I85PREP
⌘ K
AIGP study materials
AIGP study resources

AIGP practice questions

One question at a time. Build your AIGP understanding.

First 50 of 200 questions. Each one lists the choices, the correct answer, and an explanation. The checker at the top of this page saves progress in this browser when it is available.

Continue in the practice tool above to load more questions, 50 at a time.

Question 1Foundations of AI systems, their use cases and impacts

Which statement best characterizes the AI systems widely deployed in organizations today?

  1. They can reason across any domain at human level.
  2. They learn with no need for data or feedback.
  3. They possess self-awareness and independent intent.
  4. They are typically designed for specific, narrow tasks. Correct answer.

Explanation

AI systems in widespread commercial use are narrow: they are trained or configured for a defined task such as classification, ranking, or generation. General-purpose, self-directed reasoning at human level remains a research goal rather than a deployed capability, and every current learning approach still depends on data or feedback.

Question 2Foundations of AI systems, their use cases and impacts

A team trains a model on 50,000 emails, each already tagged 'spam' or 'not spam'. Which learning paradigm is being used?

  1. Supervised learning Correct answer.
  2. Unsupervised learning
  3. Reinforcement learning
  4. Self-supervised clustering

Explanation

Supervised learning trains on examples where each input is paired with the correct output, so the model learns a mapping from input to label. Spam detection using pre-tagged emails is the canonical case of that setup.

Question 3Foundations of AI systems, their use cases and impacts

A retailer groups its customers by purchasing behavior without defining any categories in advance. Which approach does this describe?

  1. Reinforcement learning
  2. Unsupervised clustering Correct answer.
  3. Supervised classification
  4. Rule-based expert system

Explanation

Unsupervised learning finds structure in data that carries no pre-assigned labels, and clustering customers by behavior is a typical use. The absence of predefined categories is the decisive signal that the approach is unsupervised.

Question 4Foundations of AI systems, their use cases and impacts

After a model has been trained and deployed, it scores incoming loan applications. What is this phase called?

  1. Inference Correct answer.
  2. Feature engineering
  3. Backpropagation
  4. Data labelling

Explanation

Inference is the phase in which a trained model is applied to new inputs to produce outputs. Training, by contrast, is the earlier phase in which model parameters are estimated from data.

Question 5Foundations of AI systems, their use cases and impacts

What distinguishes deep learning from classical machine learning?

  1. It produces only rule-based decision trees.
  2. It only works with structured tabular data held in spreadsheets.
  3. It requires no labelled training data and no human oversight.
  4. Layered neural networks learn representations automatically. Correct answer.

Explanation

Deep learning refers to neural networks with many hidden layers that learn feature representations directly from raw inputs rather than relying on hand-engineered features. It still requires substantial labelled data and computing power.

Question 6Foundations of AI systems, their use cases and impacts

In supervised learning, what is a label?

  1. A weight adjusted during training
  2. A metric used to score model accuracy
  3. The raw input values that are fed into the model during training
  4. The known target answer paired with a training example Correct answer.

Explanation

A label is the ground-truth output attached to a training example, which the model is optimized to reproduce. Inputs, learned weights, and evaluation metrics are separate concepts in the pipeline.

Question 7Foundations of AI systems, their use cases and impacts

Which task is the clearest example of computer vision?

  1. Generating a written summary of a meeting transcript
  2. Flagging pedestrians in vehicle camera frames Correct answer.
  3. Ranking search results by click-through rate
  4. Predicting next quarter's sales from historical tables

Explanation

Computer vision extracts information from images and video, such as detecting objects or people in camera frames. The other options involve tabular forecasting, text summarization, and ranking, which belong to different task families.

Question 8Foundations of AI systems, their use cases and impacts

How does a large language model generate text?

  1. By predicting the next token from prior context Correct answer.
  2. By applying hand-written grammar rules in sequence
  3. By ranking documents retrieved from the internet
  4. By looking up stored answers in a fixed answer database

Explanation

Large language models are autoregressive next-token predictors: they estimate a probability distribution over the next token given preceding text and sample from it repeatedly. They do not retrieve canned answers or follow hand-written grammar rules.

Question 9Foundations of AI systems, their use cases and impacts

A model scores 99% on its training data but only 62% on new customer records. What is the most likely cause?

  1. Underfitting caused by an overly simple model
  2. Overfitting to the training data Correct answer.
  3. Inference latency being too high
  4. A held-out test set that is too large

Explanation

A large gap between very high training performance and much lower performance on unseen records is the signature of overfitting, where the model memorizes idiosyncrasies of the training set. Typical remedies include more data, regularization, or a simpler model.

Question 10Foundations of AI systems, their use cases and impacts

Why is a held-out test set kept separate during model development?

  1. To speed up the training process on very large datasets
  2. To tune hyperparameters before the final training run
  3. To estimate performance on unseen data Correct answer.
  4. To store backups of the training data

Explanation

A held-out test set is never used for fitting, so its results approximate how the model will behave on genuinely new data. Repeatedly tuning against the test set leaks information and inflates the reported performance.

Question 11Foundations of AI systems, their use cases and impacts

A robotics team rewards a simulated agent for each successful grasp and penalizes collisions. Which paradigm is being used?

  1. Rule-based expert system
  2. Supervised learning
  3. Reinforcement learning Correct answer.
  4. Unsupervised learning

Explanation

Reinforcement learning trains an agent by rewarding desired actions and penalizing undesired ones through interaction with an environment. No dataset of labelled correct actions is supplied to the agent.

Question 12Foundations of AI systems, their use cases and impacts

What is a defining characteristic of a foundation model?

  1. It is pretrained broadly and adapted to many tasks Correct answer.
  2. It requires no compute resources to run
  3. It is trained from scratch on data from one narrow task only
  4. It produces only numerical outputs, never text

Explanation

Foundation models are trained on broad data at scale and then adapted, through prompting or fine-tuning, to a wide range of downstream tasks. This contrasts with task-specific models built for a single purpose.

Question 13Foundations of AI systems, their use cases and impacts

A streaming service suggests films by identifying users whose viewing histories resemble a given user's history. What is this technique called?

  1. Optical character recognition
  2. Speech synthesis
  3. Collaborative filtering Correct answer.
  4. Content moderation

Explanation

Collaborative filtering recommends items by exploiting similarity between users' or items' interaction histories. It can operate without any understanding of the content of the items themselves.

Question 14Foundations of AI systems, their use cases and impacts

A hiring model trained on a decade of past decisions consistently ranks candidates from one demographic lower. What is the most direct explanation?

  1. Bias in historical labels was learned by the model Correct answer.
  2. The model has too few parameters to learn properly
  3. The inference hardware is misconfigured
  4. The model was trained for too many epochs

Explanation

Models learn whatever patterns are present in their training data, including historical inequities embedded in the labels. Disparate outcomes across groups therefore usually trace back to the training signal rather than to hardware or capacity.

Question 15Foundations of AI systems, their use cases and impacts

What is the primary purpose of human-in-the-loop review in a high-stakes AI workflow?

  1. To convert unstructured data into a structured format
  2. To reduce the volume of training data needed
  3. To let a person check or override high-risk outputs Correct answer.
  4. To eliminate the need for any model monitoring

Explanation

Human-in-the-loop review places a person in the decision path to verify, correct, or escalate model outputs where errors carry serious consequences. It complements rather than replaces model monitoring and governance.

Question 16Foundations of AI systems, their use cases and impacts

A fraud detection model degrades sharply after customers begin paying in a new way. Which phenomenon does this illustrate?

  1. A feature scaling error
  2. Underfitting of the training data
  3. Overfitting to the test set
  4. Model drift Correct answer.

Explanation

Model drift is the deterioration of predictive performance as the world diverges from the conditions captured during training. It can be driven by changes in the inputs, in the input-to-output relationship, or in both.

Question 17Foundations of AI systems, their use cases and impacts

Why is explainability especially important for AI used in credit decisions?

  1. It removes the need for training data
  2. Applicants may need reasons for a denial Correct answer.
  3. It makes the model train faster on GPUs
  4. It increases the raw accuracy of the model

Explanation

Explainability supports accountability: decision subjects and reviewers need to understand the basis of an adverse decision and to challenge it where appropriate. It does not itself improve accuracy, reduce data needs, or accelerate training.

Question 18Foundations of AI systems, their use cases and impacts

Which activity belongs to the post-deployment stage of the AI lifecycle?

  1. Monitoring live performance and retraining when needed Correct answer.
  2. Defining the business problem and success metrics
  3. Collecting and labelling the initial training data set for the project
  4. Selecting the model architecture and loss function

Explanation

Post-deployment work centers on operating the model in production, which includes monitoring performance and data quality and retraining when conditions change. Data collection, architecture selection, and problem definition all precede deployment.

Question 19Foundations of AI systems, their use cases and impacts

Classifying product reviews as positive or negative is a task within which area?

  1. Speech synthesis
  2. Natural language processing Correct answer.
  3. Computer vision
  4. Robotic process automation

Explanation

Sentiment classification operates on written text, which places it within natural language processing. Vision, robotic process automation, and speech synthesis address different data modalities or process functions.

Question 20Foundations of AI systems, their use cases and impacts

A model predicts house prices from size, number of bedrooms, and location. These inputs are called:

  1. Hyperparameters
  2. Features Correct answer.
  3. Labels
  4. Epochs

Explanation

Features are the input variables a model uses to make a prediction. Labels are the target outputs, epochs are passes through the training data, and hyperparameters configure the learning process.

Question 21Foundations of AI systems, their use cases and impacts

A hospital has only 2,000 labelled X-rays. Fine-tuning a model pretrained on millions of general images outperforms training from scratch. Why?

  1. Pretraining encodes visual features that transfer to the new task Correct answer.
  2. Pretraining removes all bias from the original image data
  3. Fine-tuning eliminates the need for any labelled X-rays
  4. Fine-tuning guarantees the model cannot overfit small datasets

Explanation

Pretraining on large image corpora produces generic visual features, such as edge and texture detectors, that fine-tuning can repurpose with modest amounts of labelled data. Fine-tuning still requires some labels and offers no guarantee against overfitting or inherited bias.

Question 22Foundations of AI systems, their use cases and impacts

How does retrieval-augmented generation reduce hallucination in a language model assistant?

  1. It retrains all of the model's weights from scratch on every user query
  2. It blocks the model from producing any text output
  3. It lowers the sampling temperature all the way to exactly zero
  4. It places retrieved documents in the prompt so answers are grounded Correct answer.

Explanation

RAG retrieves relevant passages at query time and inserts them into the prompt, so the model conditions its answer on source material. This grounds outputs and reduces fabrication without changing the model's weights.

Question 23Foundations of AI systems, their use cases and impacts

How does a classical expert system differ from a neural network?

  1. It requires GPUs to perform inference at scale
  2. It applies rules written explicitly by domain experts Correct answer.
  3. It learns patterns from large datasets of labelled examples
  4. It cannot represent domain knowledge in any form

Explanation

Expert systems encode human-authored if-then rules and a curated knowledge base, so their reasoning is explicit and inspectable. Neural networks instead learn statistical patterns from data and are not hand-programmed with rules.

Question 24Foundations of AI systems, their use cases and impacts

A factory runs defect detection on a device attached to the production line rather than sending video to the cloud. What is the main benefit?

  1. It eliminates all hardware costs
  2. Model accuracy always increases when inference runs at the edge
  3. Lower latency and less exposure of sensitive footage Correct answer.
  4. It removes the need to train the model at all

Explanation

Running inference locally on the device cuts the round trip to a remote server and keeps sensitive video on the premises. Edge deployment changes where computation happens, not whether training is required.

Question 25Foundations of AI systems, their use cases and impacts

A model's inputs keep the same distribution, but the relationship between those inputs and the target changes after a policy change. What is this called?

  1. Data drift in the input features
  2. Overfitting to the training set
  3. Class imbalance in the labels
  4. Concept drift Correct answer.

Explanation

Concept drift is a change in the relationship between inputs and the target, so identical inputs now imply different outcomes. Data drift, by contrast, is a change in the distribution of the inputs themselves.

Question 26Foundations of AI systems, their use cases and impacts

Which description best characterizes a 'narrow' AI system?

  1. A system that can rewrite its own objectives without human input
  2. A system that matches human performance across every cognitive task
  3. A system built for one bounded task, such as transcribing speech Correct answer.
  4. A system that needs no data to reach expert-level accuracy

Explanation

Nearly all deployed AI is narrow: it is optimized for a defined task and performs poorly outside that task's scope. Generality across all cognitive tasks, self-authored goals, and data-free competence are not properties of current production systems.

Question 27Foundations of AI systems, their use cases and impacts

After a model has been trained and deployed, the process of using it to score new inputs is called:

  1. Annotation
  2. Tokenization
  3. Inference Correct answer.
  4. Backpropagation

Explanation

Training adjusts the model's parameters; inference is the forward pass that applies the trained parameters to new data to produce an output. In production, inference dominates compute cost and latency.

Question 28Foundations of AI systems, their use cases and impacts

Which characteristic distinguishes unsupervised learning from supervised learning?

  1. It can only be applied to image data
  2. It works with data that carries no target labels Correct answer.
  3. It always requires a human to label every input
  4. It guarantees higher accuracy than supervised methods

Explanation

Unsupervised methods look for structure such as clusters, groupings, or lower-dimensional representations without a target variable. Supervised learning, by contrast, requires labelled examples pairing inputs with the desired output.

Question 29Foundations of AI systems, their use cases and impacts

Converting a recording of a customer's spoken complaint into written text relies on which capability?

  1. Machine translation
  2. Automatic speech recognition Correct answer.
  3. Image segmentation
  4. Optical character recognition

Explanation

Automatic speech recognition maps an audio waveform onto a sequence of words. The other options operate on images or on text in a different language.

Question 30Foundations of AI systems, their use cases and impacts

What distinguishes generative AI models from traditional predictive models?

  1. They are trained without any labelled data at all
  2. They never make errors on unseen inputs
  3. They can only produce numeric risk scores
  4. They generate new content such as text or images Correct answer.

Explanation

Generative models learn the distribution of their training data well enough to synthesize new samples, such as passages of text or pictures. Classical predictive models instead output a discrete class or a numeric estimate.

Question 31Foundations of AI systems, their use cases and impacts

A team wants to analyze customer records stored as rows and columns in a relational database. How should this data be described?

  1. Synthetic data generated by a model
  2. Structured, tabular data Correct answer.
  3. Unstructured data such as free-text notes
  4. Streaming sensor telemetry

Explanation

Rows and columns with a predefined schema are the defining form of structured data, which maps cleanly onto spreadsheets and database tables. Free text, images, and audio are unstructured because they lack that fixed schema.

Question 32Foundations of AI systems, their use cases and impacts

Which statement correctly describes the relationship between AI, machine learning, and deep learning?

  1. Deep learning is the broadest of the three categories overall
  2. AI is a subset of machine learning restricted to neural networks
  3. Machine learning and deep learning are unrelated disciplines
  4. Deep learning sits inside machine learning, which sits inside AI Correct answer.

Explanation

The terms form nested sets: AI is the broad goal of machine intelligence, machine learning is the subset that learns from data, and deep learning is the subset of machine learning built on multi-layer neural networks.

Question 33Foundations of AI systems, their use cases and impacts

Which of the following is a hyperparameter?

  1. A single pixel value in a training image
  2. The learning rate chosen before training begins Correct answer.
  3. The weight a model learns for one input feature
  4. The value of the target in a test record

Explanation

Hyperparameters are settings fixed by the practitioner before training, such as learning rate, batch size, or tree depth. Parameters, by contrast, are learned from the data during training.

Question 34Foundations of AI systems, their use cases and impacts

A bank wants a model that assigns each loan applicant to one of two risk grades. Which task type is this?

  1. Clustering
  2. Dimensionality reduction
  3. Regression
  4. Binary classification Correct answer.

Explanation

Assigning each case to one of two discrete categories is binary classification. Regression would apply if the output were a continuous quantity such as an exact interest rate.

Question 35Foundations of AI systems, their use cases and impacts

A model flagged 200 transactions, and 150 of those were genuinely fraudulent. Over the same period there were 300 actual frauds, of which the model caught 150. What is the model's precision?

  1. 0.33
  2. 0.75 Correct answer.
  3. 0.60
  4. 0.50

Explanation

Precision is the share of positive predictions that are correct, here 150 true positives out of 200 flagged, or 0.75. The 150 of 300 caught figure is recall, a different metric.

Question 36Foundations of AI systems, their use cases and impacts

A dataset contains 1,000 fraudulent transactions among 1,000,000 total. A model that always predicts 'not fraud' reports 99.9% accuracy. What does this demonstrate?

  1. The model has learned the fraud pattern with high precision
  2. The dataset is too small to train any reliable model
  3. Accuracy hides failure when one class dominates Correct answer.
  4. Accuracy is only a valid metric for regression tasks

Explanation

When the negative class is overwhelming, a trivial predictor achieves high accuracy while catching zero frauds. Metrics focused on the minority class, such as recall, precision, or F1, are needed alongside accuracy.

Question 37Foundations of AI systems, their use cases and impacts

In semi-supervised learning, the training data consists of:

  1. Labels that are withheld from the model at prediction time
  2. Some labelled data plus a larger unlabelled set Correct answer.
  3. Synthetic examples that have no relation to real data
  4. Only unlabelled examples, with no target values at all

Explanation

Semi-supervised learning combines a small labelled set with a much larger unlabelled set, using the unlabelled data to learn structure that improves the decision boundary. It is useful when labelling is costly but raw data is plentiful.

Question 38Foundations of AI systems, their use cases and impacts

Why does an ensemble of decision trees often generalize better than a single tree?

  1. It removes the need for a labelled training set altogether
  2. Averaging varied trees reduces the variance of predictions Correct answer.
  3. It guarantees the absence of bias in every prediction
  4. It makes each individual tree easier to interpret

Explanation

Individual deep trees are unstable: small changes in the data can alter them substantially. Combining many decorrelated trees cancels much of that variability, lowering variance and improving generalization.

Question 39Foundations of AI systems, their use cases and impacts

A team discovers that 12% of the dates in its training records are in the wrong format, producing impossible values. What is the most appropriate first step?

  1. Profile and clean the data before training any model Correct answer.
  2. Deploy the model and monitor for errors
  3. Add more layers to the neural network to compensate
  4. Increase the learning rate so the model adapts

Explanation

Corrupted fields violate the assumptions of both the data pipeline and the model, so they must be profiled and corrected before training. Model-side changes cannot recover signal from values that are simply wrong.

Question 40Foundations of AI systems, their use cases and impacts

Why can including personal data in a model's training set create a privacy risk even if the raw records are never published?

  1. Training data is always publicly downloadable by anyone
  2. Models can memorize and later reveal training records Correct answer.
  3. Personal data may not legally be stored in any database
  4. Privacy law prohibits the use of any purely numeric data

Explanation

Models, especially large ones, can retain enough detail about individual examples to reproduce them when prompted or probed. This means protection must extend beyond access controls on the original dataset.

Question 41Foundations of AI systems, their use cases and impacts

A dermatology model is trained only on images from a single hospital serving one demographic group. What is the main risk?

  1. The model will store copies of every image it saw
  2. It may perform poorly for under-represented skin types Correct answer.
  3. The model will need fewer images than comparable models
  4. The model will train too slowly to be practical

Explanation

Training data that is not representative produces a model tuned to the patterns of the sampled population. Accuracy typically degrades for groups whose appearance or conditions were rare or absent during training.

Question 42Foundations of AI systems, their use cases and impacts

A radiologist accepts nearly every AI suggestion without independent review, including some confident errors. What does this illustrate?

  1. Automation bias: the reviewer defers to the system Correct answer.
  2. Data drift affecting the imaging pipeline over time
  3. Overfitting of the underlying diagnostic model
  4. A limitation inherent to unsupervised learning methods

Explanation

Automation bias is the human tendency to over-trust automated advice, which erodes the independent check that human oversight is meant to provide. Mitigations include requiring the reviewer to record their own assessment before seeing the model output.

Question 43Foundations of AI systems, their use cases and impacts

In a shadow deployment, a new model:

  1. Receives live inputs but its outputs are only logged Correct answer.
  2. Is trained on live production traffic and never evaluated offline
  3. Is deployed to all users only after each one opts in
  4. Replaces the current model for a random half of all users

Explanation

Shadow mode runs the candidate model on real traffic while the incumbent continues to serve decisions, so the team can compare outputs without user risk. It is a common step between offline evaluation and full rollout.

Question 44Foundations of AI systems, their use cases and impacts

A support assistant reads user messages and also follows any instruction found inside retrieved documents, including ones telling it to reveal its system prompt. Which risk does this describe?

  1. Feature scaling error
  2. Class imbalance
  3. Model underfitting
  4. Indirect prompt injection Correct answer.

Explanation

Indirect prompt injection occurs when hostile instructions arrive through content the model ingests, such as retrieved pages or documents, rather than from the user directly. Defences include treating retrieved content as data and validating actions before execution.

Question 45Foundations of AI systems, their use cases and impacts

What is an embedding?

  1. A label attached to each training example by an annotator
  2. A numeric vector representing an item's meaning Correct answer.
  3. A filter rule that blocks unsafe model outputs
  4. A human-written summary of a model's overall behaviour

Explanation

An embedding maps an item such as a word, sentence, or image into a dense numeric vector so that similar items land close together in that space. Distances between embeddings power search, clustering, and recommendation.

Question 46Foundations of AI systems, their use cases and impacts

A model that detects pneumonia from X-rays works well at one hospital but fails elsewhere. Analysis shows that images containing a chest drain or a portable-scanner marker were nearly always positive in the training set. What best explains the failure?

  1. The model was trained with a learning rate that was too high
  2. The test images had a much higher resolution than the training images
  3. The model was accidentally trained with unsupervised learning
  4. The model learned a spurious cue correlated with the label Correct answer.

Explanation

The model latched onto incidental features that happened to correlate with the label in one setting, a shortcut rather than the clinical signal. Such spurious correlates do not transfer to sites where the correlation is absent or reversed.

Question 47Foundations of AI systems, their use cases and impacts

Why can a model rarely satisfy demographic parity and equalized odds simultaneously when group base rates differ?

  1. Equalized odds requires removing all protected attributes entirely
  2. Fairness can only be measured after a model is deployed
  3. The two criteria conflict when group base rates differ Correct answer.
  4. One of the metrics is defined only for regression models

Explanation

Formal results show that, apart from degenerate cases, equalized odds and demographic parity cannot both hold when the groups have different prevalence of the positive outcome. Teams must therefore choose a fairness criterion explicitly and justify the trade-off.

Question 48Foundations of AI systems, their use cases and impacts

A city deploys a model that sends patrols to predicted crime hotspots. Arrests rise in those areas, which supplies more training data from them. What dynamic does this describe?

  1. Regularization
  2. Catastrophic forgetting
  3. Data leakage
  4. A self-reinforcing feedback loop Correct answer.

Explanation

The model's own actions change the data that will train its successor, so predictions and enforcement reinforce each other. This endogenous loop can entrench initial patterns and make measured performance diverge from reality.

Question 49Foundations of AI systems, their use cases and impacts

A model reports exceptional scores on a public benchmark, but its training corpus contained the benchmark's questions and answers. What is this problem and why does it matter?

  1. Underfitting; the model has learned too little from its data
  2. Regularization; it improves generalization on unseen data
  3. Contamination; the score overstates real capability Correct answer.
  4. Pruning; it removes redundant parameters from the network

Explanation

When evaluation items appear in training data, the benchmark measures memorization rather than generalization. Reported results then overstate capability and can mislead procurement, deployment, and safety decisions.

Question 50Foundations of AI systems, their use cases and impacts

In AI safety, what does the alignment problem refer to?

  1. Aligning the database schemas that feed the training pipeline
  2. Matching the size of the training and test datasets exactly
  3. Synchronizing GPU clocks across distributed training workers
  4. Keeping systems' behaviour consistent with intended goals Correct answer.

Explanation

Alignment concerns whether a system actually optimizes what its designers intended, including in situations not anticipated during development. It covers specification errors, reward hacking, and behaviour that looks correct in testing but diverges in deployment.