Who Wrote Your AI Interview Questions, and What They Test

Who Picks the Questions on a Recorded Interview?

Someone at the employer does. HireVue's candidate help center says everyone applying for a given position gets identical questions, selected by the employer, with the same time to review and prepare each one ¹. The platform also applies security measures to stop candidates seeing those questions before the interview ¹. That is one vendor describing its own product, and for that product it settles who owns the list.

The settings around each question belong to the employer too. Spark Hire's help article for employers shows the allowed takes being chosen per question while the question set is built, from a menu of one to five attempts or unlimited ². VidCruiter says recruiters set thinking time, recording time, number of attempts and review timers per question ³.

The box you are answering into was configured, one question at a time.

Where the employer got the wording is a different question, and none of that help text answers it. The employer-side guidance on this site treats a question set as an owned internal asset: a shared question bank with a row per question, and questions built from the role's own weekly tasks instead of a borrowed list. Both pieces exist because the borrowed list is the easy default.

Picked from somewhere else and picked deliberately are not opposites.

Why Does a Generic Question Still Have a Specific Answer?

Because the wording carries less than the thing sitting behind it. HireVue's 2025 explainability statement describes its score as benchmarked against expert human rater evaluations of answers to similar competency-based questions ⁴. So two questions worded differently can be aimed at the same competency. That is the vendor's own account of its own scoring rather than an audit, and it covers HireVue's AI-scored interviews only. What an AI interview scores elsewhere is a wider question.

Behavioural prompts are shared property. "Tell me about a time you handled a disagreement" belongs to nobody, and two employers can run the same sentence against two different definitions of a strong answer. The employer-side guidance on this site puts that definition in its own field on the row, next to the question.

The wording travels between companies, and the standard a good answer has to meet stays home.

On at least one platform, somebody is still supposed to read the answer after the model has. VidCruiter's page says that where AI-assisted scoring is used, reviewers receive AI-generated scores as a starting point and then apply their own judgment with built-in scorecards, with humans described as in control of every hiring decision ³. That page is undated, and it describes intended use rather than showing how much weight any particular reviewer gives that starting score.

What Changes When the AI Interviewer Is Live?

The list can stop being fixed. Alex, the AI recruiter formerly called Apriora, publishes candidate testimonials describing the AI asking typical role questions, summarising the answers back, and asking follow-ups based on what the candidate just said ⁵. Mercor's engineering blog says its interviewer, Monty, is personalised before the conversation starts by processing the candidate's resume ⁶. Both accounts come from the vendors, about their own products.

Hold each of them at arm's length. Alex's page is marketing assembled from selected testimonials, and the mechanics it describes may vary with how a given employer configures the product ⁵. Mercor's volumes are self-reported: around 10,000 conversations a day, fifteen minutes each, on Mercor's own marketplace, which vets contractors for AI-training and expert-data work rather than filling an employer's own roles ⁶.

In that format a question can come from your resume or your last answer, not only from a list.

The two formats also fail differently for you. On a recorded platform, misreading a question costs a take, and where only one take was allowed it costs the answer. In a live conversation there is no take to burn, and the recovery is a sentence: say what you took the question to mean, then answer that. Knowing which of the two is behind the link decides which of those habits is worth rehearsing.

A rehearsed block is built for a fixed list.

A follow-up aimed at the detail you skipped asks for what the block left out, which is the practical argument for knowing your own examples down to their numbers, their dates and the disagreements inside them.

Read the Posting Back Into the Question

Start from the posting, because it is the one public statement you already hold of what this employer decided the role needs. Take the two or three duties it names most insistently, and assume each question is aimed at one of them. Then answer with an episode from your own work that sits inside that duty, carrying the detail only somebody who was there could supply.

Episodes also have some evidence behind them. In a 2007 experiment, past-behaviour questions ("tell me about a time when") were more resistant to faking than situational ones ("what would you do if") ⁷. That study used undergraduates in an in-person format, reports no effect sizes in its abstract, and notes that most candidates fake to some degree ⁷.

Whether the finding transfers to a recorded video answer the paper does not test.

The posting, your own episodes and the recruiter each carry a job before you open the link:

• read the duties in the posting and name the two the questions are most likely aimed at
• write one real episode per duty, with the numbers and the outcome you can defend under a follow-up
• ask the recruiter which competencies the round covers, in writing, in the reply to the invitation

Asking is not presumptuous. An interviewer working off an AI-drafted script usually means a structured interview is in progress, and a recruiter who can name what the round measures is telling you the set has something behind it. One who cannot may not have been told either.

You do not get to choose the questions, only which piece of your own work each one gets.

See how it works >