Most machine learning interview guides were written for a job that no longer exists. They will drill you on bias-variance and random forests, then leave you silent when the interviewer asks how you would stop a retrieval system from hallucinating, or why your fine-tuned model still invents facts.
Interviews for AI and data roles in 2026 test both. You still need the fundamentals — a hiring manager will absolutely ask you to explain regularisation or defend your choice of evaluation metric. But increasingly you also need to hold your own on embeddings, context windows, retrieval pipelines, drift monitoring and the trade-offs of shipping a model that talks back to customers.
So we built a practice set that covers the whole surface: 250 questions across 25 topics, every one with a full written answer — the intuition, the formula, working code where it earns its place, and a note on what the interviewer is actually testing.
→ Open the 250 interview questions (free, no sign-up)
What has changed in AI interviews
Three shifts are worth preparing for specifically.
Generative AI questions are no longer optional. Even for classical data science roles, expect at least one question about how large language models work, when retrieval beats fine-tuning, or how you would evaluate a system whose output has no single correct answer. Candidates who can only discuss supervised learning now look narrow.
Production reasoning is weighted more heavily than modelling. The industry has collectively noticed that the model is a small part of the system. Interviewers ask about data leakage, training-serving skew, drift monitoring and rollback plans, because those are what actually break. If you can explain why your offline AUC was 0.91 and your production performance was not, you are ahead of most candidates.
Judgement beats recall. Nobody is impressed that you can recite the formula for Gini impurity. They want to know when you would choose a tree over a linear model, what you would give up, and how you would tell whether the choice was right. Every answer in our set closes with that framing.
The 25 topics, and what each one is for
Each link below opens the practice set filtered to that topic.
Foundations — the questions that open almost every interview
- ML Foundations — supervised vs unsupervised, the curse of dimensionality, parametric vs non-parametric models
- Probability and Statistics — Bayes theorem, p-values, the Central Limit Theorem, correlation versus causation
- Feature Engineering — encoding, scaling, missing data, and the many faces of data leakage
Core models — the bread and butter
- Linear Regression and Logistic Regression
- Decision Trees and Ensembles: Bagging and Boosting
- Support Vector Machines and Naive Bayes and KNN
- Clustering and Dimensionality Reduction
- Time Series Forecasting — the topic most guides skip, and the one that catches people out
Evaluation and diagnosis — where candidates are separated
- Model Evaluation and Metrics — why accuracy lies, when ROC-AUC misleads, how to pick a threshold
- Overfitting and the Bias-Variance Tradeoff
- Optimization and Regularization
Deep learning and generative AI
- Neural Network Foundations, CNNs and Computer Vision, Sequence Models and Attention
- NLP Fundamentals
- Large Language Models — decoding, hallucination, LoRA, RLHF, prompt injection
- RAG and Vector Search — chunking, hybrid retrieval, reranking, and how to debug a RAG system that confidently lies
Shipping it — the production half
- MLOps and Deployment — versioning, drift, monitoring, CI/CD for models
- Responsible AI and Fairness — fairness metrics, the impossibility result, model cards
- Recommenders and Ranking — cold start, candidate generation, position bias
- Applied ML System Design — the open-ended design questions that close senior interviews
What a good answer looks like
Take a question that comes up constantly: why does accuracy become useless when one class is rare?
A weak answer says “because the data is imbalanced”. A strong answer shows the arithmetic. With 10,000 transactions of which 100 are fraudulent, a model that predicts “not fraud” every single time scores 99% accuracy, catches zero fraud, and has undefined precision. Accuracy is a weighted average of per-class recall, weighted by class frequency — so a rare class contributes almost nothing to it.
Then it moves to what you would do instead: precision and recall on the minority class, PR-AUC rather than ROC-AUC because the enormous true-negative count flatters the latter, and a threshold chosen by expected cost rather than by the default 0.5. That is the shape of every answer in the set — the mechanism, the numbers, and the decision it drives.
A four-week preparation plan
Working through 250 questions cold is a poor use of time. This ordering works better:
Week 1 — establish the base. Foundations, Probability and Statistics, Feature Engineering, and both regression topics. Do not skip the statistics: it is where interviews expose people who learned the library but not the reasoning.
Week 2 — models and evaluation together. Trees, ensembles, SVM, KNN, clustering, then Model Evaluation and Overfitting. Studying evaluation immediately after the models makes both stick, because every model question eventually becomes a metric question.
Week 3 — deep learning and generative AI. Neural networks through to LLMs and RAG. If you are targeting an AI engineer or GenAI role, invert this and start here.
Week 4 — production and rehearsal. MLOps, Responsible AI, Recommenders, System Design. Then use the Random question button and answer out loud, timed. Speaking an answer is a different skill from recognising one, and the gap between them is where interviews are lost.
Mark each question complete as you go — progress is saved in your browser, so you can close the tab and pick up where you left off.
Three mistakes that cost people offers
Reciting instead of deciding. Interviewers rarely want a definition. They want to hear you weigh options. Attach a “when I would use this, and what I would give up” sentence to every answer.
Ignoring the production half. Candidates over-prepare on modelling and under-prepare on everything after it. Data leakage, drift, monitoring and rollback come up constantly, and strong answers there are memorable because so few people give them.
Claiming certainty you do not have. “I am not sure, but here is how I would find out” is a genuinely good answer. Confidently wrong is much worse than honestly uncertain — and the same is true of the models you build, which is a point worth making out loud.
Start here
Pick the topic closest to the role you are interviewing for, answer five questions out loud before reading the answers, then compare. The gap between what you said and what is written is your study plan.
→ Open the 250 interview questions with answers
If you are still deciding which role to target, start with our guide to AI career paths and what each one actually requires.
Frequently asked questions
Are these questions free?
Yes. No sign-up, no email, no paywall. Your progress is stored in your own browser only.
How current are the topics?
The set includes LLMs, RAG and vector search, responsible AI and MLOps alongside the classical material, which reflects what is actually being asked in 2026 rather than what was asked five years ago.
Do I need to know all 250 for an interview?
No. Filter to the topics relevant to the role. A data analyst role will lean on statistics and evaluation; an AI engineer role on LLMs, RAG and system design; a data scientist role on most of the middle.
Is this suitable for freshers?
Yes. Every topic has Easy, Medium and Hard questions, so filter to Easy first and work up. The answers assume no prior familiarity with the specific technique.
How should I use it the day before an interview?
Do not try to read everything. Filter to Hard in your two weakest topics, answer them out loud, and re-read the “what the interviewer is testing” note at the end of each answer.
