HireVue interview questions, and how they are scored

What kind of questions

HireVue's candidate FAQ groups them into three shapes: situational judgement questions about a hypothetical scenario, scenario-based simulations where you act out a response, and past-behaviour questions about what you have previously done. For a finance or consulting graduate scheme, the third is overwhelmingly the most common — some version of "tell me about a time when".

What the FAQ also says, and what is worth internalising, is that the model considers the whole context of a response rather than certain words here or there. Keyword-stuffing an answer with the firm's values is a strategy for a system that does not exist.

Nobody publishes the real questions. Including us.

This is the part of the page where most competing sites make something up. There is no legitimate source anywhere for any employer's current one-way video interview questions. Lists presented as "the real JP Morgan HireVue questions" are reconstructed from candidate forum posts, which is both unreliable and, for several of those forums, a breach of their terms of use.

What can be sourced is different and, for preparation purposes, nearly as useful: the competency frameworks employers publish about themselves, and the process facts they state in their own words. Three of the eight employers we cover publish actual question text on their own careers sites. The other five publish a competency framework, and the practice prompts on those pages are derived from it and labelled as derived. That distinction is enforced by a build check, not by good intentions.

How a score is produced

HireVue's 2025 Explainability Statement is the most detailed account it publishes. It describes three steps: convert speech to text, analyse the text, then score the answer against how expert human raters scored similar answers to similar competency questions. On what is excluded, it is unusually direct:

The same document names the components. Transcription is done by a third party, Rev.ai. The language model is built on RoBERTa, refined on roughly one and a half million unlabelled asynchronous job interview transcripts, then fine-tuned to predict expert human interview ratings across tens of thousands of rated interviews. In plain terms: it is a text model trained to imitate how trained human interviewers scored the same kind of answer.

Two practical consequences follow from "transcript only", and they are the two most useful things on this page:

• Being nervous on camera is not directly scored. Trembling, blinking, looking away and sitting stiffly are not inputs. If they cost you anything, it is indirectly, by making the sentences worse.
• Being inaudible is fatal in a way that being awkward is not. Everything downstream runs on a transcript. A bad microphone, a mumble or a noisy room degrades the only input the model has.

What changed, and why you will read contradictory things

HireVue's public position on what it assesses has moved, so old articles describe a product that no longer exists. The sequence, as far as it can be established:

• Until 2020, HireVue's assessments included a facial analysis component that evaluated facial muscle movements. SHRM reported in February 2021 that HireVue discontinued it in March 2020, after concluding that visual analysis no longer significantly added value.
• In the same 2021 report, HireVue's then chief data scientist described analysis of speech pauses and tonality as small factors in the model that were being reviewed for bias.
• The 2025 statement says the AI uses no audio characteristics at all, tone of voice included.

Those last two are not compatible. Either the model changed between 2021 and 2025, which is likely and would be an improvement, or one of the two descriptions was loose. HireVue has not published anything reconciling them, so we are not going to pretend to know which. What you can rely on is the current statement, because it is the one HireVue is on the record for now.

Is a human involved at all?

Sometimes, and in more circumstances than people assume. Three are documented:

• If the AI cannot score you, a person looks. HireVue's 2025 statement says AI text scoring needs two things: your consent to AI scoring, and enough detectable content in the response. If someone gives too short an answer, runs out of time, does not speak clearly enough, or has a microphone problem, the response is not scored by AI and is flagged for human review instead.
• You can decline AI scoring. The same document says candidates who opt out of AI evaluation take the same assessment, but recruiters manually review and rate their responses against the same rating scales.
• Shortlisted candidates get read. HireVue's FAQ describes the interview as a first cut, with the top group going on to person-to-person interviews before hiring decisions are made by people.

Two rights you may have, depending on where you are

• Illinois. The Artificial Intelligence Video Interview Act requires an employer using AI analysis of recorded interviews for an Illinois-based position to tell you before the interview that AI may be used, explain how it works and what general types of characteristics it evaluates, and get your consent. It also says an employer may not use AI to evaluate applicants who have not consented, and that on your request they must delete your interviews within 30 days, including backup copies.
• New York City. Local Law 144 requires a bias audit of an automated employment decision tool within the previous year, a public summary of that audit, and notice to candidates. The city's own guidance states that notice must be given 10 business days before the tool is used. Complaints go to the Department of Consumer and Worker Protection.

Neither is legal advice, and neither applies everywhere. They are worth knowing because they are the two places where the disclosure you are owed is written down.