AI Interview Questions and Answers for Freshers (BTech, BCA, MCA)

Artificial intelligence (AI) is the field of computer science that builds systems able to do tasks that usually need human intelligence, such as understanding language, recognising images and making decisions. Most AI today learns patterns from data instead of following only hand-written rules. Map apps predicting traffic, fraud checks on digital payments and voice assistants that understand Hindi are everyday examples. For a beginner walkthrough, see what AI is.

AI is the broad goal of making machines act intelligently, machine learning is a subset of AI where systems learn from data, and deep learning is a subset of machine learning that uses neural networks with many layers. A rule-based chess program is AI without ML. A spam filter trained on labelled emails is ML. Face unlock on a phone usually runs on deep learning, which needs more data and compute. See AI vs ML vs deep learning.

The three main types are supervised, unsupervised and reinforcement learning. Supervised learning trains on labelled examples, like emails marked spam or not spam. Unsupervised learning finds structure in unlabelled data, like grouping shoppers by buying behaviour. Reinforcement learning trains an agent through rewards and penalties, like a game-playing bot. Strong answers also mention self-supervised learning: LLMs are pretrained by predicting the next token in raw text, so the labels come from the data itself.

Classification predicts a category, while regression predicts a continuous number. Predicting whether a loan applicant will default (yes or no) is classification; predicting the price of a flat in Pune is regression. The difference changes your metrics: classification uses accuracy, precision, recall and F1, while regression uses errors such as MAE or RMSE. Logistic regression, despite its name, is a classification algorithm.

We split data to measure how a model performs on data it has never seen. The training set fits the model, the validation set helps tune settings such as tree depth, and the test set gives one final, honest score at the end. Splits like 80/10/10 are common starting points, though the right split depends on data size. If you keep tuning against the test set, your reported score becomes too optimistic.

Overfitting is when a model learns the training data too closely, including its noise, so it scores well in training but poorly on new data. Underfitting is when a model is too simple to capture the real pattern, so it performs poorly everywhere. You spot overfitting when training accuracy is high but validation accuracy is much lower. Fixes include more data, a simpler model, regularisation, dropout and early stopping. For underfitting, add better features or a more expressive model.

A neural network is a model made of layers of connected units, called neurons, that turn inputs into outputs using weighted sums and activation functions. Hidden layers learn intermediate patterns and the output layer gives the prediction. During training, backpropagation works out how much each weight contributed to the error, and gradient descent adjusts the weights to reduce it. Networks with many hidden layers are called deep networks. See what a neural network is.

Narrow AI, sometimes called weak AI, is built for specific tasks such as recommending videos, translating text or spotting defects on a factory line. Artificial general intelligence (AGI) would match humans across almost any intellectual task. AGI does not exist today, and researchers disagree about how far away it is. Chat assistants can feel general in conversation, but most experts still place them short of AGI. Keep this answer balanced; interviewers are wary of hype.

The bias-variance trade-off describes two sources of error that pull in opposite directions. Bias is error from overly simple assumptions, which causes underfitting. Variance is error from being too sensitive to the training data, which causes overfitting. A more complex model usually lowers bias but raises variance, so you look for the balance with the lowest error on unseen data. A random forest, for example, reduces the high variance of one deep tree by averaging many trees.

Choose an algorithm based on the problem type, the data size and how explainable the result must be. The one-line versions interviewers expect:

• Linear regression: predicts a number from a mostly linear relationship.
• Logistic regression: predicts the probability of a class; a strong baseline.
• Decision tree: splits data with if-else rules; easy to explain.
• Random forest: averages many trees; accurate with little tuning.
• Gradient boosting (XGBoost, LightGBM): adds trees that correct earlier errors; often best on tabular data.
• K-nearest neighbours: labels a point by its closest examples.
• SVM: finds the widest margin between classes.
• K-means: groups unlabelled data into k clusters.

A confusion matrix is a table that compares a classifier's predictions with the true labels, showing true positives, false positives, true negatives and false negatives. It tells you not just how often the model is wrong, but how. For disease screening, a false negative (missing a sick patient) is usually far more costly than a false positive (one extra test for a healthy person). Precision, recall and F1 are all calculated from these four counts.

Precision is the share of predicted positives that are truly positive, TP / (TP + FP); recall is the share of real positives the model catches, TP / (TP + FN). Favour precision when false alarms are costly, like a spam filter hiding an important email. Favour recall when missing a case is dangerous, like fraud or cancer screening. The F1 score is their harmonic mean. Raising the decision threshold usually increases precision and lowers recall.

Cross-validation estimates how well a model will generalise by training and testing it on several different splits of the data. In k-fold cross-validation, you divide data into k parts (often 5 or 10), train on k minus 1 parts, test on the remaining part, repeat k times and average the scores. It is more reliable than a single split, especially on small datasets. For time-series data, use time-ordered splits so the model never trains on the future.

Gradient descent is an optimisation method that adjusts model parameters step by step in the direction that reduces the loss. At each step it calculates the gradient, or slope, of the loss for each parameter and moves a small amount the opposite way. The learning rate sets the step size: too large and training can diverge, too small and it crawls. Neural networks are usually trained with mini-batch gradient descent, using a small sample of data per step.

An activation function adds non-linearity to a neural network so it can learn complex patterns, not just straight-line relationships. Without it, stacking many layers would still collapse into one linear function. ReLU is the common default in hidden layers because it is simple and trains well. Sigmoid squashes output to between 0 and 1 for binary classification, and softmax turns outputs into probabilities across several classes.

Feature engineering is creating and transforming input variables so a model can learn patterns more easily, for example extracting day of week from a timestamp or counting failed logins in the last hour for fraud detection. Feature scaling, through normalisation or standardisation, puts numeric features on similar ranges. It matters for distance-based models like KNN and for gradient-based training, but tree-based models rarely need it. Good features often beat a fancier algorithm.

Generative AI creates new content such as text, images, code or audio, while traditional machine learning mostly predicts a label or a number. A classic model might classify a customer review as positive or negative; a generative model can write a polite reply to it. Both learn from data, but generative models learn the patterns well enough to produce new samples. See what generative AI is.

A large language model is a neural network, usually based on the transformer architecture, trained on very large amounts of text to predict the next token. That one objective, at huge scale, lets it summarise, translate, answer questions and write code. GPT, Gemini, Claude and Llama are well-known model families. Also know tokens (the chunks of text a model reads) and the context window (how much it can read at once). See what an LLM is.

Prompt engineering is writing and testing instructions so an AI model gives useful, accurate and consistently formatted output. A good prompt states the role, the task, the context, the output format and any limits. "Summarise this placement brochure in 5 bullet points for a second-year student" works far better than "summarise this". The trade-off: prompts fix a lot, but not missing knowledge. Read more about prompt engineering.

AI hallucination is when a model gives a confident answer that is false or made up, such as a fake reference or a wrong date. It happens because language models generate likely-sounding text, not verified facts. You reduce it by grounding answers in trusted documents, telling the model to say when it does not know, using a lower temperature for factual tasks and checking important outputs. See AI hallucination.

RAG, or retrieval-augmented generation, first searches your own documents for relevant passages and then gives them to the LLM, so it answers from that context. Think of it as an open-book exam for the model. It keeps answers current and specific without retraining, and it can show sources. A good fresher project is a RAG chatbot over your college handbook. Learn the basics in what RAG is.

Bias in AI is when a model produces systematically unfair results for certain groups, usually because the training data is unrepresentative or reflects past discrimination. A resume screening model trained mostly on past hires from a few colleges may unfairly score candidates from other colleges lower. Reduce it by checking data coverage, measuring performance separately for each group, handling sensitive features and their proxies carefully, and keeping humans involved in high-impact decisions.

Explainable AI means being able to show why a model made a particular prediction, in terms people can understand. It matters most in lending, healthcare, insurance and hiring, where affected people deserve a reason and regulators may ask for one. Logistic regression and small decision trees are explainable by design. For complex models, methods such as SHAP and LIME estimate each feature's contribution. The trade-off: the most accurate model is not always the easiest to explain.

Answer in about 90 seconds using five parts: problem, data, approach, result and lesson. A sample shape: "I predicted student dropouts using anonymised college records. I compared logistic regression with a random forest and chose the forest for its higher recall on the dropout class. The hardest part was class imbalance, which I handled with class weights." Quote your real metric on a held-out test set, and expect follow-ups on every number and choice you mention.

Give a specific reason linked to something you actually did, not a general line about AI being the future. For example, a college project on crop disease detection might have shown you how a model can help farmers, and you enjoyed the loop of cleaning data, training and testing. Then connect it to the role: what you want to learn in your first year and why this team fits. Interviewers are checking genuine motivation.

Not necessarily, because accuracy is misleading on imbalanced data. If only 1% of transactions are fraud, a model that always predicts "not fraud" scores 99% accuracy and catches zero fraud. I would check the confusion matrix, recall and precision on the fraud class, and the precision-recall curve. Then I would handle imbalance with class weights, resampling or threshold tuning, and pick the threshold based on the cost of missed fraud versus false alarms.

The usual suspects are data drift, data leakage or a mismatch between training and production pipelines. Drift means real-world inputs changed, for example shopping behaviour shifting during a festival sale. Leakage means a training feature secretly contained information about the answer that is unavailable at prediction time. A pipeline mismatch means preprocessing differs between training and serving. I would compare input distributions, re-check features for leakage and set up monitoring.

Start simple and squeeze value from every example. I would begin with a strong baseline such as logistic regression, use cross-validation instead of one split, and invest in feature engineering. For text or images, transfer learning from a pretrained model usually beats training from scratch. Data augmentation, active learning (labelling the most uncertain examples next) or a zero-shot LLM baseline can also help. Report results with their variation across folds, not one lucky score.

I would first clarify scope: which questions, which scripts (Hindi in Devanagari and Hinglish in Roman script), and when to hand over to a human. Then I would build a RAG assistant over the official admission documents, using a model and embeddings tested on Hindi, and evaluate it with real student questions in both languages. I would add a fallback reply for anything not in the documents and measure accuracy on a reviewed test set before launch.