đ Table of Contents
- How to Use These Questions
- Domain 1: Foundational Concepts (10 Questions)
- Domain 2: AI Development Lifecycle (10 Questions)
- Domain 3: Responsible AI (10 Questions)
- Domain 4: Risk Management (10 Questions)
- Domain 5: Governance & Regulation (10 Questions)
- Interpret Your Score
- Next Steps
- Frequently Asked Questions
The best way to prepare for the AIGP exam is to practice with realistic questions. This collection includes 50 scenario-based questions modeled after actual exam content, covering all five domains with detailed explanations for every answer.
Use these questions to assess your current knowledge, identify weak areas, and build the pattern recognition skills you'll need on exam day.
How to Use These Practice Questions
Simulate Exam Conditions
Allow ~1.8 minutes per question (exam pace). Don't look at answers until you've committed to a choice.
Track Your Score
Note which questions you get wrong. Pattern of errors reveals which domains need more study time.
Read All Explanations
Even for questions you got right, read the explanation to reinforce your understanding and learn nuances.
Retake After Study
After focused study on weak areas, retake these questions to measure improvement.
đŻ Target Score
Aim for 75%+ correct (38/50) on these practice questions before scheduling your exam. This buffer accounts for the fact that actual exam questions may differ from practice materials. If you're scoring below 70%, spend more time studying before attempting the real exam.
Domain 1: Foundational Concepts of AI (Questions 1-10)
Question 1 of 50
Domain 1
Easy
A machine learning model is trained on historical loan approval data that includes the outcomes (approved/denied). The model learns to predict loan approval based on applicant features. What type of machine learning is this?
Correct Answer: A
Supervised learning uses labeled training dataâdata that includes both inputs and known outputs (in this case, whether loans were approved or denied). The model learns to map inputs to outputs based on these examples.
Why other options are incorrect:
⢠B) Unsupervised learning works with unlabeled data, finding patterns without known outcomes.
⢠C) Reinforcement learning learns through trial and error with rewards/penalties, not from labeled historical data.
⢠D) Transfer learning is a technique for applying knowledge from one task to another, not a fundamental learning type.
Why other options are incorrect:
⢠B) Unsupervised learning works with unlabeled data, finding patterns without known outcomes.
⢠C) Reinforcement learning learns through trial and error with rewards/penalties, not from labeled historical data.
⢠D) Transfer learning is a technique for applying knowledge from one task to another, not a fundamental learning type.
Question 2 of 50
Domain 1
Medium
Scenario: A financial services company implements a neural network model for fraud detection. The model achieves 99.5% accuracy in testing but the compliance team struggles to explain to regulators why specific transactions are flagged as fraudulent.
What governance challenge does this scenario BEST illustrate?
Correct Answer: B
The "black box" problem is a fundamental governance challenge with complex neural networks. While these models can achieve high accuracy, their decision-making process is difficult to interpret and explainâa critical issue when regulatory accountability requires transparent reasoning.
Key governance implication: In regulated industries like financial services, the inability to explain AI decisions can violate regulatory requirements (e.g., GDPR Article 22's right to explanation, fair lending laws requiring reason codes).
Why other options are incorrect:
⢠A) Overfitting would cause poor generalization to new data, but the scenario shows high accuracy.
⢠C) Insufficient data typically causes poor performance, not explainability issues.
⢠D) Concept drift occurs post-deployment when data distributions change; not the issue described.
Key governance implication: In regulated industries like financial services, the inability to explain AI decisions can violate regulatory requirements (e.g., GDPR Article 22's right to explanation, fair lending laws requiring reason codes).
Why other options are incorrect:
⢠A) Overfitting would cause poor generalization to new data, but the scenario shows high accuracy.
⢠C) Insufficient data typically causes poor performance, not explainability issues.
⢠D) Concept drift occurs post-deployment when data distributions change; not the issue described.
Question 3 of 50
Domain 1
Easy
Which type of AI technology is MOST commonly used for automated document classification and sentiment analysis of customer feedback?
Correct Answer: C
Natural Language Processing (NLP) is the branch of AI focused on enabling computers to understand, interpret, and generate human language. Document classification and sentiment analysis are core NLP applications.
Common NLP applications include: Text classification, sentiment analysis, named entity recognition, machine translation, chatbots, and text summarization.
Why other options are incorrect:
⢠A) Computer vision processes images and video, not text.
⢠B) Robotics involves physical machines and movement.
⢠D) Reinforcement learning is a learning paradigm, not a specific application area.
Common NLP applications include: Text classification, sentiment analysis, named entity recognition, machine translation, chatbots, and text summarization.
Why other options are incorrect:
⢠A) Computer vision processes images and video, not text.
⢠B) Robotics involves physical machines and movement.
⢠D) Reinforcement learning is a learning paradigm, not a specific application area.
Question 4 of 50
Domain 1
Medium
A company is using a large language model (LLM) for customer service automation. What is the PRIMARY governance concern specific to generative AI systems like LLMs?
Correct Answer: D
Hallucinationsâgenerating content that sounds plausible but is factually incorrectâare a distinctive governance challenge for generative AI and LLMs. Unlike traditional ML that classifies or predicts from existing data, generative AI creates new content that may be completely fabricated.
Governance implications: For customer service, hallucinations could provide incorrect information about products, policies, or legal rights. This creates liability, reputational, and compliance risks that require specific controls (human review, grounding techniques, output verification).
Why other options are less correct:
While A, B, and C are legitimate concerns for many AI systems, they're not specific to generative AI. Hallucination is the unique challenge that distinguishes generative AI governance.
Governance implications: For customer service, hallucinations could provide incorrect information about products, policies, or legal rights. This creates liability, reputational, and compliance risks that require specific controls (human review, grounding techniques, output verification).
Why other options are less correct:
While A, B, and C are legitimate concerns for many AI systems, they're not specific to generative AI. Hallucination is the unique challenge that distinguishes generative AI governance.
Question 5 of 50
Domain 1
Hard
Scenario: A healthcare AI system trained to detect skin cancer from images performs excellently on the test dataset (98% accuracy). However, when deployed in a clinic serving primarily patients with darker skin tones, detection accuracy drops significantly.
What is the MOST likely technical cause of this performance degradation?
Correct Answer: B
Training data bias is the most likely cause. If the training dataset predominantly contained images of lighter skin tones, the model learned patterns that don't generalize well to darker skin tones. This is a well-documented problem in dermatology AI.
This illustrates a critical governance principle: AI systems can only be as good as their training data. Non-representative training data leads to discriminatory outcomes, even when no one intended discrimination. This is why data governanceâensuring diverse, representative training dataâis fundamental to AI governance.
Why other options are incorrect:
⢠A) Overfitting would show poor performance on ANY new data, not specifically one demographic group.
⢠C) Equipment quality is unlikely given the systematic pattern (affecting one demographic group consistently).
⢠D) Architecture issues would affect all predictions equally, not show demographic disparities.
This illustrates a critical governance principle: AI systems can only be as good as their training data. Non-representative training data leads to discriminatory outcomes, even when no one intended discrimination. This is why data governanceâensuring diverse, representative training dataâis fundamental to AI governance.
Why other options are incorrect:
⢠A) Overfitting would show poor performance on ANY new data, not specifically one demographic group.
⢠C) Equipment quality is unlikely given the systematic pattern (affecting one demographic group consistently).
⢠D) Architecture issues would affect all predictions equally, not show demographic disparities.
Question 6 of 50
Domain 1
Easy
What is the PRIMARY difference between deep learning and traditional machine learning?
Correct Answer: C
Deep learning is characterized by neural networks with multiple hidden layers (hence "deep") that can automatically learn hierarchical feature representations from raw data. Traditional ML typically requires manual feature engineering.
Governance implication: Deep learning's automatic feature learning makes it powerful but also makes it harder to understand what features the model is actually usingâcontributing to the "black box" problem.
Why other options are incorrect:
⢠A) Both can use labeled or unlabeled data depending on the task.
⢠B) Deep learning is used across domains (NLP, audio, tabular data).
⢠D) Traditional ML can outperform deep learning, especially with limited data.
Governance implication: Deep learning's automatic feature learning makes it powerful but also makes it harder to understand what features the model is actually usingâcontributing to the "black box" problem.
Why other options are incorrect:
⢠A) Both can use labeled or unlabeled data depending on the task.
⢠B) Deep learning is used across domains (NLP, audio, tabular data).
⢠D) Traditional ML can outperform deep learning, especially with limited data.
Question 7 of 50
Domain 1
Medium
A retail company wants to group customers into segments based on purchasing behavior WITHOUT predefined categories. Which machine learning approach is MOST appropriate?
Correct Answer: B
Unsupervised clustering is designed to discover natural groupings in data without predefined categories. Common algorithms include K-means, hierarchical clustering, and DBSCAN.
Key distinction: "Without predefined categories" signals unsupervised learning. If categories already existed (e.g., "high-value," "at-risk," "new"), you'd use supervised classification.
Why other options are incorrect:
⢠A) Supervised classification requires labeled categories to train on.
⢠C) Reinforcement learning learns through actions and rewards, not pattern discovery.
⢠D) Supervised regression predicts continuous values, not group membership.
Key distinction: "Without predefined categories" signals unsupervised learning. If categories already existed (e.g., "high-value," "at-risk," "new"), you'd use supervised classification.
Why other options are incorrect:
⢠A) Supervised classification requires labeled categories to train on.
⢠C) Reinforcement learning learns through actions and rewards, not pattern discovery.
⢠D) Supervised regression predicts continuous values, not group membership.
Question 8 of 50
Domain 1
Medium
An AI governance professional is reviewing a computer vision system used for quality control in manufacturing. What is a PRIMARY governance concern for such systems?
Correct Answer: D
Reliability and accuracy under varying conditions is the primary governance concern for manufacturing quality control AI. Changes in lighting, camera angles, product variations, or environmental conditions can cause performance degradation.
Context matters: Different AI applications have different primary risks. For quality control (not involving people), the main concern is operational reliability, not bias or privacy.
Why other options are less appropriate:
⢠A) Quality control systems inspect products, not employeesâno biometric data involved.
⢠B) Discrimination against protected groups isn't relevant when inspecting manufactured items.
⢠C) Manufacturing quality control doesn't typically involve financial services regulations.
Context matters: Different AI applications have different primary risks. For quality control (not involving people), the main concern is operational reliability, not bias or privacy.
Why other options are less appropriate:
⢠A) Quality control systems inspect products, not employeesâno biometric data involved.
⢠B) Discrimination against protected groups isn't relevant when inspecting manufactured items.
⢠C) Manufacturing quality control doesn't typically involve financial services regulations.
Question 9 of 50
Domain 1
Hard
A company is implementing transfer learning, using a pre-trained language model and fine-tuning it on their proprietary customer service data. From an AI governance perspective, what is the MOST significant risk of this approach?
Correct Answer: C
Inherited biases and problematic content are a significant governance risk with transfer learning. Pre-trained models carry patternsâincluding biasesâfrom their original training data. Fine-tuning on proprietary data doesn't necessarily remove these inherited issues.
Governance implications:
⢠You may not know what data the base model was trained on
⢠Due diligence on third-party models is essential
⢠Testing for inherited biases must be part of deployment criteria
Why other options are incorrect:
⢠A) Transfer learning typically requires LESS compute than training from scratch.
⢠B) Pre-trained models CAN be adapted; that's the purpose of fine-tuning.
⢠D) Transfer learning often IMPROVES performance, especially with limited domain data.
Governance implications:
⢠You may not know what data the base model was trained on
⢠Due diligence on third-party models is essential
⢠Testing for inherited biases must be part of deployment criteria
Why other options are incorrect:
⢠A) Transfer learning typically requires LESS compute than training from scratch.
⢠B) Pre-trained models CAN be adapted; that's the purpose of fine-tuning.
⢠D) Transfer learning often IMPROVES performance, especially with limited domain data.
Question 10 of 50
Domain 1
Easy
What term describes the process by which humans categorize data examples (e.g., marking emails as spam or not spam) to create training data for supervised learning?
Correct Answer: A
Data labeling (or annotation) is the process of adding labels or tags to raw data so it can be used for supervised learning. Quality of labels directly affects model quality.
Governance considerations for data labeling:
⢠Labeler consistency and quality control
⢠Labeler bias affecting ground truth
⢠Clear labeling guidelines and training
⢠Privacy of data being labeled
Why other options are incorrect:
⢠B) Feature engineering is creating input variables from raw data, not labeling outputs.
⢠C) Model inference is generating predictions from a trained model.
⢠D) Hyperparameter tuning is optimizing model configuration settings.
Governance considerations for data labeling:
⢠Labeler consistency and quality control
⢠Labeler bias affecting ground truth
⢠Clear labeling guidelines and training
⢠Privacy of data being labeled
Why other options are incorrect:
⢠B) Feature engineering is creating input variables from raw data, not labeling outputs.
⢠C) Model inference is generating predictions from a trained model.
⢠D) Hyperparameter tuning is optimizing model configuration settings.
Domain 2: AI Development Life Cycle (Questions 11-20)
Question 11 of 50
Domain 2
Medium
Scenario: A company deployed an AI model for demand forecasting 18 months ago. Recently, predictions have become significantly less accurate, even though the model code hasn't changed.
What is the MOST likely cause of this degradation?
Correct Answer: B
Data drift (also called concept drift) occurs when the statistical properties of the data the model was trained on differ from the data it encounters in production. This is a natural phenomenonâcustomer behaviors, market conditions, and patterns change over time.
Key governance requirement: Ongoing monitoring for drift is essential. Models should not be deployed and forgottenâthey need continuous performance evaluation and periodic retraining.
Why other options are incorrect:
⢠A) The scenario states the model code hasn't changed, ruling out software bugs.
⢠C) Poor design would cause problems from the start, not emerge after 18 months.
⢠D) Resource issues would affect speed/availability, not accuracy patterns.
Key governance requirement: Ongoing monitoring for drift is essential. Models should not be deployed and forgottenâthey need continuous performance evaluation and periodic retraining.
Why other options are incorrect:
⢠A) The scenario states the model code hasn't changed, ruling out software bugs.
⢠C) Poor design would cause problems from the start, not emerge after 18 months.
⢠D) Resource issues would affect speed/availability, not accuracy patterns.
Question 12 of 50
Domain 2
Easy
What is the PRIMARY purpose of a "model card" in AI documentation?
Correct Answer: C
Model cards are standardized documentation that provides essential information about an ML model, including: intended use cases, training data description, performance metrics across different groups, known limitations, and ethical considerations.
Governance value: Model cards promote transparency and informed decision-making. They help downstream users understand whether a model is appropriate for their use case and what risks to consider.
The concept was introduced by Mitchell et al. (Google) in 2019 and has become a best practice in responsible AI development.
Governance value: Model cards promote transparency and informed decision-making. They help downstream users understand whether a model is appropriate for their use case and what risks to consider.
The concept was introduced by Mitchell et al. (Google) in 2019 and has become a best practice in responsible AI development.
Domain 3: Implementing Responsible AI (Questions 21-30)
Question 21 of 50
Domain 3
Hard
Scenario: A hiring AI system shows the following metrics:
⢠80% of qualified male applicants are recommended for interviews
⢠80% of qualified female applicants are recommended for interviews
⢠However, 60% of total male applicants are recommended vs. 40% of total female applicants
⢠80% of qualified male applicants are recommended for interviews
⢠80% of qualified female applicants are recommended for interviews
⢠However, 60% of total male applicants are recommended vs. 40% of total female applicants
This system satisfies which fairness criterion while potentially violating another?
Correct Answer: D
Equal opportunity requires equal true positive rates across groupsâhere, qualified applicants of both genders have an 80% recommendation rate. This criterion is satisfied.
Demographic parity requires equal overall selection rates regardless of qualifications. The 60% vs 40% overall rates violate this criterion.
Key insight: Different fairness definitions can conflict. This scenario illustrates why choosing the appropriate fairness metric depends on context and valuesâthere's no single "correct" definition of fairness that works for all situations.
Demographic parity requires equal overall selection rates regardless of qualifications. The 60% vs 40% overall rates violate this criterion.
Key insight: Different fairness definitions can conflict. This scenario illustrates why choosing the appropriate fairness metric depends on context and valuesâthere's no single "correct" definition of fairness that works for all situations.
Domain 4: Risk Management for AI (Questions 31-40)
Question 31 of 50
Domain 4
Medium
According to the NIST AI Risk Management Framework, which function focuses on establishing AI governance policies, defining roles and responsibilities, and cultivating a risk-aware organizational culture?
Correct Answer: A
GOVERN is the foundational function of NIST AI RMF. It establishes the organizational context for AI risk management, including policies, accountability structures, and risk culture.
Key GOVERN activities include:
⢠Establishing AI governance policies and procedures
⢠Defining roles, responsibilities, and accountability
⢠Setting risk tolerance thresholds
⢠Building organizational risk culture
⢠Ensuring resources for risk management
Remember the four functions:
⢠GOVERN: Policies, roles, culture (organizational foundation)
⢠MAP: Context, stakeholders, impacts (understanding the AI system)
⢠MEASURE: Metrics, testing, monitoring (quantifying risks)
⢠MANAGE: Respond, prioritize, improve (taking action)
Key GOVERN activities include:
⢠Establishing AI governance policies and procedures
⢠Defining roles, responsibilities, and accountability
⢠Setting risk tolerance thresholds
⢠Building organizational risk culture
⢠Ensuring resources for risk management
Remember the four functions:
⢠GOVERN: Policies, roles, culture (organizational foundation)
⢠MAP: Context, stakeholders, impacts (understanding the AI system)
⢠MEASURE: Metrics, testing, monitoring (quantifying risks)
⢠MANAGE: Respond, prioritize, improve (taking action)
Domain 5: AI Governance & Regulatory Landscape (Questions 41-50)
Question 41 of 50
Domain 5
Medium
Under the EU AI Act, which of the following AI applications is classified as PROHIBITED?
Correct Answer: B
Subliminal manipulation techniques that materially distort behavior and cause significant harm are explicitly prohibited under EU AI Act Article 5.
Other prohibited AI practices include:
⢠Exploiting vulnerabilities of specific groups (age, disability)
⢠Social scoring by public authorities
⢠Real-time remote biometric identification in public spaces (with limited exceptions)
⢠Emotion recognition in workplace and education (with exceptions)
Why other options are not prohibited:
⢠A) Credit scoring is HIGH-RISK, not prohibited
⢠C) Recruitment AI is HIGH-RISK, not prohibited
⢠D) Chatbots have LIMITED RISK transparency requirements only
Other prohibited AI practices include:
⢠Exploiting vulnerabilities of specific groups (age, disability)
⢠Social scoring by public authorities
⢠Real-time remote biometric identification in public spaces (with limited exceptions)
⢠Emotion recognition in workplace and education (with exceptions)
Why other options are not prohibited:
⢠A) Credit scoring is HIGH-RISK, not prohibited
⢠C) Recruitment AI is HIGH-RISK, not prohibited
⢠D) Chatbots have LIMITED RISK transparency requirements only
Question 42 of 50
Domain 5
Hard
Scenario: A European company is deploying an AI system to assist judges in criminal sentencing recommendations. Under the EU AI Act, this system would be classified as:
What is the correct EU AI Act classification for this system?
Correct Answer: C
High-risk is correct. AI systems used in the administration of justice and democratic processesâincluding assisting judicial decisionsâare explicitly listed as high-risk in Annex III of the EU AI Act.
High-risk AI in law enforcement/justice includes:
⢠AI assisting judges in researching and interpreting facts and law
⢠AI assisting in applying the law to concrete facts
⢠AI for crime prediction (individual risk assessment)
⢠AI for parole and probation decisions
Why not prohibited? The Act distinguishes between assistance (high-risk) and fully automated decisions affecting fundamental rights without human oversight (which could be prohibited depending on implementation).
High-risk AI in law enforcement/justice includes:
⢠AI assisting judges in researching and interpreting facts and law
⢠AI assisting in applying the law to concrete facts
⢠AI for crime prediction (individual risk assessment)
⢠AI for parole and probation decisions
Why not prohibited? The Act distinguishes between assistance (high-risk) and fully automated decisions affecting fundamental rights without human oversight (which could be prohibited depending on implementation).
đ Note: Sample Questions
This page includes representative questions from each domain. For the complete 50-question set with detailed explanations across all domains, plus hundreds more practice questions, access our full question bank through the main practice quiz section.
Interpret Your Score
| Score Range | Interpretation | Recommendation |
|---|---|---|
| 80-100% (40-50 correct) | Excellentâlikely ready to test | Schedule your exam; continue light review |
| 70-79% (35-39 correct) | Good foundationâsome gaps remain | Focus on weak domains; more practice needed |
| 60-69% (30-34 correct) | Borderlineâsignificant study needed | Do not schedule exam yet; intensive study required |
| Below 60% (< 30 correct) | Not readyâfundamental gaps | Full study program needed before retesting |
Analyze Your Domain Performance
Beyond overall score, track which domains you struggled with:
- Weak in Domain 1? Review AI/ML fundamentalsâtechnical concepts through a governance lens
- Weak in Domain 2? Study the AI development lifecycle and governance checkpoints at each stage
- Weak in Domain 3? Focus on responsible AI principles, fairness definitions, and organizational implementation
- Weak in Domain 4? Deep dive into NIST AI RMFâknow all four functions and their activities
- Weak in Domain 5? Study EU AI Act classifications and requirements extensively
Next Steps
â
How to Continue Your Preparation
Based on your performance on these questions:
- Review all explanationsâeven for questions you got right
- Identify your weakest domain(s) and allocate extra study time there
- Practice more questionsâaim for 300+ total before the exam
- Retake these questions after additional study to measure improvement
- Schedule your exam when consistently scoring 75%+
Frequently Asked Questions
Where can I find more free AIGP practice questions?
Free AIGP practice questions are available from IAPP (for members), this guide's 50 questions, and various exam prep platforms offer free trials. For comprehensive preparation with hundreds of questions, consider a dedicated question bank that tracks your performance by domain.
How many practice questions should I complete before the AIGP exam?
We recommend completing at least 300-500 practice questions before your exam. This volume helps build pattern recognition for exam-style questions and ensures you've seen the breadth of topics covered. Quality matters tooâfocus on understanding explanations, not just memorizing answers.
Are these questions representative of the actual AIGP exam?
These questions are modeled after the AIGP exam format and content based on the Body of Knowledge, but are not actual exam questions. The real exam may include different scenarios and question styles. Use these for learning and assessment, but don't assume exam questions will be identical.
What score should I aim for on practice questions?
Aim for 75%+ on practice questions before scheduling your exam. This buffer accounts for exam-day nerves and differences between practice and real questions. If you're consistently scoring 70% or below, spend more time studying before attempting the exam.
Ready for More Practice?
Access our complete question bank with 500+ questions, detailed explanations, and performance tracking across all domains.