Interview questions AI answers badly, with examples

The interview questions AI answers badly are those that depend on something the assistant was never given: the candidate's own history, or what has been said in the last ten minutes. This guide describes five kinds of question, with examples for roles outside engineering, and gives the reason each one works. It is advice based on how the tools operate. A weak answer to one of these questions is a reason to ask another, and it proves nothing alone.

Why standard questions are easy for an assistant

An AI interview assistant turns the interviewer's speech into text, reads the shared screen and sends both to a language model. The steps are in how AI interview assistants work. A question that many employers ask has many published answers, and the model returns a good one.

interviewing.io tested this on coding questions. It reported that cheating was "quite easy" when companies asked LeetCode questions word for word or with a small twist, and that only fully custom questions were "largely impervious to cheating." In its 2025 survey, 58% of interviewers at the largest technology companies said they had changed the kinds of algorithmic question they ask.

An assistant has two limits. It knows only what it hears and reads, and it needs time. Independent testing of one such app, reported in Wikipedia's article on Cluely, found delays of 5 to 90 seconds and suggestions that were often generic. Each kind of question below uses one of those limits.

Questions about the candidate's own work

The first kind asks for details of work the candidate did. The model has no record of what happened on the candidate's projects, so the answer has to come from memory.

• Finance: "You say you shortened the month-end close. Which step took longest before the change, and who objected to changing it?"
• Marketing: "Which campaign from last year would you not run again, and what in the numbers told you so?"
• Operations: "Tell me about a supplier you stopped using. What was the last thing that went wrong?"

Follow each answer twice. "Who else was in that meeting?" and "What did it cost?" are enough. A model can invent a project story, and an invented story has to stay consistent through the second and third follow-up. An FBI notice from October 2023 on North Korean IT workers in remote jobs lists "incorrect yet plausible-sounding answers" among the signs of cheating in interviews, beside excessive pausing and eye movements that suggest reading.

Questions that depend on the conversation

The second kind changes one condition after the candidate has answered.

• Customer success: "The client has now told you they are also talking to a competitor. What changes in your plan for the renewal call?"
• Data analysis: "Suppose a third of the rows have no date. What does that do to the method you just described?"
• Design: "Legal says the form must collect a date of birth. What do you move?"

A candidate who built the first answer can adjust it while you are still talking. An assistant has to hear the new condition and write a new answer, so the pause comes back each time you change something.

The third kind refers to something the candidate said earlier.

• "Earlier you said you would start with the largest accounts. Does that still hold, given what you just told me about churn?"
• "You gave two reasons for choosing that vendor. Which one would you drop if you had to?"

These questions cannot be answered without the earlier answer. A candidate who read that answer from a screen may not remember what it said.

Short questions and plain explanations

The fourth kind expects one sentence.

• "Which of those two would you do first?"
• "Would you ship it this week, yes or no?"

A delay of several seconds before a one-sentence answer is easy to see. Ask "why?" straight after, and the candidate has to give a reason for a choice made a moment ago.

The fifth kind asks the candidate to explain their own answer to a newcomer.

• "Explain that to a new hire in their first week, without the jargon."
• "Give me an example of that from a shop or a kitchen."

A person who understands the answer can restate it more simply. A person who read it out has to wait for a second version and read that one too.

How to use these questions fairly

Tell candidates at the start that you will ask follow-up questions and that a pause to think is fine. People know their own work, so these questions are often easier than trivia.

Ask every candidate for the role the same kinds of question. You are then comparing answers to like questions, and no candidate is singled out.

Judge the pattern across several questions. Nervous candidates pause, and people forget the details of a project from five years ago. The other things to look for are in signs a candidate is using AI.

Write down the question and what you noticed, and discuss it with the panel before deciding. When you are unsure, a short second interview with a different interviewer is fairer than a rejection. The steps before and after the interview are in how to stop AI use in interviews.

Try them in your next interview

These questions need no preparation beyond reading the résumé and listening to the first answer. SecureInterview is a video interview in the browser with a shared screen and chat, free for up to 3 people and 30 minutes. Start an interview with your work email.