Where the Scorer Reads Your Words, the Format Takes Your Interviewer

What Do the Vendors Say Their Scorers Read?

Two of them publish the answer, and both say the transcript. HireVue's 2025 explainability statement says its AI relies only on what the candidate says, using no video analysis and no other audio characteristics: not facial expressions, not body language, not the background, not tone of voice ¹. Vervoe's candidate help page says its AI reviews only the transcript of an audio or video answer ².

HireVue and Vervoe are describing their own products here, and neither account is an audit. Each covers one system in one configuration. HireVue's document calls itself a living document that gets updated over time, so it describes the 2025 build, not necessarily what a candidate met in an earlier year ¹. Vervoe's is help-centre text about Vervoe's own scorer ². Nothing in either sentence tells you what a third platform reads, and nothing in either one is a finding by somebody outside the company.

It is not a promise that nobody watches.

Vervoe is explicit about that: the video and audio are transcribed and provided to the employer along with the recording ². A person at the company can open it and form whatever impression a person forms from a video.

The words-only scorer sits upstream of that, and does not undo it.

For the rest of what these systems take in and leave out, what an AI interview scores and what it does not covers what is published and how to ask about your own interview.

Where the Format Actually Costs a Quiet Candidate

In the missing interviewer. A live one hears a two-sentence answer and asks a follow-up, and the follow-up can be where a quiet candidate finally says the useful thing. A one-way recording has nobody to notice. The timer runs, the answer stops where the nerves stopped it, and nothing goes back for more.

The belief that a machine is listening also changes what people say, and that part has been measured. In an online mock interview, 124 participants recorded spoken answers after being told either that an algorithm or that a human rater would evaluate them. The algorithm group reported less deceptive impression management, felt they had fewer opportunities to perform, and gave shorter answers ³.

Those were mostly German students with no job on the line, and the impression-management part was self-reported ³. Read it as a description of a pull, not a measurement of your interview.

On a scorer that reads what you said, a short answer is less evidence.

A second line of work has people reporting that they trim one thing in particular. Across six studies in Scientific Reports, participants believed AI is worse than humans at assessing interpersonal skills, with no equivalent belief about analytical ones. In one 201-person study they rated it 3.28 against 5.24 for humans on a seven-point scale ⁴.

The doubt was specific to interpersonal skill.

In a later study in the same paper, 107 Korean participants who had been through both AI and human selection said they had leaned away from interpersonal skills and toward analytical ones when the assessor was an AI ⁴. That is recalled behaviour rather than observed behaviour, and AI video interviewing is unusually established in Korea's market.

The adaptation is understandable, and it deletes evidence.

Being given information about advanced AI selection technology softened the negative belief in the same paper ⁴. On the hiring side, what follow-up questions expose whether someone understands the answer they gave describes the questions this format removes, and reading it shows what a good interviewer would have asked you next.

What the Research Found When Models Did Read Delivery

Extraversion was the one trait where delivery carried the most weight, and the finding argues against confidence coaching rather than for it. In an academic study of automated video personality models, the models put the most weight on the words for four of the five Big Five traits. Extraversion was the exception: on average 38% of the weight went to voice cues, 34% to words and 28% to face and body ⁵.

The same paper shows why that number is not a tip. The share given to words swung from 17.5% in one sample to 60.0% in another ⁵. Weights that move that far between samples are describing the training set, not a stable property of people.

Four cues did predict extraversion in every sample: word count, average loudness, loudness peaks per second and variability in the smile muscle at the corner of the mouth ⁵.

Turning that into a plan to speak louder would be wrong twice over. These models were built to reproduce how raters perceived students, not hiring outcomes, and no vendor's system is described in them ⁵.

The paper adds that vendor accounts of that split have been inconsistent ⁵.

Give the Scorer More Words, Not More Volume

The durable move on a content-scored interview is to say more of substance: the situation, what you actually did, what changed, and what you would do differently. A thin answer is the failure mode the vendor statements and the two candidate studies above point to. Length alone proves nothing. But a scorer reading a transcript can only work with what is in it.

Four things worth doing, none of them a performance:

• Answer the question that was asked, then give one concrete example with a beginning and an end.
• Use any thinking time you are given to pick the example rather than to rehearse a tone.
• Say the interpersonal part out loud, because a transcript cannot infer the meeting you defused.
• Finish the answer rather than the clock.

For candidates who present differently on camera, the nearest measurement of the format question comes from first-impression research. A meta-analysis of 21 articles found that non-autistic observers formed less favourable first impressions of autistic people. The size of that penalty tracked the medium. On a standardized scale, it was 0.63 when observers had audio and video, 0.56 for video alone, and 0.22 for a transcript, where the effect could no longer be told apart from zero ⁶.

None of the included studies measured an actual hiring decision, the transcript group was the smallest, and a null result is not proof that text is neutral ⁶. What it supports is an ordering. A person watching a video judges presentation more than a person reading the transcript does.

Employers weigh the same format from the other side. Whether the async video round is worth keeping when answers are scripted is written for the hiring team, and what the record shows about AI interviews, accents and disability covers the accommodation route when the format itself is the barrier.

The format is the employer's choice, and asking about it is open to you.

Ask who watches the recording. It is a fair question for a recruiter, and the answer tells you which half of this article applies to you.

See how it works >