How to use artificial intelligence viva questions
An artificial intelligence viva tests whether you can explain a technical decision aloud, not whether you can recite a glossary. A strong answer usually does four things:
- defines the term or problem precisely;
- states the assumptions and relevant data;
- compares reasonable alternatives;
- identifies limitations, risks and ways to evaluate the system.
You should practise answering in a fixed order. Start with a short definition, then give the mechanism, then apply it to the case. Finish by stating what you would measure or check next. This prevents a technically correct answer from becoming a list of disconnected facts.
For example, if asked why a model overfits, do not stop at “it memorises the training data”. Explain that the model has fitted noise or idiosyncrasies in the training sample, so its generalisation error rises on unseen data. Then mention a suitable response, such as stronger regularisation, more data, early stopping or a simpler model. Say how you would check the result: use a validation set or cross-validation, while keeping the test set untouched until final evaluation.
The questions below are written as oral stations rather than as a particular university’s examination. Your own course or examining body may use different timings, topics or marking rules, so check its official guidance. The method remains useful: answer aloud, record yourself, and review whether you made a decision and justified it.
A useful answer structure is define, explain, apply, evaluate:
- Define the concept in one or two sentences.
- Explain the relevant algorithm, assumption or trade-off.
- Apply it to the scenario in the question.
- Evaluate performance, uncertainty, safety, fairness or deployment risk.
An examiner will usually learn more from a clear answer with one relevant limitation than from a long answer containing every related term.
Set up an artificial intelligence viva board
Before practising, gather your own lecture notes, programming exercises, project documentation and reading summaries. Organise them into sections that can generate questions. A useful board separates foundations from applied decisions, because a viva can move from “what is gradient descent?” to “why did you use it here?” within one conversation.
In MySummaries, a board for this topic could be arranged as follows:
The examiner emphasis is deliberately broader than naming algorithms. You need to show that you understand when a method is appropriate, how its performance should be measured and what could go wrong. If you have built a model, be ready to defend each important choice: data collection, preprocessing, feature selection, model family, loss function, validation design, threshold selection and deployment monitoring.
A station list helps you practise the opening question without guessing what comes next. The best stations combine a core concept with a decision or failure case.
A set of stations on artificial intelligence might look like this:
Do not practise only the stations you already know. The untouched station is valuable because it tests whether your method works without a prepared script. After each recording, note the first point at which your answer became vague. That is usually the point where a revision card or a second attempt is needed.
Station 1: explain overfitting and generalisation
This question tests whether you can distinguish memorisation from learning a pattern. It also tests whether you know how to diagnose the problem rather than reaching immediately for a favourite remedy.
Examiner
Your training accuracy is 98% but validation accuracy is 71%. Explain what may be happening and what you would investigate first.
You correctly identified a generalisation gap and described overfitting as fitting noise or training-specific detail.
ImproveState that training and validation performance must be compared using the same metric and relevant preprocessing.
You mentioned model complexity and dataset size, but you did not check for leakage or a mismatched validation distribution.
ImproveName leakage, duplicate records, class distribution and train-validation distribution as early checks.
You proposed regularisation and more data, both reasonable interventions.
ImproveLink the intervention to the diagnosis: simplify the model or increase regularisation only after confirming the gap is genuine.
You recommended a validation set but did not explain how to protect the final test set.
ImproveUse validation or cross-validation for choices, then use the untouched test set once for final estimation.
A strong answerThe most likely explanation is overfitting: the model has fitted noise or details specific to the training sample, so it performs poorly on unseen data. First I would check that the metrics are calculated consistently, that preprocessing was fitted only on the training data, and that there is no duplicate record or target leakage. I would also compare the class distribution and feature distribution between the splits. If the gap is genuine, I could collect more representative data, simplify the model, increase regularisation or use early stopping, then select the intervention using validation data and reserve the test set for final evaluation.
Notice that the model answer does not claim that regularisation is always correct. A low validation score can also result from distribution shift, a weak feature representation, label noise, leakage in the wrong direction or an unsuitable validation split. In a viva, naming the first diagnostic checks often earns more credit than listing five remedies without prioritisation.
Station 2: choose classification metrics
This station tests whether you can match a metric to the cost of errors. Accuracy is not automatically useless, but it can conceal poor performance on a minority class or at an unsuitable decision threshold.
Examiner
A screening model has high accuracy but misses many positive cases. Which metrics and threshold decisions would you examine?
You correctly prioritised false negatives because the scenario is screening for positive cases.
ImproveState that the acceptable balance depends on the consequence of a missed positive and a false alarm.
You named recall, precision, a confusion matrix and precision-recall analysis.
ImproveExplain that the chosen metric should be reported with the operating threshold and confidence or uncertainty where appropriate.
You explained that lowering the threshold may increase recall and reduce precision.
ImproveAdd that the threshold must be selected on validation data and checked on representative subgroups.
You explained why accuracy alone is inadequate, but did not mention calibration or prevalence changes.
ImproveSay whether predicted probabilities are calibrated and how changing prevalence could affect predictive values.
A strong answerI would begin with the confusion matrix and quantify false negatives, false positives, recall and precision at the current threshold. Because this is a screening use case, missed positives may be especially costly, so recall is important, but increasing recall may create more false positives and unnecessary follow-up. I would inspect the precision-recall curve, choose a threshold using validation data and check calibration and performance across relevant subgroups. I would report the operating threshold and expected error trade-off rather than presenting accuracy alone, and I would reassess it if prevalence or the input population changes.
A common weak answer lists accuracy, precision, recall and F1 without saying which one drives the decision. Improve it by using the words false positive, false negative, threshold, prevalence and calibration only when they are relevant to the scenario. Technical vocabulary should carry an explanation.
Station 3: defend a responsible deployment decision
This station tests whether you treat deployment as a sociotechnical decision. A model can have strong average performance and still be unsuitable because of biased data, different error rates, weak documentation, privacy concerns or a lack of human oversight.
Examiner
A model performs differently across demographic groups. How would you investigate and respond before deployment?
You proposed subgroup evaluation and checking label quality, but did not first define the decision and affected groups.
ImproveSpecify the intended use, protected or relevant groups, outcome label and the harm represented by each error.
You mentioned false-positive and false-negative rates, which is a useful start.
ImproveCompare calibration, recall, precision and error rates where appropriate, while recognising that fairness criteria can conflict.
You suggested better data and threshold changes but treated them as purely technical fixes.
ImproveInclude documentation, stakeholder review, human oversight, appeal routes and a decision not to deploy if risk remains unacceptable.
You ended with further testing but did not give a deployment recommendation.
ImproveState the conditions that must be met before a limited, monitored deployment or a decision not to deploy.
A strong answerI would first define the decision the model supports, the outcome label, the affected groups and the harm associated with each type of error. I would then audit data coverage and label quality, and compare relevant performance measures such as recall, precision, calibration and false-positive or false-negative rates across groups. Mitigation might include better representative data, revised labelling, a different threshold policy or a different model, but those choices can involve competing fairness criteria. Before deployment I would require documentation, human oversight, monitoring, an appeal or correction route and a clear stop condition; if the remaining risk is not acceptable for the use case, I would not deploy the model.
The final sentence matters. “Deploy with monitoring” is not a complete answer unless you explain what will be monitored and what action follows a failure. Name the owner of the decision, the relevant data quality checks, the performance measures and the point at which the system is withdrawn or reviewed.
Improve the spoken wording of your answers
After marking, inspect the exact words that cost you marks. Vague verbs such as “check”, “fix” and “use” often hide an incomplete plan. Replace them with a named measurement or action.
I would check whether the model is fair across the groups. Then I would fix the threshold if needed. The model could be biased because of the data. I would not deploy it until the results were reviewed.
Then I would fix the threshold if needed.
This states an intervention without saying which error, metric or decision threshold is being changed, and it does not explain who decides whether the residual risk is acceptable.
Say: I would compare false-positive and false-negative rates, calibration and recall across the relevant groups. If the difference is caused by the operating threshold, I would test an alternative threshold on validation data, document the trade-off and obtain stakeholder approval before any limited deployment.Practise replacing general statements with this pattern:
- “I would measure…” followed by the metric or data check;
- “I would compare…” followed by the relevant groups, splits or baselines;
- “I would change…” followed by the technical intervention;
- “I would stop or review…” followed by a concrete safety condition.
Do not turn every answer into a memorised paragraph. Learn the structure and the decision points. An examiner may alter the scenario, and a flexible answer will survive that change better than a script.
A five-minute debrief after every station
Use the same debrief each time:
- Definition: Did I define the central concept accurately and early?
- Mechanism: Did I explain how the algorithm or failure mode works?
- Application: Did I answer this scenario rather than reciting a textbook section?
- Evaluation: Did I identify a metric, baseline, uncertainty or limitation?
- Decision: Did I state what I would do next and why?
Score each item from zero to two. A low score is more useful than a vague impression that the answer “felt poor”. Convert the missed point into a short card, then retake the station after revising it. If you repeatedly miss the same concept, return to the relevant notes and rebuild the explanation from first principles.
You can also vary the listener. First answer to a technical peer. Then answer as if the listener understands the application but not the mathematics. Finally answer to a non-specialist decision-maker. The underlying facts should stay consistent, while the amount of notation and jargon changes.
How MySummaries helps
MySummaries can turn your own artificial intelligence notes into a revision board, then use that board to create oral stations, record your answer and return examiner-style feedback on technical explanation, trade-offs, evaluation and safety. You can also revise the underlying definitions with flashcards and listen to a spoken explanation before retaking a weak station. Start at portal.mysummaries.app.