AI Interview Questions for Freshers

Prepare to explain a small project in depth. An interviewer can learn more from how you checked one assistant than from a memorised list of model names. The questions below are practice prompts grouped by theme, not leaked or promised interview questions. For each one we describe what a strong answer shows and what a weak answer sounds like.

Most fresher interviews for AI-related roles in India mix three things: basic concepts, judgement about when AI is appropriate, and a detailed discussion of whatever project you mention. Prepare the project first. It is the part you control.

Concepts

What is the difference between an LLM and an assistant?

Strong answer: the model generates language; an assistant combines the model with application instructions, context, interfaces and possibly tools. You use your own project to identify each layer. Weak answer: treating the chat product and the model as the same thing, or claiming a model can send an email simply because its response says an email was sent.

What is a hallucination and why does it happen?

Strong answer: the model produces fluent text that is not supported by its input or training, because it predicts likely language rather than checking facts. You give an example from your test set and say how you detected it. Weak answer: "the model lies sometimes", with no example and no detection method.

What are tokens and why do they matter?

Strong answer: tokens are the units the model reads and writes; they set the context limit and drive cost and latency. You mention a time your input was too long or your bill surprised you. Weak answer: a definition with no practical consequence.

What does a system instruction do?

Strong answer: it sets the role, constraints and format before the user's message, and you have evidence of what changed when you edited it. Weak answer: assuming an instruction is always followed, with no mention of testing.

Evaluation and quality

How would you evaluate an AI summary?

Strong answer: start with the purpose and source; check factual support, coverage of important points and preservation of uncertainty; then assess readability. In a fictional ten-point handbook, a summary that keeps eight points but reverses a safety condition is not "mostly correct" in any useful sense. You explain how you weight serious errors. Weak answer: judging quality by how natural the paragraph sounds.

How did you improve a prompt?

Strong answer: a before and after example with the failure, your hypothesis, the changed instruction and the result on the same test set. You say what you held constant and whether you repeated cases. Weak answer: "I added 'be accurate' and it got better", tested on one favourable example.

How do you know your improvement is real and not luck?

Strong answer: you kept the same cases, ran each more than once, looked at which cases moved in each direction and stayed modest about a small set. Weak answer: quoting one number from one run.

Reliability and judgement

What would you do when the source lacks an answer?

Strong answer: show a missing-information case from your project. The application should say it does not have enough information, request clarification or hand off to a person. You explain whether the limitation came from the source, retrieval or interpretation, because each needs a different fix. Weak answer: letting the model fill the gap and calling it helpful.

When should a task stay manual?

Strong answer: consider consequences, ambiguity, volume and how easily the result can be checked. A low-volume task with serious consequences and weak evidence may be better served by human judgement. Weak answer: "everything can be automated eventually".

How would you handle personal information?

Strong answer: use only necessary, authorised information; check the product's data controls and retention; keep private data out of public test sets and repositories; practise with synthetic examples. Weak answer: "I removed the names", implying that this alone makes a record anonymous.

What is retrieval and when would you add it?

Strong answer: find relevant documents first and pass them to the model, so answers rest on a known source; you add it when the task depends on information the model does not have or that changes often. You note that retrieval has its own failures. Weak answer: describing retrieval as a way to make the model "know everything".

Working with others

How would you explain an AI limitation to a non-technical manager?

Strong answer: a concrete example, the business consequence, the check you put in place and what remains a residual risk. Weak answer: jargon, or pretending the risk does not exist.

What would you do if a teammate wanted to ship a prompt you had not tested?

Strong answer: ask for the test set, offer to run it together, agree on what failing looks like before shipping. Weak answer: either refusing outright or going along without evidence.

Which tools have you used, and what did each one do badly?

Strong answer: two or three tools you actually used, with one specific weakness each that you saw yourself. Weak answer: a long list with no weaknesses.

How to talk about your project

Use a fixed shape and practise it until it takes about two minutes. State the problem and who had it. Describe the inputs and what was fictional or real. Say exactly what you did versus what a tutorial or teammate did. Give one result on your test set and one limitation that still exists. Then stop and invite questions.

Expect the follow-ups to go deep: which case failed most often, what the fix cost you, what you would do with more time. If you do not know a detail, say what you would inspect to find out. Clear reasoning and honest boundaries are more credible than invented experience.

Questions to ask the interviewer

• Which tasks does the team currently use AI tools for, and which have they decided to keep manual?
• How does the team check output quality before something reaches a customer or colleague?
• What does a new joiner typically work on in the first three months?
• Who reviews prompts or workflows before they change?
• How does the team handle customer data when using external AI services?

Asking about evaluation and data handling shows judgement. Avoid asking only about salary and work from home in the first conversation; raise those once there is mutual interest.

Mistakes freshers make in these interviews

• Reciting model names and parameter counts instead of describing one tested task.
• Claiming a project was production-grade when it was a notebook, then being caught on a follow-up.
• Saying AI "understands" without being able to explain what it actually does.
• Not having a single example of something that went wrong.

How GullyMentor approaches it

A GullyMentor mentor runs mock interviews on the project you actually built, pushes on the weak answers above and helps you rewrite your two-minute explanation until it is specific. See the AI and Generative AI track at /career-tracks/ai-generative-ai. We do not guarantee jobs, interviews or salaries.

Sources & further reading

• NIST AI Risk Management Framework
• Microsoft: Generative AI for Beginners (free course on GitHub)
• OpenAI: prompt engineering guide
• Google: prompt design strategies (Gemini API docs)
• Hugging Face Learn (free LLM course)

Examples and practice recommendations are educational illustrations. Tool documentation and programme inclusions should be checked for your intended use.