What a HireVue Score Reads, and What No Trick Reaches

What does HireVue's AI actually score?

It scores the words. HireVue's 2025 explainability statement describes three stages: speech-to-text transcription, then a natural-language model, then a score benchmarked against expert human raters' evaluations of answers to similar competency-based questions ¹. The same document says the AI uses no video analysis and no other audio characteristics ¹. What it names as not assessed: facial expressions, body language, background and surroundings, tone of voice.

Every claim in it comes from the company that sells the software.

No audit is attached to the sentence. The document calls itself a living document, so it describes the 2025 configuration and not necessarily whatever ran in an earlier year.

The history is on the record too. In January 2021 HireVue's chief data scientist wrote that its algorithms saw no significant additional predictive power when non-verbal data was added to language data ². On that basis the company said it would use no visual analysis in its pre-hire algorithms going forward, and a same-day press release dates the removal itself to early in 2020 ². No data was published alongside the claim. That scope is HireVue's new assessments, not other vendors' models.

By the vendor's own account, the visual channel went because it was not carrying the prediction.

One caution sits inside "only what is said". The words reach the model through a speech recognizer, which has failure modes of its own. So the transcript rule describes what the scorer looks at, and not whether the system heard you correctly.

The recording is a separate matter from the score.

Facial recognition is what people sometimes mean when they ask, and HireVue's candidate FAQ says the video is never used to identify a candidate ³. Nothing in it says the recording goes unwatched. What an AI interview scores, and what it does not covers the wider question, since no published source describes every vendor at once.

What each popular trick is actually aimed at

The delivery tricks aim at a channel this scorer does not read. Ring lights, practiced eye contact with the lens, a level and confident tone: on HireVue's published account, none of it reaches the model ¹. That account is about the scorer, though, not about the audience. The recording exists, and the employer has it.

Seeing the questions early does not work by stealth either.

HireVue's candidate help center says the real questions cannot be viewed before the interview ⁴. The stated ground is that every candidate gets the same amount of time to review, prepare and respond. The practice questions are generic, they can be redone without limit, and they are never uploaded to anyone ⁴.

The default has one stated exception, and it runs through the employer. HireVue's candidate FAQ says that telling an employer you are on the spectrum is a personal choice, and that a candidate who does may also be able to ask for the interview questions in advance ³. The default is a default, not a bar on an accommodation.

Stuffing an answer with words lifted from the job description does aim at the channel that is read, and nothing published shows it pays there. A list of the right nouns is not an account of anything you did, and the score is benchmarked against how expert raters marked answers to similar competency questions ¹.

You can rewatch a practice answer, and you cannot rewatch a submitted one ⁴.

One more check comes before any of this matters, and it is the invite itself. A statement written about HireVue's models says nothing about what the next vendor's system reads, and plenty of platforms publish far less than this one does. The pages a vendor writes for candidates are worth ten minutes before the link expires. Read them for what the employer switched on, not for reassurance.

What happens after the screen?

A person does, on the vendor's intended design. HireVue's candidate FAQ says its interviews and assessments are meant as first steps that help recruiters make the first cut ³. The top group then moves on to person-to-person interviews, and people make the hiring decisions ³. That is candidate-facing copy about an intended design, not a promise about one employer's configuration.

The question is not whether a script clears one filter, but what it costs on the other side.

Motivated self-presentation does move interview scores, and the honest thing is to say so. In one lab study, 111 students answered the same 20-question structured interview twice, once honestly and once instructed to answer the way an applicant would ⁵. The mean rating rose from 3.20 to 3.62, which the authors read as a clear faking effect ⁵. Students, nothing at stake, human raters in the room: that is a long way from a recorded video round scored by a model.

The part worth carrying is who gained. How much a person's rating rose tracked how well they could work out what each question was assessing ⁵.

Reading the question is the skill, and honest preparation is how it gets built.

The documented downside is not subtle. CNBC reported in March 2025 that a candidate posted a video of himself using a coding-interview copilot to pass an Amazon interview ⁷. Amazon and the other companies that had made him offers then rescinded them ⁷. None of the companies confirmed that on the record, and he says he had already turned Amazon's down. He publicised the use himself, so it is one case rather than a rate.

Scripted answers are a known problem on the hiring side too. The piece on whether the async video round is worth keeping when answers are scripted tells hiring teams to keep the round for scheduling reach. The deciding moves into twenty minutes a human watches. A screen you cleared on a script hands you that conversation anyway.

Write the answers the format was built to reward

Write one real example for each competency the posting names, each carrying a decision you made and something a person could check. In a ten-country study of asynchronous video interviews, 582 people answered eight questions for a mock bank manager role ⁶. Honest tactics such as self-promotion were associated with higher rated performance, deceptive ones such as extensive image creation with lower ⁶.

That was a mock interview with no job at stake, tactic use was self-reported, and human raters did the rating ⁶.

An association in one study is not a rule about any scoring model. It is a reason to spend the preparation on real material rather than invented material.

Checkable means a figure you can source, a date, the name of the person who disagreed with you. It is the detail a generic answer cannot produce, and it survives being asked about in the next round.

Preparing hard is not the thing this question is worried about. Researchers compared coached with uncoached interviewees on a situational panel interview, in a combined sample of 146 public safety job incumbents ⁸. Predictive validity and reliability were higher in the coached group ⁸. The abstract reports no validity coefficients ⁸. The coaching studied was the content kind: focus people on the interview-relevant substance and help them convey it accurately, rather than teach them to move a score by peripheral means.

One setting, one interview format, no random assignment.

It still points the opposite way from the folk theory that preparing is a species of cheating. Use the practice round for the mechanics. It is private, unlimited and submitted to nobody ⁴. That makes it the place to find out that the microphone clips, or that the first answer runs long.

For the evening before, How to prepare for a one-way video interview in an hour is the compressed version.

The answer you can give twice, once into a recorder and once to a person who interrupts halfway, is the one worth writing down.

See how it works >