limited to technical defects. Regular review allows the project team to identify changes in exposure, implement planned responses, and escalate risks when their potential impact exceeds established tolerance levels.

Question 224

Which approach is most appropriate for handling a sensitive dataset used to train an AI model?

1. Give every project member unrestricted access
2. Publish the dataset publicly for transparency
3. Apply appropriate access controls, minimization, protection, and governance
4. Remove all security requirements during development

Correct Answer: 3

Explanation

Sensitive datasets should be protected through controls appropriate to their classification and intended use. These controls may include data minimization, role-based access, encryption, secure storage, retention limits, monitoring, and appropriate governance procedures. Access should be limited to authorized personnel who require the information for legitimate project responsibilities. Publishing sensitive data or granting unrestricted access can create unnecessary privacy and security exposure. Teams should also document how data is collected, processed, stored, shared, and eventually disposed of. Applying these protections throughout the AI lifecycle helps reduce unauthorized disclosure and supports responsible data management while allowing legitimate development and testing activities to proceed.

Question 225

Which metric is generally most appropriate when a regression model’s average absolute prediction error is important to stakeholders?

1. Recall
2. Mean Absolute Error
3. Precision
4. F1 score

Correct Answer: 2

Explanation

Mean Absolute Error, or MAE, calculates the average absolute difference between predicted and actual values. It provides an intuitive measure of the typical magnitude of prediction error in the same units as the target variable. For example, if a model predicts demand, MAE can indicate the average number of units by which predictions differ from actual demand. Precision, recall, and F1 score are primarily classification metrics. MAE does not heavily amplify large errors in the same way that squared-error metrics do, which can make it useful when stakeholders want a straightforward representation of average prediction error. Metric selection should always reflect the business consequences of model errors.

Question 226

What is the main purpose of human-in-the-loop oversight in an AI system?

1. To eliminate the need for model monitoring
2. To allow people to review, approve, or intervene in relevant AI decisions
3. To guarantee perfect model accuracy
4. To prevent the system from collecting any data

Correct Answer: 2

Explanation

Human-in-the-loop oversight incorporates human review or intervention into appropriate points of an AI workflow. Depending on the use case, a person may approve recommendations, review uncertain outputs, handle exceptions, or make the final decision when consequences are significant. Human involvement does not guarantee perfect accuracy and does not eliminate the need for technical monitoring. The appropriate level of oversight depends on factors such as risk, decision impact, model reliability, regulatory expectations, and operational requirements. Effective human oversight also requires clear roles, sufficient training, usable interfaces, and procedures that allow reviewers to challenge or override AI outputs when necessary.

Question 227

Which practice best supports responsible use of generative AI outputs?

1. Treat every generated response as automatically correct
2. Remove all human review
3. Validate important outputs against reliable information and defined requirements
4. Allow generated content to bypass organizational controls

Correct Answer: 3

Explanation

Generative AI systems can produce plausible but incorrect, incomplete, outdated, or misleading outputs. For important applications, outputs should therefore be evaluated against reliable information, business requirements, and appropriate quality criteria. The required level of validation depends on the consequences of errors and the intended use. Human review may be appropriate for high-impact decisions or sensitive content. Organizations can also use grounding, retrieval mechanisms, structured evaluation, output controls, and clear usage policies. Treating generated content as automatically correct can introduce significant operational and reputational risks. Responsible use requires users to understand system limitations and apply suitable validation before relying on outputs.

Question 228

Which technique helps identify unusual observations that may require further investigation before model training?

1. Outlier detection
2. Tokenization
3. Encryption
4. Deployment orchestration

Correct Answer: 1

Explanation

Outlier detection identifies observations that differ substantially from expected patterns within a dataset. Such observations may represent genuine unusual events, data-entry errors, measurement problems, fraud, system failures, or other meaningful cases. Teams should investigate outliers rather than automatically deleting them because some unusual observations may contain valuable information. Depending on the use case, techniques such as statistical rules, clustering methods, distance-based approaches, or model-based methods may be used. Understanding outliers can improve data preparation and help prevent unexpected observations from disproportionately affecting model training. However, removing an outlier should be justified by evidence and the intended purpose of the AI system.

Question 229

What is the primary purpose of a model card or similar model documentation?

1. To provide information about the model’s purpose, limitations, evaluation, and intended use
2. To replace all security controls
3. To guarantee that the model will never drift
4. To automatically retrain the model

Correct Answer: 1

Explanation

Model documentation, including model cards or comparable records, provides structured information about an AI model and its intended use. It may describe the model’s purpose, training context, evaluation results, limitations, known risks, relevant datasets, and appropriate usage conditions. Such documentation improves transparency and helps users understand what the model was designed to do and where caution may be required. It does not automatically prevent model drift, replace security controls, or perform retraining. Maintaining documentation throughout the lifecycle is useful because models, data, environments, and intended applications may change. Updated documentation helps stakeholders make informed decisions about deployment and continued use.

Question 230

Which project-management practice is most useful when an AI stakeholder requests a major change to an approved requirement?

1. Implement the change immediately without analysis
2. Ignore the request
3. Evaluate the change through the established change-control process
4. Delete the original requirement

Correct Answer: 3

Explanation

A major requirement change should normally pass through an established change-control process. The project team can assess how the proposed change affects scope, schedule, budget, resources, architecture, data, testing, risks, and expected outcomes. Appropriate stakeholders can then determine whether the change should be approved, rejected, deferred, or modified. Implementing significant changes without analysis can introduce scope creep and unexpected project impacts. Ignoring valid requests can also prevent the project from responding to legitimate business needs. Change control provides a structured mechanism for managing evolving requirements while preserving visibility and accountability across the AI project lifecycle.

Question 231

Which characteristic describes a baseline model?

1. A simple reference approach used to establish a performance point for comparison
2. A production model that can never be changed
3. A model that always has the highest possible accuracy
4. A model used only for data encryption

Correct Answer: 1

Explanation

A baseline model provides a reference point against which more sophisticated approaches can be evaluated. It may be a simple statistical method, heuristic, or relatively straightforward machine learning model. Establishing a baseline helps determine whether additional complexity actually provides meaningful improvement. Without a baseline, teams may have difficulty judging whether a sophisticated model delivers sufficient incremental value. A baseline does not need to be the final production solution and can be replaced when a better approach is validated. Comparing candidate models with an appropriate baseline supports evidence-based decisions about performance, complexity, operational cost, and whether the additional sophistication is justified.

Question 232

Which practice is most useful for identifying whether a model’s performance has degraded after deployment?

1. Comparing current monitored performance with established benchmarks and thresholds
2. Deleting historical metrics
3. Disabling production monitoring
4. Changing the model without evaluation

Correct Answer: 1

Explanation

Post-deployment monitoring should compare current system behavior against established performance expectations. Relevant indicators may include accuracy, precision, recall, error rates, latency, data distributions, business outcomes, and other use-case-specific measures. Historical benchmarks provide a reference for identifying meaningful degradation. When monitoring detects a problem, the team can investigate whether the cause is data drift, concept drift, infrastructure issues, changing user behavior, implementation errors, or another factor. Deleting historical metrics makes comparison more difficult, while changing the model without evaluation can introduce additional risk. Continuous monitoring provides evidence for corrective actions such as recalibration, retraining, rollback, or controlled model replacement.

Question 233

Which data-management activity establishes where data originated and how it moved through processing stages?

1. Data lineage
2. Hyperparameter tuning
3. Model calibration
4. Feature scaling

Correct Answer: 1

Explanation

Data lineage documents the origin, movement, transformation, and processing history of data. It can show where information was collected, which systems processed it, what transformations were applied, and where the resulting data was used. Lineage supports traceability, data governance, troubleshooting, auditing, and reproducibility. In AI projects, it can also help teams understand which datasets contributed to model development and determine the potential impact of changes to upstream data sources. Hyperparameter tuning and calibration are model-development activities, while feature scaling transforms numerical variables. Maintaining reliable lineage becomes increasingly important as AI systems use multiple data sources and complex automated pipelines.

Question 234

What is a key benefit of using automated testing in an AI deployment pipeline?

1. It eliminates all model risks
2. It can consistently verify defined technical and functional conditions before release
3. It guarantees business success
4. It removes the need for human governance

Correct Answer: 2

Explanation

Automated testing allows teams to repeatedly verify predefined conditions as models and supporting software move through development and deployment processes. Tests may cover data validation, model behavior, API functionality, integration, security controls, performance, and other requirements. Automation improves consistency and can identify regressions earlier than relying entirely on manual testing. However, automated testing does not eliminate all AI risks or guarantee business success. Human governance remains important for evaluating context, risk, ethics, and business suitability. A strong deployment pipeline combines automated checks with appropriate human approvals and monitoring so that technical changes are controlled before they reach production.

Question 235

Which situation best illustrates underfitting?

1. A model performs extremely well on training data but poorly on unseen data
2. A model is too simple to capture important patterns and performs poorly on both training and validation data
3. A model has excessive access permissions
4. A model experiences unauthorized network access

Correct Answer: 2

Explanation

Underfitting occurs when a model is too simple or insufficiently capable of representing meaningful patterns in the data. As a result, it may perform poorly on both the training dataset and unseen validation or test data. Overfitting presents a different pattern, where a model performs very well on training data but significantly worse on unseen data. Underfitting may be addressed through improved feature engineering, an appropriate increase in model complexity, better training procedures, or improved data representation. However, increasing complexity should not be done automatically because unnecessary complexity can introduce other risks. Model behavior should be evaluated using suitable datasets and metrics before changes are made.

Question 236

Why is stakeholder communication important during an AI project?

1. It helps maintain shared understanding of objectives, progress, risks, decisions, and changes
2. It eliminates the need for documentation
3. It guarantees stakeholder agreement on every issue
4. It replaces technical validation

Correct Answer: 1

Explanation

AI projects involve stakeholders with different responsibilities, expectations, and levels of technical knowledge. Regular communication helps maintain a shared understanding of project objectives, progress, risks, assumptions, decisions, dependencies, and changes. It can also surface concerns early and provide opportunities to clarify requirements or adjust expectations. Communication does not eliminate the need for formal documentation or technical validation, and stakeholders may still disagree on particular issues. Effective communication should therefore be structured around the audience and project needs. Clear reporting, appropriate terminology, defined escalation paths, and timely updates help improve coordination and support informed decisions throughout the AI lifecycle.

Question 237

Which security threat involves manipulating training data so that the resulting model learns undesirable patterns?

1. Data poisoning
2. Load balancing
3. Data compression
4. Feature scaling

Correct Answer: 1

Explanation

Data poisoning occurs when malicious or otherwise harmful information is deliberately introduced into training data with the goal of influencing model behavior. Depending on the attack, poisoned data may cause incorrect classifications, biased behavior, hidden vulnerabilities, or degraded performance. Protection can include controlled data acquisition, source validation, access controls, data-quality checks, anomaly detection, provenance tracking, and review of unexpected changes in training datasets. Data poisoning is different from ordinary data-quality problems because it involves intentional manipulation or exploitation of the training process. Organizations should consider data integrity as part of AI security because model behavior depends heavily on the quality and trustworthiness of training information.

Question 238

Which approach is most appropriate when an AI model must operate within strict response-time requirements?

1. Evaluate and optimize inference latency as part of system performance testing
2. Focus only on training accuracy
3. Ignore infrastructure capacity
4. Remove all production monitoring

Correct Answer: 1

Explanation

When an AI system has strict response-time requirements, inference latency should be treated as an explicit performance requirement. Teams can measure end-to-end response time and identify whether delays arise from the model, data retrieval, network communication, infrastructure, or application components. Optimization techniques may include selecting an appropriate model architecture, improving hardware utilization, caching, efficient data pipelines, or other suitable engineering approaches. High training accuracy alone does not guarantee acceptable production performance. Testing should occur under realistic workload conditions because latency may increase with concurrent users or larger data volumes. Monitoring latency after deployment is also important because infrastructure and workload changes can affect performance.

Question 239

Which activity helps determine whether an AI initiative has delivered the expected organizational benefits after implementation?

1. Benefits realization monitoring
2. Removing project objectives
3. Ignoring user feedback
4. Deleting financial assumptions

Correct Answer: 1

Explanation

Benefits realization monitoring evaluates whether the AI initiative is producing the expected business or organizational outcomes. Depending on the project, benefits may include reduced processing time, improved forecasting, increased productivity, lower costs, better customer experience, or improved decision quality. Teams should define measurable benefit indicators and compare actual results with the expectations established during planning. Monitoring benefits can reveal whether the system is delivering value, whether adoption is sufficient, or whether additional changes are necessary. Technical model performance alone does not prove business success. An AI model can meet technical targets while failing to produce meaningful organizational benefits if users do not adopt it or the solution does not address the intended problem.

Question 240

What should an organization do when an AI system reaches the end of its useful lifecycle?

1. Continue operating it indefinitely without review
2. Retire it through a controlled process that addresses dependencies, data, records, access, and replacement needs
3. Delete every record immediately
4. Disable monitoring before retirement

Correct Answer: 2

Explanation

AI model retirement should be handled through a controlled lifecycle process rather than an immediate or undocumented shutdown. The organization should identify system dependencies, communicate the retirement decision, preserve required records, revoke or modify access appropriately, address data-retention obligations, and transition users or processes to a replacement solution when necessary. Historical model information may need to be retained for audit, regulatory, operational, or analytical purposes. Monitoring and security controls should remain appropriate until retirement activities are complete. A structured retirement process reduces the risk of disrupting dependent systems or losing important records and ensures that the organization can demonstrate how and why the AI system was removed from service.