Artificial intelligence · Practice test · Updated November 2026
An artificial intelligence practice test checks the core ideas taught in university AI courses: intelligent agents, search, games, constraint satisfaction, logic, planning, probabilistic reasoning and machine learning, plus modern topics such as Transformers, large language models and retrieval-augmented generation. There is no single official AI certification for this material, so the reference point is the standard textbook, Artificial Intelligence: A Modern Approach, and the course syllabi built around it. This guide shows what a good AI practice test covers, who it suits, how to prepare in two to three weeks and the mistakes that cost marks. It includes 10 free artificial intelligence practice questions with answers and explanations.
At a glance
| Topic | Artificial intelligence: classic AI plus machine learning and modern AI |
| Levels covered | Students, job seekers and working developers |
| Core topics | Agents, search, games, CSPs, logic, planning, probability, ML, LLMs |
| Standard textbook | Russell and Norvig, AIMA, 4th edition |
| Official certification | No single one; vendor exams such as AWS AI Practitioner and AI-901 |
| Exametics set | 5 practice tests, 550 questions |
- The AIMA website says the book has been adopted by over 1,500 schools, so its chapter list is a fair map of what “core AI” means.
- Expect calculation and reasoning questions (Bayes’ rule, search costs, game trees), not just definitions.
Source: Artificial Intelligence: A Modern Approach (official book site) and UC Berkeley CS 188 Introduction to Artificial Intelligence. Checked on 9 November 2026.
What a good artificial intelligence practice test covers
A good artificial intelligence practice test follows the classic AI curriculum and then adds machine learning and modern AI. The clearest outline is the table of contents of Artificial Intelligence: A Modern Approach (AIMA, 4th edition, by Stuart Russell and Peter Norvig).
| AIMA part | What practice questions test |
|---|---|
| Problem-solving | Search algorithms, heuristics, adversarial search and games, constraint satisfaction |
| Knowledge, reasoning and planning | Propositional and first-order logic, inference, knowledge representation, automated planning |
| Uncertain knowledge and reasoning | Probability, Bayesian networks, reasoning over time, decisions and utilities |
| Machine learning | Learning from examples, probabilistic models, deep learning, reinforcement learning |
| Communicating, perceiving and acting | Natural language processing, computer vision, robotics |
| Conclusions | Philosophy, ethics and safety of AI |
UC Berkeley’s CS 188 course, which cites AIMA for its readings, follows a similar order: search and A*, constraint satisfaction, game trees, Markov decision processes, reinforcement learning, Bayes nets, hidden Markov models and machine learning.
Agents, search and games
AI courses start with an agent that perceives an environment and acts in it. Know how to describe an environment: fully or partially observable, deterministic or stochastic, episodic or sequential, static or dynamic, discrete or continuous, single-agent or multiagent.
Search questions check that you can compare algorithms. Know when breadth-first, uniform-cost, depth-first and iterative deepening search are complete and optimal, and why A* needs a good heuristic.
- Heuristics: what admissible and consistent mean, and why an overestimating heuristic can make A* return a worse path.
- Adversarial search: minimax, alpha-beta pruning, evaluation functions and expectimax for games with chance.
Constraint satisfaction, logic, knowledge and planning
Constraint satisfaction problems (CSPs) such as map colouring and Sudoku test several ideas at once: variables, domains and constraints, backtracking search, forward checking, arc consistency (AC-3) and the variable and value ordering heuristics.
Logic questions cover propositional and first-order logic: entailment, satisfiability, conjunctive normal form, resolution, and forward and backward chaining.
Planning questions use the classical planning model, where each action has preconditions and effects. Be ready to read a simple action schema and trace a forward search over states.
Uncertainty, probabilistic reasoning and machine learning
A large part of any AI syllabus is reasoning under uncertainty: conditional probability, Bayes’ rule, independence, Bayesian networks, hidden Markov models and decision making with utilities and Markov decision processes.
The machine learning part of AIMA covers learning from examples, probabilistic models, deep learning and reinforcement learning. Expect questions on:
- Supervised learning: decision trees, linear and logistic regression, training, validation and test sets.
- Generalisation: overfitting and underfitting, regularisation, cross-validation and the bias–variance trade-off.
- Neural networks: layers, activation functions, gradient descent and backpropagation.
- Reinforcement learning: rewards, policies, value functions, Q-learning and exploration versus exploitation.
Language, perception and modern AI
Language and perception questions start with natural language processing and computer vision. The key modern paper is “Attention Is All You Need” (Vaswani et al., 2017), which describes the Transformer as a model “based solely on attention mechanisms, dispensing with recurrence and convolutions entirely”, according to the original Transformer paper on arXiv.
Large language models are built on that architecture, and retrieval-augmented generation (RAG) can keep their answers current. The RAG paper by Lewis et al. (2020) combines a pre-trained sequence-to-sequence model with a dense vector index searched by a neural retriever, and shows that the index can be swapped to update the model’s knowledge without retraining.
Responsible AI
Two official responsible AI frameworks come up most often.
- NIST AI Risk Management Framework: released on 26 January 2023 for voluntary use. Its core has four functions, govern, map, measure and manage, and govern is described as cross-cutting, according to the NIST AI RMF Core. NIST’s Generative AI Profile (NIST AI 600-1, July 2024) lists 12 risks; one is confabulation, its term for confidently stated but false output, often called hallucination.
- EU AI Act: Regulation (EU) 2024/1689 entered into force on 1 August 2024 and sorts AI systems into unacceptable, high, transparency and minimal risk levels, according to the European Commission’s AI Act page. The first eight prohibitions have applied since February 2025; a ninth, added by the AI Omnibus, applies from December 2026.
Exametics practice tests
Artificial Intelligence Practice Tests
- 5 practice tests with 110 questions each: 550 questions in total
- Classic AI curriculum plus machine learning, Transformers, LLMs, RAG and responsible AI
- Every answer explained, with a hint, and tagged with one of twelve topic areas
- Timed Exam Mode test; rebuilt in November 2026 with every answer key checked again
- Lifetime access
30-day refund. Instant access after payment.
Who these artificial intelligence practice tests are for
These practice tests suit computer science students, job seekers and developers who want to check how well they know the core ideas of AI. They are for knowledge and interview preparation, not a copy of any vendor exam.
| Learner | What to focus on |
|---|---|
| Computer science students | Search, CSPs, logic, Bayes nets and MDPs: the topics that university AI exams test with worked problems. |
| Job seekers | Explaining trade-offs out loud: A* versus greedy search, overfitting fixes, when to use RAG instead of fine-tuning. |
| Working developers | Machine learning basics, Transformers, LLM limits and responsible AI, to fill gaps left by learning on the job. |
Is there an official artificial intelligence certification?
There is no single, vendor-neutral certification that covers this whole syllabus. Vendor AI exams test AI concepts on their own platforms, so they are a next step after the core ideas.
- AWS Certified AI Practitioner (AIF-C01): a Foundational exam with 65 questions in 90 minutes, costing US$100, taken at a Pearson VUE centre or online, according to the AWS Certified AI Practitioner page. Read our AWS AI Practitioner (AIF-C01) domains and study plan.
- Microsoft Certified: Azure AI Fundamentals (AI-901): Microsoft lists two skill areas, “Identify AI concepts and capabilities” and “Implement AI solutions with Microsoft Foundry”, and says candidates need knowledge of Python coding syntax, according to the Microsoft Learn Azure AI Fundamentals page. See our AI-901 Azure AI Fundamentals guide with free practice questions.
Next on your AI path
- Artificial Intelligence and Machine Learning practice tests: 240 more AI and machine learning questions, each with an explanation and a hint.
- AWS Certified AI Practitioner (AIF-C01) practice tests: turn your AI fundamentals into a recognised entry-level certification.
- Generative AI Interview Prep practice tests: practise the generative AI and LLM questions that AI job interviews ask.
For background reading, see our guides to AI practice exams for job seekers and machine learning certification core concepts.
How to prepare for an AI practice test in 2–3 weeks
With basic programming and probability, two to three weeks is realistic: classic AI first, then uncertainty and learning, then modern AI and timed practice.
- Days 1–4: agents and search Classify five environments. Trace breadth-first, uniform-cost and A* search by hand on a small graph, and check whether your heuristic is admissible.
- Days 5–7: games and CSPs Draw a small minimax tree and apply alpha-beta pruning. Solve a map-colouring CSP with backtracking, forward checking and the minimum-remaining-values heuristic.
- Days 8–10: logic and planning Convert sentences to conjunctive normal form and prove one entailment by resolution. Write one planning action with preconditions and effects.
- Days 11–14: probability and machine learning Work through Bayes’ rule problems until they are routine, then cover Bayesian networks, MDPs, overfitting, regularisation, neural networks and Q-learning.
- Days 15–18: modern AI and timed practice Read the abstracts of the Transformer and RAG papers and the NIST and EU AI Act summaries. Then take the Exametics artificial intelligence practice tests in timed Exam Mode and revisit your weakest topic areas.
Tip
For every algorithm you study, write one line each on whether it is complete, whether it is optimal, and what it costs in time and memory.
Planning more than one exam?
If AI fundamentals are the first step towards a cloud AI certification or several technical interviews, the All-Access Pass covers every practice test on Exametics, including new exams as they are added, with lifetime access. It costs less than buying about six single practice tests.
Free artificial intelligence practice questions (10 with answers)
Try these 10 free artificial intelligence practice questions before opening the answers. They are original AI sample questions, each labelled with its topic area; the full Exametics artificial intelligence practice tests have 550 questions in the same style.
Practice question 1 · Intelligent agents
An automated taxi drives passengers through city traffic. Which description of its task environment is correct?
- A. Partially observable, stochastic, multiagent, sequential, dynamic and continuous.
- B. Fully observable, deterministic, single-agent, episodic, static and discrete.
- C. Partially observable, deterministic, single-agent, sequential, static and discrete.
- D. Fully observable, stochastic, multiagent, episodic, dynamic and continuous.
Show answer
Answer: A. Sensors cannot see everything, outcomes are uncertain and other drivers are agents too. Each decision affects later ones, the world changes while the taxi decides, and speed and steering are continuous.
B describes a simple puzzle, not traffic. C is wrong because traffic is neither deterministic nor static, and there are other agents. D is wrong because the taxi cannot observe the full state and its decisions are not independent episodes.
Practice question 2 · Search
Which condition guarantees that A* tree search returns an optimal (lowest-cost) solution?
- A. Every action has the same step cost of 1.
- B. The frontier is ordered by path cost g(n) only, ignoring the heuristic.
- C. The heuristic is admissible: it never overestimates the true cost from a node to the nearest goal.
- D. The heuristic is always strictly greater than the true remaining cost.
Show answer
Answer: C. A* expands nodes in order of f(n) = g(n) + h(n). If h never overestimates, a goal reached by a worse path cannot be selected before the optimal path is fully explored, so the first goal expanded is optimal. Graph search needs the stronger property of consistency.
A is not enough: with equal step costs, an overestimating heuristic can still lead A* to a worse goal first. B describes uniform-cost search, which is optimal but is not A*. D describes an overestimating heuristic, which is exactly what breaks optimality.
Practice question 3 · Adversarial search
Which statement about alpha-beta pruning in a two-player, zero-sum game tree is true?
- A. It returns an approximate value that can differ from the minimax value, in exchange for speed.
- B. It returns the same root value as minimax, and with perfect move ordering it examines about bm/2 nodes instead of bm.
- C. It only works in games with chance nodes, such as backgammon.
- D. It prunes the same branches whatever order the moves are examined in.
Show answer
Answer: B. Alpha-beta only skips branches that cannot affect the final decision, so the root value is identical to plain minimax. With the best moves examined first, the effective branching factor drops to about the square root of b.
A is wrong because alpha-beta is exact, not approximate. C is wrong because alpha-beta is designed for deterministic minimax trees; chance nodes need expectimax-style handling. D is wrong because pruning depends heavily on move ordering.
Practice question 4 · Constraint satisfaction
During backtracking search on a map-colouring problem, which variable does the minimum-remaining-values (MRV) heuristic choose next?
- A. The variable with the most legal values left, so that it is least likely to fail.
- B. The next variable in alphabetical order.
- C. The value that rules out the fewest choices for neighbouring variables.
- D. The unassigned variable with the fewest legal values left in its domain.
Show answer
Answer: D. MRV picks the most constrained variable. If a variable has very few values left, it is the most likely to fail, so choosing it first detects dead ends early and prunes the search tree.
A is the opposite of MRV. B is a fixed ordering that ignores the current state of the search. C describes the least-constraining-value heuristic, which orders values for a variable, not which variable to assign next.
Practice question 5 · Logic and inference
You want to prove by resolution that a knowledge base KB entails a sentence α. What is the correct procedure?
- A. Add α to KB, convert to conjunctive normal form, and derive the empty clause.
- B. Check that α is true in at least one model of KB.
- C. Add ¬α to KB, convert to conjunctive normal form, and derive the empty clause.
- D. Show that ¬α is satisfiable on its own.
Show answer
Answer: C. KB entails α exactly when KB ∧ ¬α is unsatisfiable. Resolution is a refutation method: you add the negated goal, convert everything to clauses and keep resolving. Deriving the empty clause shows a contradiction, so α must follow from KB.
A would show that KB ∧ α is unsatisfiable, which means KB entails ¬α. B is too weak: entailment needs α to be true in every model of KB, not just one. D says nothing about KB at all.
Practice question 6 · Probabilistic reasoning
A condition affects 1% of a population. A test detects it in 90% of people who have it, and gives a false positive for 10% of people who do not. A randomly chosen person tests positive. Roughly how likely is it that they have the condition?
- A. About 90%
- B. About 8%
- C. About 50%
- D. About 1%
Show answer
Answer: B. By Bayes’ rule, P(condition | positive) = (0.90 × 0.01) / (0.90 × 0.01 + 0.10 × 0.99) = 0.009 / 0.108 ≈ 0.083, or about 8%. False positives from the large healthy group outnumber true positives.
A confuses P(positive | condition) with P(condition | positive), the classic base-rate mistake. C has no basis in the numbers. D is the prior before the test; a positive result does raise the probability.
Practice question 7 · Machine learning
A neural network classifier reaches 99% accuracy on its training set but only 72% on a separate validation set drawn from the same data. What is the most likely problem and a sensible response?
- A. Underfitting; add more layers and train for more epochs.
- B. The learning rate is too low; increase it tenfold.
- C. The validation set is too large; move most of it into the training set and stop measuring validation accuracy.
- D. Overfitting; use regularisation such as weight decay or dropout, early stopping, more training data or a simpler model.
Show answer
Answer: D. A large gap between training and validation accuracy means the model has memorised details that do not generalise. Regularisation, early stopping, more data and a smaller model all reduce this variance.
A is wrong because underfitting shows poor training accuracy too, and a bigger model would make overfitting worse. B is wrong because a model that already fits the training data almost perfectly is not held back by a slow learning rate. C removes the only signal that showed the problem, so you could no longer measure generalisation.
Practice question 8 · Transformers
In the Transformer’s scaled dot-product attention, why are the query–key dot products divided by √dk before the softmax?
- A. For large dk the dot products grow large, pushing the softmax into regions with extremely small gradients; scaling counteracts this.
- B. It adds information about the position of each token in the sequence.
- C. It makes the attention weights sum to 1 across the keys.
- D. It reduces the cost of attention from quadratic to linear in the sequence length.
Show answer
Answer: A. The “Attention Is All You Need” paper suggests that, for large key dimensions, dot products grow in magnitude and push the softmax into regions with extremely small gradients, so the authors scale them by 1/√dk.
B describes positional encodings, which are added to the input embeddings because the model has no recurrence or convolution. C is what the softmax itself does. D is wrong because scaling by a constant does not change the number of query–key pairs, so standard attention stays quadratic in sequence length.
Practice question 9 · Large language models and RAG
A company’s internal policies change every week, and its LLM assistant keeps giving answers based on old versions. Which approach best keeps answers current?
- A. Pre-train a new language model from scratch every week.
- B. Increase the sampling temperature so the model gives more varied answers.
- C. Fine-tune the model once on this week’s policies and leave it unchanged afterwards.
- D. Use retrieval-augmented generation: update a searchable index of the policies and pass the passages retrieved for each question to the model.
Show answer
Answer: D. RAG separates knowledge from the model’s weights. When a policy changes, you update the index; the next question retrieves the new text and the model answers from it. The RAG paper shows that swapping the index updates the model’s knowledge without retraining.
A is far too expensive for a weekly change. B only changes how random the output is; it adds no new knowledge and can increase errors. C is out of date again a week later, because the knowledge stays frozen in the weights.
Practice question 10 · Responsible AI
Under the EU AI Act’s risk-based approach, how does the European Commission classify an AI-based email spam filter?
- A. Unacceptable risk: it is banned in the EU.
- B. Minimal or no risk: the AI Act sets no specific rules for it.
- C. High risk: it needs a risk management system, logging and human oversight before use.
- D. Transparency risk: every filtered email must be labelled as AI-generated content.
Show answer
Answer: B. The European Commission’s AI Act page gives spam filters and AI-enabled video games as examples of minimal or no risk, where the Act sets no specific rules.
A covers practices such as social scoring, which are prohibited. C describes the strict obligations for high-risk systems, which the Commission does not apply to spam filters. D covers disclosure duties, such as telling people they are talking to an AI system and labelling deepfakes; filtering email does not generate content.
How did you score?
8–10 correct: you are ready for a full timed test, so try Exam Mode next. Under 8: go back to the AIMA chapters for the topic areas you missed, then practise them again. The Exametics artificial intelligence practice tests have 550 questions like these, each with a hint and a full explanation.
Common mistakes on artificial intelligence practice tests
Most lost marks come from mixing up similar terms or knowing a definition but not the property it guarantees.
- Confusing admissible and consistent heuristics. Admissible means never overestimating. Consistent is the stronger triangle-inequality condition, and every consistent heuristic is also admissible.
- Mixing up variable and value ordering in CSPs. MRV and the degree heuristic choose which variable to assign; least-constraining-value chooses which value to try.
- Getting entailment backwards. To prove α, you add ¬α and look for a contradiction. Adding α proves nothing about α.
- Ignoring base rates. P(evidence | cause) is not P(cause | evidence). Write out Bayes’ rule with the prior every time.
- Treating LLM output as fact. NIST calls confidently stated false output confabulation. Retrieval, citations and human review reduce it; a higher temperature does not.
Artificial intelligence practice test FAQs
Are there free artificial intelligence practice questions?
Yes. This post includes 10 free artificial intelligence practice questions with answers and explanations, each labelled with its topic area. For a full practice test, the Exametics artificial intelligence practice tests have 550 questions across 5 tests.
What does an artificial intelligence practice test cover?
A good artificial intelligence practice test follows the classic university AI curriculum: agents, search, games, constraint satisfaction, logic, knowledge representation, planning, probabilistic reasoning, language and perception. It should also cover machine learning and modern AI, such as neural networks, Transformers, large language models, retrieval-augmented generation and responsible AI.
Is there an official artificial intelligence certification?
No single, vendor-neutral certification covers the whole AI syllabus. Vendors offer entry-level AI exams instead, such as AWS Certified AI Practitioner and Microsoft’s AI-901 Azure AI Fundamentals, which test AI concepts on their own platforms, according to the AWS and Microsoft Learn certification pages.
How many questions are in the Exametics artificial intelligence practice tests?
There are 5 practice tests with 110 questions each, 550 questions in total, including a timed Exam Mode test. Every question explains why the right answer is right and why the others are wrong, and includes a hint.
Is artificial intelligence hard to learn, and how long should I prepare?
The core ideas are manageable if you know basic programming and probability, but questions often ask you to calculate or trace an algorithm rather than recall a definition. Two to three weeks of steady study is a realistic plan if you have taken an introductory computer science course.
Which should I study first: classic AI or machine learning?
Start with classic AI (agents, search, logic and probability), because machine learning and modern AI build on those ideas. Then move on to machine learning, neural networks, Transformers and large language models.
Official sources: Artificial Intelligence: A Modern Approach, NIST AI Risk Management Framework and the European Commission’s AI Act page (linked above). Checked on 9 November 2026.
Practise artificial intelligence with 550 explained questions
Test yourself across classic AI, machine learning and modern AI in timed Exam Mode, with an explanation and hint for every question.
✓ Every answer explained · ✓ Timed exam mode · ✓ 30-day refund