AIGP Practice Test

AIGP practice test: what's covered

• Foundations
• EU AI Act
• NIST AI RMF
• ISO 42001
• AI Lifecycle
• Responsible AI

Free AIGP sample questions

1. A governance lead is briefing executives on terminology before launching an AI program. They need to define 'machine learning' precisely so the board does not conflate it with rule-based automation. Which description most accurately distinguishes machine learning from a traditional deterministic program?

   • A. A system whose behavior is explicitly hand-coded as fixed if-then rules by developers
   • B. A system that learns patterns and infers a function from training data rather than being explicitly programmed for each decision
   • C. Any software that automates a repetitive human task on a schedule
   • D. A database query engine that returns stored records matching a condition
   Show answer

   Correct: B. Machine learning derives a model from data so it can generalize to new inputs, unlike deterministic programs whose logic is hand-coded. Fixed if-then rules, scheduled automation, and query engines are all explicitly programmed and do not 'learn' from data.

2. An organization is categorizing its AI portfolio. One system is a large language model that generates novel marketing copy and images; another classifies incoming tickets into known categories. How should the governance team label these two systems respectively?

   • A. Both are discriminative AI
   • B. Generative AI and discriminative (predictive) AI
   • C. Both are generative AI
   • D. Symbolic AI and expert-system AI
   Show answer

   Correct: B. A model that produces new content (text, images) is generative AI; a model that assigns inputs to existing categories is discriminative or predictive AI. Labeling both the same way, or as symbolic/expert systems, misrepresents how each works and the distinct risks each carries.

3. During a model design review, a data scientist warns that the model fits the training data almost perfectly but performs poorly on a held-out validation set. Which condition is being described, and what is the governance concern?

   • A. Underfitting — the model is too simple and must be made more complex before release
   • B. Overfitting — the model memorized noise and may not generalize, raising reliability and fairness risk in production
   • C. Data minimization — too little data was collected to train the model
   • D. Differential privacy — noise was added so the model cannot reproduce the data
   Show answer

   Correct: B. High training accuracy with poor validation performance is the classic signature of overfitting: the model learned noise instead of the underlying signal. Underfitting is the opposite (poor on both), while data minimization and differential privacy are unrelated data-governance/privacy concepts.

4. A bank's credit model was trained on historical loan decisions that reflected past discriminatory lending. The model now systematically scores applicants from a protected group lower. What is the primary source of this harm?

   • A. Hardware acceleration bias introduced by the GPU
   • B. Historical (societal) bias embedded in the training data being learned and amplified by the model
   • C. Overfitting to the validation set
   • D. A poorly chosen random seed during training
   Show answer

   Correct: B. When training data encodes past discriminatory outcomes, the model learns and can amplify that historical bias. GPU acceleration and random seeds do not create societal bias, and overfitting to validation data is a generalization issue, not the cause of group-level discrimination here.

5. A product team wants to deploy a deep neural network for a high-stakes medical triage decision, but clinicians cannot understand why the model recommends what it does. Which responsible-AI principle is most directly at risk, and what mitigation best addresses it?

   • A. Privacy — apply data minimization to the training set
   • B. Transparency/explainability — provide post-hoc explanation techniques and clear documentation so affected parties can understand and contest decisions
   • C. Availability — add redundant servers to keep the model online
   • D. Scalability — retrain the model on a larger cluster
   Show answer

   Correct: B. A 'black box' in high-stakes use threatens transparency and explainability; explanation techniques plus documentation let clinicians understand and contest outputs. Privacy, availability and scalability address other properties and do not resolve the inability to understand the decision.

Last updated: 2026-10-19

AIGP practice test FAQ