Artificial Intelligence interview questions

25 questions from the Artificial Intelligence course, each with the correct answer and an explanation. Try to answer before you open the answer.

What Artificial Intelligence Is

1. Which approach dominated early AI research?

  1. Deep convolutional networks
  2. Large language models
  3. Symbolic methods: search, logic and hand-written rules
  4. GPU clusters
Show answer

Answer: Symbolic methods: search, logic and hand-written rules

Learning-based methods rose later.

Lesson: Definitions and a Short History

2. What does a simple reflex agent base its action on?

  1. Only the current percept, using condition-action rules
  2. A full model of the future
  3. A learned value function
  4. A conversation history
Show answer

Answer: Only the current percept, using condition-action rules

No memory, no planning.

Lesson: Intelligent Agents and Environments

3. What kind of AI exists in deployed systems today?

  1. AI with no data or rules
  2. Proven artificial general intelligence
  3. Conscious machines
  4. Narrow AI focused on specific tasks
Show answer

Answer: Narrow AI focused on specific tasks

Even broad models are evaluated task by task.

Lesson: Kinds of AI

Solving Problems by Search

4. What is the frontier in search?

  1. The goal state
  2. The set of discovered states waiting to be expanded
  3. The cheapest path found
  4. The list of all possible states
Show answer

Answer: The set of discovered states waiting to be expanded

Algorithms differ in how they order the frontier.

Lesson: Formulating a Search Problem

5. Which algorithm guarantees a shortest path when all steps cost the same?

  1. Depth-first search
  2. Breadth-first search
  3. Random walk
  4. Hill climbing
Show answer

Answer: Breadth-first search

BFS expands in order of distance.

Lesson: Breadth-First and Depth-First Search

6. What does A* minimise when choosing the next state?

  1. The number of neighbours
  2. Only the cost so far
  3. Only the heuristic
  4. g(n) + h(n): cost so far plus estimated cost to go
Show answer

Answer: g(n) + h(n): cost so far plus estimated cost to go

Combining both guides the search efficiently.

Lesson: A* Search and Heuristics

7. Why can hill climbing return a poor solution?

  1. It stops at a local optimum when no neighbour is better
  2. It never moves
  3. It always picks random states
  4. It requires a goal test
Show answer

Answer: It stops at a local optimum when no neighbour is better

Accepting some worse moves helps escape.

Lesson: Local Search: Hill Climbing and Simulated Annealing

Games and Constraints

8. What does minimax assume about the opponent?

  1. The opponent never moves
  2. The opponent plays randomly
  3. The opponent always chooses the move that is best for them
  4. The opponent helps you win
Show answer

Answer: The opponent always chooses the move that is best for them

Plan for the strongest reply.

Lesson: Minimax: Playing Against an Opponent

9. How does alpha-beta's result compare with minimax?

  1. It searches more nodes
  2. It is always less accurate
  3. It returns the same value while visiting fewer nodes
  4. It ignores the opponent
Show answer

Answer: It returns the same value while visiting fewer nodes

Pruning removes only irrelevant branches.

Lesson: Alpha-Beta Pruning

10. Which heuristic picks the next variable in backtracking?

  1. Most constrained variable first
  2. Alphabetical order always
  3. The variable with the most values left
  4. A random variable each time, ignoring constraints
Show answer

Answer: Most constrained variable first

Tackle the hardest choice early to fail fast.

Lesson: Constraint Satisfaction Problems

Knowledge and Reasoning

11. What does forward chaining do?

  1. Searches a game tree
  2. Starts from a goal and works backwards
  3. Trains a neural network
  4. Applies rules to known facts repeatedly to derive new facts
Show answer

Answer: Applies rules to known facts repeatedly to derive new facts

Data-driven inference.

Lesson: Logic and Rule-Based Reasoning

12. How does a knowledge graph store facts?

  1. As pixel arrays
  2. As subject-relation-object triples
  3. As neural network weights only
  4. As unstructured paragraphs
Show answer

Answer: As subject-relation-object triples

Triples make relationships queryable.

Lesson: Knowledge Graphs

13. What is the knowledge acquisition bottleneck?

  1. A shortage of GPUs
  2. The difficulty of writing and maintaining enough explicit rules
  3. Slow internet
  4. Too many training examples
Show answer

Answer: The difficulty of writing and maintaining enough explicit rules

Encoding knowledge by hand does not scale.

Lesson: The Limits of Symbolic AI

Reasoning Under Uncertainty

14. Why can a positive result from an accurate test still be unlikely to be correct?

  1. Probabilities cannot be updated
  2. Accurate tests are always wrong
  3. Bayes' rule ignores priors
  4. When the condition is rare, false positives can outnumber true positives
Show answer

Answer: When the condition is rare, false positives can outnumber true positives

Base rates matter.

Lesson: Probability and Bayes' Rule

15. What is the "naive" assumption in naive Bayes?

  1. Features are independent of each other given the class
  2. All classes are equally likely
  3. There is only one feature
  4. Words never repeat
Show answer

Answer: Features are independent of each other given the class

A strong but useful simplification.

Lesson: Naive Bayes Classification

16. How does a rational agent choose an action under uncertainty?

  1. At random
  2. By always choosing the most likely outcome
  3. By maximising expected utility
  4. By minimising the number of outcomes
Show answer

Answer: By maximising expected utility

Weigh outcomes by probability and value.

Lesson: Decisions and Expected Utility

Learning Agents

17. When is the perceptron guaranteed to converge?

  1. Only on images
  2. Always, on any data
  3. When the classes are linearly separable
  4. Only with one example
Show answer

Answer: When the classes are linearly separable

Separable data guarantees convergence.

Lesson: Learning From Examples: The Perceptron

18. What does value iteration need that Q-learning does not?

  1. A known model of transitions and rewards
  2. A reward signal
  3. States
  4. Actions
Show answer

Answer: A known model of transitions and rewards

Value iteration plans with a model; Q-learning learns from experience.

Lesson: Markov Decision Processes and Value Iteration

19. Why does a Q-learning agent sometimes take random actions?

  1. To avoid using rewards
  2. To slow learning down on purpose
  3. Because Q-values are random
  4. To explore actions that might be better than its current estimates
Show answer

Answer: To explore actions that might be better than its current estimates

Exploration versus exploitation.

Lesson: Reinforcement Learning: Q-Learning

Modern AI: Language Models and Agents

20. What do language models fundamentally learn to do?

  1. Store a database of all facts
  2. Predict the next token given previous text
  3. Search game trees
  4. Colour maps
Show answer

Answer: Predict the next token given previous text

Next-token prediction at scale.

Lesson: Language Models: Predicting the Next Word

21. Which combination drove recent AI progress?

  1. Hand-written expert rules only
  2. Large datasets, massive compute and scalable architectures like the transformer
  3. Smaller models and less data
  4. Removing evaluation
Show answer

Answer: Large datasets, massive compute and scalable architectures like the transformer

Scale plus architecture plus tuning.

Lesson: Deep Learning at Scale

22. What do modern AI agents add to a language model?

  1. Removal of all safety limits
  2. A larger font
  3. A faster keyboard
  4. A loop of deciding, calling tools and observing results
Show answer

Answer: A loop of deciding, calling tools and observing results

Perceive, decide, act, with tools.

Lesson: AI Agents With Tools

AI, Ethics and Society

23. Which is a common source of AI bias?

  1. Historical data that reflects past discriminatory decisions
  2. Using too many GPUs
  3. Fast inference
  4. Writing tests
Show answer

Answer: Historical data that reflects past discriminatory decisions

Models learn what the data shows.

Lesson: Bias and Fairness

24. What is the idea behind risk-based AI regulation?

  1. No AI system needs controls
  2. All AI is banned
  3. Higher-risk uses face stronger obligations and controls
  4. Only model size matters
Show answer

Answer: Higher-risk uses face stronger obligations and controls

Controls scale with potential harm.

Lesson: Safety, Privacy and Regulation

25. What should come first in an AI project?

  1. Deploying to all users
  2. Choosing the biggest model
  3. Checking whether AI is the right tool and defining the goal and costly errors
  4. Skipping evaluation
Show answer

Answer: Checking whether AI is the right tool and defining the goal and costly errors

Start with the problem, not the technology.

Lesson: A Responsible AI Checklist

← Back to the course