20 AI Interview Questions for Freshers with Example Answers

How to practise these AI interview questions

Use each answer as a starting point, then explain it in your own words and connect it to a small example. These are practice questions, not a claim about any employer’s interview process. A clear definition, a practical example and a validation step make a stronger answer than memorised terminology.

1. What is the difference between AI and machine learning?

AI is the broader field of systems performing tasks associated with intelligence. Machine learning is an approach within AI that learns patterns from data. A rule-based assistant can be an AI system without learning a model from examples.

2. How are classification and regression different?

Classification predicts a category, such as a support-ticket type. Regression predicts a numeric value, such as delivery time. Choose the target and evaluation method from the problem rather than the algorithm you want to try.

3. What are features and labels?

Features are the inputs used by a model. A label or target is the outcome it is trained to predict in supervised learning. Check that every feature would be available when the prediction is actually made.

4. Why split data into training, validation and test sets?

Training data is used to fit a model. Validation data or cross-validation supports model choices. A separate test set estimates performance after those choices. Repeatedly tuning against the test result makes that estimate less independent.

5. What is overfitting?

A model overfits when it captures training-specific details that do not generalise well. A gap between strong training performance and weaker validation performance can be a warning. Consider simpler models, regularisation, better data or an appropriate early-stopping setup.

6. What is data leakage?

Leakage occurs when information unavailable at prediction time influences training or evaluation. One example is fitting a scaler or imputer on the entire dataset before splitting. Another is using a column recorded only after the outcome.

7. Why can accuracy be misleading?

If 95 of 100 examples belong to one class, always predicting that class gives 95% accuracy while missing every example of the other class. This is an illustrative example. Inspect the confusion matrix and choose metrics that reflect the cost of mistakes.

8. What do precision and recall measure?

Precision asks how many predicted positives are actually positive. Recall asks how many actual positives were found. For a support classifier, decide which missed cases and false alarms matter before choosing a threshold.

9. What is a baseline model?

A baseline is a simple comparison, such as a majority-class prediction or a straightforward rule. Evaluate it using the same split and metric as the more complex approach. Improvement should be measured rather than assumed.

10. How would you handle missing values?

First check what is missing and why. The approach may be a documented fill value, imputation, a missingness indicator or excluding unsuitable rows. Learn imputation parameters only from training data and apply the same transformation consistently.

11. When would you use a Python list or dictionary?

A list stores an ordered sequence, such as messages to process. A dictionary maps keys to values, such as counts by message category. Explain your choice through the operations you need, including lookup, iteration and updates.

12. What should you do when a model API call fails?

Distinguish invalid input, authentication, rate limits and temporary service errors. Show a useful error, avoid exposing credentials and apply bounded retries only where appropriate. Do not silently present a placeholder as a real model response.

13. What is a large language model?

An LLM is a language model trained on large amounts of data to model token sequences and generate text. It can produce useful responses, but fluent wording does not guarantee factual accuracy or correct reasoning.

14. What makes a useful prompt?

State the task, relevant context, constraints and expected output. Include examples when they clarify the requirement. Compare changes on a fixed set of inputs, including difficult cases, rather than choosing a prompt from one appealing answer.

15. What is an embedding?

An embedding represents an item such as text as a numeric vector. Similarity between vectors can support retrieval or grouping. Similarity does not prove that a retrieved passage answers the question or is factually correct.

16. How does retrieval-augmented generation work?

A RAG workflow retrieves relevant material and supplies it as context for generating an answer. Evaluate retrieval quality and answer support separately. If the source does not contain the answer, the application should handle that absence explicitly.

17. How is RAG different from fine-tuning?

RAG adds retrieved context at the time a question is answered. Fine-tuning changes model parameters using training examples. The right choice depends on whether the problem concerns access to information, consistent behaviour or another measurable requirement.

18. How would you evaluate an AI assistant?

Use representative questions with expected facts or source passages. Include unanswerable questions, ambiguous inputs and format requirements. Track wrong facts, unsupported claims, missing details and inappropriate actions separately.

19. What makes an AI agent different from a chatbot?

An agent can select steps and use tools to pursue a task, while a basic chatbot may only return text. Tool access creates additional requirements: permissions, validation, action limits, logs and approval for consequential changes.

20. How should you explain a project in an interview?

State the problem, data, baseline, approach and measured result. Describe one failure and what you changed after investigating it. Distinguish your own contribution from a tutorial and say clearly what the project does not demonstrate.

Scenario: an assistant gives an unsupported answer

Start by saving the input, retrieved sources and output. Check whether the right passage was retrieved. If it was, investigate whether the answer used it correctly. If the answer is absent, test the application’s missing-information response. Re-run the same cases after a change so you can identify what improved.

Scenario: test performance is unexpectedly high

Investigate leakage before celebrating the score. Look for duplicate records across splits, a target-derived feature, preprocessing fitted on all rows or a time split that uses future information. Explain what you would inspect and which corrected experiment would test the concern.

A final revision checklist

• Explain your dataset, target and baseline without opening the notebook.
• Describe one metric and why it fits the task.
• Show one normal result and one failure case.
• Reproduce the project using your README.
• State a limitation and a measurable next improvement.

Review the scikit-learn guidance on leakage and preprocessing and Google’s explanation of overfitting for the model-evaluation answers.

Use the AI portfolio guide to prepare your project evidence and the learning roadmap to revisit weak foundations.