The short answer: AI interview questions fall into four groups. Fundamentals cover what AI is, how it differs from machine learning and deep learning, and the main types of learning. Generative AI covers how large language models work, tokens, embeddings, hallucinations, retrieval-augmented generation (RAG), agents and prompt injection. Responsible AI covers bias, explainability, trustworthiness and regulation. And, in almost every kind of job now, AI skills questions ask how you use AI tools and check their work. Below are short answers checked against Google’s Machine Learning Crash Course, NIST, OWASP, the OECD and the original papers, with links to our deeper guide for each role.
Looking for interviews that are run by an AI, rather than questions about AI? See our guide to AI interviews.
AI fundamentals
“What is artificial intelligence?”
The OECD’s definition, revised in 2023, describes an AI system as a machine-based system that infers, from the input it receives, how to produce outputs such as predictions, content, recommendations or decisions that can affect physical or virtual environments. AI systems differ in how autonomous they are and whether they keep adapting after deployment. In an interview, follow the definition with an example from the job you’re applying for, such as a model that flags fraudulent transactions or drafts replies to customers.
“What’s the difference between AI, machine learning and deep learning?”
They’re nested. AI is the broad goal of systems that produce intelligent-seeming outputs, including rule-based systems. Machine learning is the subset that learns patterns from data instead of following hand-written rules. Deep learning is machine learning with neural networks of many layers, which powers modern image, speech and language models. Generative AI, in Google’s words, creates new content such as text, images or music by learning patterns from existing data.
“What are supervised, unsupervised and reinforcement learning?”
Google’s Machine Learning Crash Course describes them like this:
- Supervised learning learns from labeled examples to make predictions, usually regression (a number) or classification (a category). Example: predicting whether an email is spam.
- Unsupervised learning finds patterns in unlabeled data, often by clustering similar examples. Example: grouping customers by behavior.
- Reinforcement learning learns from rewards or penalties for actions taken in an environment, and produces a policy for getting the most reward. Example: teaching a robot to walk.
“What is the Turing test, and does it still matter?”
Alan Turing proposed it in his 1950 paper “Computing Machinery and Intelligence” as a way around the question “Can machines think?”. Instead, he asked whether a machine could play an “imitation game” well enough that a questioner couldn’t reliably tell it from a person. A good answer adds its limits: sounding human isn’t the same as being accurate, reliable or safe, which is why AI systems are now judged on task-specific tests and human evaluation.
“What is a neural network?”
Layers of simple connected units. Each unit computes a weighted sum of its inputs and applies a nonlinear function. Training adjusts the weights to reduce a loss function, using backpropagation to compute how each weight affects the error. Our machine learning interview questions explain backpropagation, vanishing gradients and the other deep learning basics.
“What is overfitting?”
Google defines it as a model that matches, or memorizes, its training data so closely that it fails on new data. You spot it when training performance keeps improving while validation performance gets worse. Fixes include more data, simpler models, regularization and early stopping.
“How do you know whether an AI model is good?”
Measure it on data it hasn’t seen, with a metric that matches the cost of its mistakes. For classifiers, that’s often precision and recall rather than accuracy. For generative AI, it means a test set of real tasks scored for correctness, faithfulness to sources and safety, often with human review. Then confirm in production, for example with an A/B test, because offline scores don’t always hold up. Our LLM engineer interview questions cover how to evaluate language models.
Generative AI and LLM questions
“How does a large language model work?”
Text is split into tokens, and the model learns to predict the next tokens in a sequence from huge amounts of text. Google’s course explains that modern LLMs use the Transformer architecture, whose self-attention lets each token weigh how much every other token in the input affects its meaning. After this pretraining, models are often fine-tuned for specific tasks or to follow instructions. The model generates text by repeatedly predicting a likely next token, which is why it can sound fluent and still be wrong.
“What are tokens and the context window?”
Tokens are the units a language model reads and writes: whole words, parts of words or punctuation. The context window is how many tokens the model can take into account at once, including your prompt, any documents you add and its own answer. Long conversations or documents can exceed it, so applications summarize, trim or retrieve only the relevant parts.
“What are embeddings?”
Embeddings are lists of numbers (vectors) that represent items such as words, documents or products. Google’s course describes them as lower-dimensional representations of sparse data, and similar items end up close together in that space. That makes them the basis of semantic search, recommendations and the retrieval step in RAG.
“What is a hallucination, and how do you reduce it?”
NIST’s generative AI profile calls it confabulation: confidently stated but false content that can mislead users. It warns of the harm when people believe and act on that content, for example a made-up detail in a summary of patient information. To reduce it:
- Ground answers in sources with retrieval, and ask for citations.
- Let the model say “I don’t know”, and constrain answers to the provided material.
- Check outputs with code, tools or a second model, and test on questions with known answers.
- Keep a person in the loop for high-stakes decisions.
“What is retrieval-augmented generation (RAG)?”
RAG was introduced in a 2020 paper by Lewis and colleagues. It combines what a model learned in training with documents it retrieves at the moment it answers. In practice, you search a document store for passages relevant to the question and include them in the prompt. The model can then answer from current or private information and cite it. Most of the hard work is in retrieval quality: how documents are split, embedded and ranked. Our LLM engineer interview questions cover RAG in depth.
“When would you use prompting, RAG or fine-tuning?”
Start with prompting: it’s the cheapest and fastest to change. Add RAG when the model needs knowledge it doesn’t have, such as your company’s documents or recent events, or when answers must cite sources. Choose fine-tuning when you need to change behavior, such as a consistent format, tone or narrow task, and prompting isn’t enough. The options combine: a fine-tuned model can still use retrieval.
“What is an AI agent?”
Anthropic’s engineering guide draws a useful line. Workflows orchestrate a language model and tools through predefined code paths. Agents let the model decide its own steps and which tools to use. The same guide recommends the simplest solution that works and adding complexity only when needed, because agentic systems often trade latency and cost for better results. In an interview, explain how you’d limit an agent’s tools, add stopping conditions and require human approval for risky actions. Our AI engineer interview questions go deeper.
“What is prompt injection, and how do you defend against it?”
OWASP ranks it first in its 2025 Top 10 for LLM applications. It happens when input changes a model’s behavior in unintended ways. Direct injection comes from the user’s own prompt. Indirect injection hides instructions in content the model reads, such as a web page or a file. OWASP says there may be no fool-proof prevention, but these measures limit the damage:
- Constrain the model’s role, and validate outputs against an expected format.
- Filter inputs and outputs.
- Give the model the least privilege it needs, and handle sensitive functions in code.
- Require human approval for high-risk actions.
- Separate and label untrusted content.
- Test the system with simulated attacks.
Responsible AI questions
“Where does bias in AI come from?”
NIST’s report on AI bias (SP 1270) identifies three categories: systemic bias from institutions and history, statistical bias from unrepresentative data or flawed methods, and human bias from the people who build and use systems. It stresses that human and systemic factors are significant and often overlooked. A hiring model trained on past hiring decisions, for example, can learn the preferences behind them.
“How would you check a model for unfair outcomes?”
Compare its results across groups: error rates, selection rates and how well each group is represented in the data. Investigate the gaps, then fix the data, the features or the decision threshold, add human review where stakes are high, and document what you tested. In hiring, rules apply. For example, New York City requires bias audits of automated hiring tools. Our guide to AI in hiring covers the rules.
“What makes an AI system trustworthy?”
NIST’s AI Risk Management Framework lists seven characteristics:
- Valid and reliable
- Safe
- Secure and resilient
- Accountable and transparent
- Explainable and interpretable
- Privacy-enhanced
- Fair, with harmful bias managed
It notes that these need balancing for each use. For example, a more interpretable model can be less accurate, and privacy-enhancing techniques can cost accuracy.
“What’s the difference between explainability and interpretability?”
In NIST’s framework, explainability describes how a system works, the mechanisms behind its output. Interpretability describes what an output means in the context of the system’s purpose. Both matter most in high-stakes uses such as lending, health and hiring, where people affected by a decision may need to understand it. Techniques range from choosing simpler models to feature-importance methods and examples.
“How is AI regulated?”
It depends on where you are. The EU AI Act takes a risk-based approach and treats AI used in recruitment as high-risk, with obligations such as risk management and human oversight. In the US, NIST’s AI Risk Management Framework is voluntary, and some states and cities have their own rules, such as New York City’s bias audit law for hiring tools. Check the current rules for your sector before the interview, since they’re changing quickly. Interviewing for a policy role? See our AI policy analyst interview questions.
AI skills questions for any job
Employers now ask these in marketing, finance, operations and support interviews, not just technical ones:
- “How do you use AI in your work?”
- “Tell me about a time an AI tool got something wrong. How did you catch it?”
- “What would you never hand over to AI, and why?”
- “How do you keep up with changes in AI?”
Answer with a real example in four parts: the task, the tool and approach, your judgment (how you checked and corrected the output), and the result, with a number if you can. Our guide to describing your experience with AI tools has example answers by job.
By role
- Machine learning engineers and data scientists: our machine learning and data scientist interview questions.
- AI engineers and LLM engineers: our AI engineer and LLM engineer interview questions.
- MLOps and platform roles: our MLOps engineer interview questions.
- AI training and data annotation work: our guides to AI trainer jobs and data annotation jobs.
- Interviewing at an AI lab: our OpenAI and Anthropic interview guides.
How to prepare
- Explain each fundamental in two sentences, then add an example from your field.
- Build or use one small AI project end to end, such as a RAG app over real documents, and be ready to say what went wrong.
- Prepare one responsible AI story: a risk you spotted, how you measured it and what you changed.
- Write two AI skills answers in the four-part structure, with numbers.
- Read the job posting for the tools and risks the team cares about, and prepare questions about them.
Practice your answers out loud
Frequently asked questions
What are the most common AI interview questions?
What is the difference between AI and machine learning?
What is an AI hallucination?
What AI questions are asked in non-technical interviews?
How do I prepare for AI interview questions as a beginner?
Is an AI interview the same as AI interview questions?
Sources
- OECD.AI, Updates to the OECD’s definition of an AI system explained, 29 November 2023. Read on 11 November 2026.
- Google for Developers, What is machine learning? and Machine Learning Crash Course: Overfitting, Embeddings, Large language models and Transformers. Read on 11 November 2026.
- Stanford Encyclopedia of Philosophy, The Turing Test. Read on 11 November 2026.
- Lewis et al., Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks, 2020.
- Anthropic, Building effective agents, 19 December 2024. Read on 11 November 2026.
- OWASP Gen AI Security Project, LLM01:2025 Prompt Injection. Read on 11 November 2026.
- NIST, Artificial Intelligence Risk Management Framework (AI RMF 1.0), January 2023; Generative AI Profile (NIST AI 600-1), July 2024; and Towards a Standard for Identifying and Managing Bias in Artificial Intelligence (SP 1270), March 2022. Read on 11 November 2026.
