In this article
How this list is ordered and what is sourced
The questions run in the order a candidate meets them: the process and the online assessment, machine learning fundamentals, coding, then behavioral questions. For where an internship loop sits among other hiring stages, read the interview process guide.
Three kinds of claim appear. Facts about Amazon's online assessment and its Leadership Principles come from Amazon's own pages. Definitions of the bias-variance tradeoff and the behavioral method come from papers and university career centers. Everything else is general machine learning knowledge, written without numbers, or a candidate report, labeled as one. Which team asks which question varies, and no public source was found for the processes at Twilio, OpenAI, Anthropic, Nvidia, Microsoft or Meta.
The process and the online assessment
What does an AI intern interview process look like?
It varies by company and team. Amazon documents one step officially: an online assessment (OA), a timed test sent before any live interview. Candidate reports for Applied Scientist intern roles describe a data-structure round and a machine learning depth and breadth round after it, plus Leadership Principles questions (Glassdoor reports, accessed October 2026). Treat those reports as anecdotes.
Why it is asked: this is a planning question, not an interview one. Read the full explanation: the interview process guide.
What is on the online assessment for an applied science intern?
Amazon says the OA consists of a coding assessment and a work styles assessment, and that the coding assessment takes approximately 70 minutes (Amazon Jobs, accessed October 2026). So the common belief that an AI intern screen is only coding is wrong for this role: the second part tests how you work.
Why it is asked: the OA is the first filter, so knowing its two parts decides how you spend preparation time. Read the full explanation: what an OA interview is.
What is the work styles assessment?
A work styles assessment is a set of questions about how you prefer to work, not a technical test. Amazon Web Services (AWS) says such assessments are centered on Amazon's culture and Leadership Principles and typically take 10-20 minutes (AWS Careers, accessed October 2026). One example item asks you to choose between "I like for things to be clearly structured" and "I look forward to the opportunity to learn new things".
Why it is asked: it screens for fit with how the company says it works. Source: AWS Careers, how we hire.
Can you use outside resources during the coding section?
For Amazon's applied science OA, the page says candidates may use publicly accessible online resources, such as a language's standard library documentation, and that browser usage is logged. It warns against sites that need a login or are private repositories (Amazon Jobs, accessed October 2026). Other companies set their own rules, so read the instructions you receive.
Why it is asked: it is a rules question, and misreading it can end an attempt. Source: Amazon Jobs, OA preparation.
Knowing it is half the job. Saying it is the other half.
Interview Clozo is an AI interview copilot for real interview rounds: it hears the question and puts a strong answer on your screen, and you say it in your own words. See the features or how it works.
Start Free Trial →Machine learning fundamentals
What is the bias-variance tradeoff?
The bias-variance tradeoff is the balance between a model that is too simple and one that is too flexible. A model that is too simple will underfit, which is high bias, while one that is too complex will overfit, which is high variance (arXiv 1912.08286, 2019). For squared error, expected test error splits into bias squared plus variance plus noise; classification losses have no such clean split (arXiv 1912.08286). In very wide neural networks the classic U-shape often does not appear: Neal et al. report that variance need not grow with width (arXiv 1810.08591).
Why it is asked: it is a common breadth question. Source: Neal et al., arXiv 1810.08591.
What is overfitting, and how do you detect and reduce it?
Overfitting is when a model fits noise in its training data, so it scores well on that data and worse on new data. You detect it by comparing training performance with performance on a held-out set: a widening gap is the sign. Cross-validation, which repeats that comparison over several splits, also helps detect it and choose between models. Typical remedies are more data, a simpler model, a penalty on large weights and early stopping.
Why it is asked: it checks whether you evaluate a model on data it has not seen. Read the full explanation: the AI engineer question list.
Why do you split data into training, validation and test sets?
Training data fits the model, validation data is used to choose settings such as model size or a penalty strength, and test data is touched once at the end to estimate performance on new data. If you tune on the test set, its score no longer measures anything independent, because your choices were shaped by it.
Why it is asked: leakage between splits is a common real-world mistake. Read the full explanation: the AI engineer question list.
How do you evaluate a classifier, and when is accuracy misleading?
Accuracy is the share of correct predictions, and it misleads when classes are imbalanced: a model that always predicts the common class can look good while finding nothing. Precision asks how many flagged items were right; recall asks how many true items were found. Choose the metric from the cost of each kind of error, and say so.
Why it is asked: it separates metric choice from metric recall. Read the full explanation: the AI engineer question list.
How do logistic regression, decision trees, random forests and gradient boosting differ?
Logistic regression fits a linear boundary and outputs a probability. A decision tree splits the data with a sequence of questions and is easy to read but overfits alone. A random forest averages many trees trained on different samples, lowering variance. Gradient boosting adds small trees one after another, each correcting the errors of the earlier ones.
Why it is asked: it is a common breadth area. Read the full explanation: the AI engineer question list.
What is a large language model, and how does it generate text?
A large language model (LLM) is a neural network trained on large amounts of text to output a probability for the next token, a small piece of text. To generate, it picks a token from those probabilities, appends it to the input and repeats. Say what you can build or have read about it, and be honest about the limits, such as made-up facts.
Why it is asked: AI roles increasingly expect a working picture of these systems. Read the full explanation: the AI engineer question list.
What is an embedding?
An embedding is a list of numbers that stands for a word, a document, an image or an item, learned so that similar things get similar numbers. Closeness between two lists, often measured by a dot product or cosine similarity, then serves as a measure of how related the two things are. Search and recommendation systems rely on this.
Why it is asked: it is the base of retrieval and many language-model systems. Read the full explanation: the AI engineer question list.
What do data science intern interviews ask?
Data science intern interviews usually mix the machine learning basics above with questions on statistics, data handling and a project you can explain. No public source for a fixed list was found, so treat this as general guidance. Be ready to describe how you cleaned a dataset, chose a metric and checked a result on data the model had not seen.
Why it is asked: it tests whether you can work with data end to end. Read the full explanation: the AI engineer question list.
Coding questions
What coding questions do AI interns get?
Candidate reports on Glassdoor describe easy data-structure questions in an early round, but no official list exists. Amazon's OA lets you answer in C, C++, C#, Go, Java, JavaScript, Kotlin, Objective C, Python or R (Amazon Jobs, accessed October 2026). Practice arrays, strings, hash maps, sorting and basic recursion, and say your reasoning aloud.
Why it is asked: it checks that you can turn an idea into working code. Read the full explanation: what an OA interview is.
Which programming language should you use?
Use the one you can write correctly under time pressure. Amazon's OA accepts many languages, including Python and R. Python is common for machine learning work and has built-in data structures that shorten solutions. Pick one language and know its sorting, dictionary and string tools well, rather than switching between several.
Why it is asked: language choice shows fluency more than preference. Read the full explanation: what an OA interview is.
Behavioral questions
What is the STAR method, and how do you use it?
Answer a behavioral question with the situation, task, action, result (STAR) method. The STAR method is a structure for answering by describing the specific situation, task, action and result. Pick one real event, not a generalized description of what you usually do. Use "I", not "we", when describing actions, and state the result without being prompted (University of Iowa Tippie College, accessed October 2026).
Text description of the diagram
- The STAR method describes the specific situation, task, action and result.
- Use I, not we, when describing actions.
Why it is asked: past behavior is used as a guide to future behavior. Read the full explanation: interview skills.
How many stories should you prepare?
Plan a small bank and reuse it. The University of Texas (UT) at Austin advises listing at least 6 to 8 stories from your experiences that match the job description (UT Austin Career Services, accessed October 2026). Lehigh's guide similarly starts with analyzing the position and reflecting on your background (Lehigh Career Center, accessed October 2026). One story can answer several questions.
Why it is asked: it is a preparation question; interviewers ask for a time you handled a challenge. Read the full explanation: interview skills.
What are Amazon's Leadership Principles, and how do they show up?
Amazon's official list has 16 principles, including "Leaders start with the customer and work backwards" and "Leaders are owners" (About Amazon, accessed October 2026). Candidate reports say these principles appear inside interviews, so tie a story to one or two.
Why it is asked: work styles questions and behavioral questions are tied to them. Source: About Amazon, Leadership Principles.
Tell me about a project you are proud of.
Choose a project where your own part is clear. Give the problem and constraint in one or two sentences, then spend most of the time on what you decided and built, what failed, and the result with a measurement if you have one. Name what you would change. For a machine learning project, include how you evaluated it.
Why it is asked: it shows ownership and whether you can explain technical work plainly. Read the full explanation: interview skills.
Why do you want this internship?
Link three things: what the team works on, a specific piece of your own experience that fits it, and what you hope to learn. Read the team's public pages and mention one real detail. A generic answer about loving AI fits every company, so it carries no signal for this one.
Why it is asked: it tests research and motivation. Read the full explanation: interview skills.
Follow-up questions
What are Amazon applied scientist intern interview questions?
Officially, only the online assessment is documented: coding plus work styles. Candidate reports add data-structure questions, machine learning depth and breadth, and Leadership Principles questions, and they vary by team and year. Use the entries above for each of those areas.
What do OpenAI and Anthropic intern interviews ask?
No public source for intern interview questions at either company was found for this page, so no questions are claimed. Searchers also ask whether these labs run internships at all; this page could not verify a public intern programme for either, so check each company's careers page for early-career programmes. For the processes the companies describe for other roles, see the OpenAI interview guide and the Anthropic interview process.
What is the Twilio SWE intern interview process?
No public source for Twilio's intern process was found for this page, so no steps are claimed. Check the company's careers page and use the coding and behavioral entries above as general preparation.
What are Nvidia software intern interview questions?
No public source was found for them either. Prepare the common ground in this list: coding fundamentals, a project you can explain, and behavioral stories. Candidate reports on review sites are anecdotes and differ by team.
What are common behavioral interview questions for a software engineer intern?
Expect prompts such as a time you worked through a disagreement, a time you learned something quickly, or a time something you built failed. Answer each with the STAR structure above and a story from your bank.
Are accounting intern interview questions covered here?
No. They are a different field; this page covers AI and machine learning internships only.
- Amazon's official applied science OA has a coding part of about 70 minutes and a work styles part; other companies' rounds are not officially documented here.
- Know the bias-variance tradeoff: too simple underfits, too complex overfits.
- Evaluate on data the model has not seen, and pick metrics from the cost of errors.
- Prepare a bank of 6 to 8 behavioral stories in the STAR structure and use "I" for the action.
- Treat Glassdoor reports as anecdotes that vary by team.
Sources
- Online assessment prep for Applied Science Interns (Amazon Jobs) — Amazon Jobs, 26 Oct 2026
- Amazon Leadership Principles — About Amazon, 26 Oct 2026
- How we hire (AWS Careers) — Amazon Web Services, 26 Oct 2026
- STAR Method, Career Guide — University of Iowa, Tippie College of Business, 26 Oct 2026
- Behavioral Interview Guide — Lehigh University Career Center, 26 Oct 2026
- Behavioral interviews — University of Texas at Austin Career Services, 26 Oct 2026
- On the Bias-Variance Tradeoff: Textbooks Need an Update (arXiv 1912.08286) — arXiv, 17 Dec 2019
- A Modern Take on the Bias-Variance Tradeoff in Neural Networks (arXiv 1810.08591) — arXiv (Neal et al.), 19 Oct 2018
- Amazon Applied Scientist Intern interview questions (Glassdoor, candidate reports) — Glassdoor, 26 Oct 2026
