AI Interview Questions (2026)
Questions of the kind actually asked in AI engineering interviews, grouped by topic. Every section links back to a full explanation of the concept, so a question you can't answer leads somewhere rather than just stopping.
These are deliberately scenario-shaped rather than definitional. "What is an embedding" is a question you can bluff; "you switch to a better embedding model and retrieval quality collapses — what happened" is not, and it's much closer to what an interviewer will actually put to you. Each answer explains the reasoning rather than giving a phrase to memorise.
New to the field, or want a broad revision pass first? AI Interview Questions and Answers covers the definitional ground — what a token, an agent, or RAG actually is — before you get to the scenario questions below.
Preparing specifically for an AI Engineer role? AI Engineer Interview Questions and Answers tests the practical side instead — how you'd actually build and operate an LLM application, not just define its parts.
Two more focused revision pages: LLM Interview Questions stays on how large language models themselves work and behave, and Generative AI Interview Questions covers the wider field — images, audio, and video, not just text.
Deeper on one specific area: RAG Interview Questions and Answers covers the full pipeline end to end, Agentic AI Interview Questions tests how much decision-making control a system actually has, and Prompt Engineering Interview Questions covers writing, debugging, and evaluating prompts. Machine Learning Interview Questions stays on the classical foundations — training, overfitting, evaluation metrics — rather than LLMs or agents.
Asked to design a whole system rather than answer a definition? AI System Design Interview Questions and Answers works through requirements, architecture, trade-offs, and failure modes across three full worked scenarios.
Topics
- RAG — 23 questions covering retrieval-augmented generation and the pieces it's built from: embeddings, chunking, vector databases and reranking. The largest section, because retrieval is the most heavily examined area in this field right now.
- Agents & Subagents — 18 questions on the agent loop, the agentic spectrum, when delegation helps, tool permissions, and where multi-agent systems stop being worth their cost.
- Model Behaviour — 16 questions on the parameters that shape output, why models hallucinate, and when a reasoning model's extra cost actually pays off — including which fixes are actually myths.
- MCP — 9 questions on the host/client/server split, tools vs resources, and when the protocol earns its cost over plain tool calling.
- Safety — 7 questions on prompt injection: why filtering doesn't fix it, the lethal trifecta, and where agent memory reopens the risk.
- Context & Memory — 7 questions on the context window and agent memory: why models seem to forget, why bigger isn't automatically better, and what actually persists across sessions.
- Tools — 4 questions on structured outputs: what schema enforcement actually guarantees, and what it doesn't.
- Evals — 4 questions on AI evals: why benchmark scores don't tell you what you need to know, and where LLM-as-judge actually breaks.
How to use these
Read the question, answer it out loud before looking, then compare. The gap between "I recognise this" and "I can explain it" is the thing an interview exposes, and it only shows up if you commit to an answer first.
Several questions have answers that contradict what's widely repeated — that newer models hallucinate less, that a reranker can rescue a retrieval miss, that lowering temperature fixes wrong answers. Those are the ones worth sitting with, because being confidently wrong is worse in an interview than admitting you don't know.