The One Setup Choice Published AI Interview Scorers Read

Which Part of Your Setup Reaches the Score?

The audio, and only the audio, on the two platforms that publish what their scorer reads. HireVue's 2025 explainability statement says its AI relies only on what a candidate says, with no video analysis. It names facial expressions, body language, background and surroundings, and tone of voice as things it does not assess ¹.

Vervoe publishes the same shape of statement in its candidate help center. For audio or video questions, the AI reviews only the transcript, and does not assess facial expressions, voice tone, accent or any other personal attribute ². The same page says the recording still goes to the employer along with the transcript ². Two vendors, two published sentences, and the same division: your words go to the model, your picture goes to a person.

Both are vendors describing their own products, not audits anybody ran on them.

A description can be accurate, incomplete, or out of date. HireVue's document calls itself a living document that will be updated from time to time, so it describes the 2025 configuration and not necessarily the one a candidate met in an earlier year ¹.

Neither statement covers the platform you have not been told about.

The name on your invitation is where to start. What an AI interview scores and what it does not covers how to find that out for the vendor in front of you. Do that before spending anything on the room.

Why Does Audio Change What the Transcript Says?

Because the score is computed from text, and the text comes from a machine listening to your room. HireVue describes a pipeline that runs speech-to-text transcription, then a natural-language model, then a score benchmarked against expert human raters on similar competency questions ¹. A sentence that arrives wrong is scored as the sentence that arrived.

The vendor also publishes what happens when the audio fails outright. An answer is machine-scored only if the candidate consented to it and the system detects enough content in the response, and the statement names a microphone problem as one way that second condition fails ³. Speech the system cannot understand clearly enough is another ³. When either condition is unmet, the response is not AI-scored and is flagged for human review instead ³.

HireVue calls that a small proportion of candidates and publishes no number for it ³.

Human review is not a good outcome or a bad one. It moves the answer to a person, which says nothing about what the person then decides. It also means an opt-out and a dead microphone arrive in the same queue ³.

A scorer cannot read a sentence the recognizer never received.

The transcript is an estimate either way. HireVue names Rev.ai as the supplier that produces it and publishes its own internal testing: US English word error under 10% on average for that system, against 15 to 25% for the others it tested ⁴.

The document gives no sample size and no independent replication, so it is a vendor's account of a supplier it chose.

Word error is also a rate across audio, not a count of ruined answers.

Where Do Lighting and Background Land?

On the two platforms that publish what they read, with the person who watches the recording, which is a separate step from the scoring. Vervoe states that video and audio are transcribed and provided to the employer along with the recording ². So the room you sit in is outside the scoring model and fully inside the process.

There, ordinary video-call judgment applies, and it is the same judgment anybody applies to any video call. A visible face, a room that doesn't pull the eye, and audio without an echo read as somebody who took the round seriously. None of that needs equipment. A lamp or a window in front of you costs nothing, and neither does a plain wall.

A person watching is not a neutral instrument either.

The hiring side has its own warning about reading a recording too closely. The employer-side page on a candidate whose eyes keep moving off-camera tells interviewers that off-camera glances also track low vision, autism, a second monitor and notes written before the call, so treating them as evidence risks screening on a disability.

What that page cannot do is tell you which employer read it.

So keep the frame plain rather than impressive. The visual channel reaches a human whose care you can't control, and a background that says nothing is the version with the least in it to be read wrong.

What If the Platform Publishes Nothing About What It Reads?

Then the visual channel cannot be ruled out, and one published test is the reason to say so. Reporters at Bayerischer Rundfunk ran several hundred video clips through Retorio, a German video-interview personality tool, in late 2020 and published the results in February 2021 ⁵. An actress repeated the same answer, and ten further test people had their videos edited one factor at a time ⁵.

A picture frame and woodchip wallpaper added behind a test person moved the personality scores by up to 15 points, and a bookshelf moved them by more ⁵. Glasses cut nearly 10 points from the actress's conscientiousness score ⁵. A headscarf moved her results by nearly 20 points, reading as more open, more conscientious and less neurotic ⁵. Repeating the answer unchanged gave stable results, and swapping the audio track left them virtually unchanged ⁵.

That is journalism rather than peer review: a handful of test people, no statistical testing, and one product as it stood in late 2020.

It still does the one job it needs to do here.

It shows a product responding to a change that had nothing to do with the answer, which is enough to stop anyone claiming the picture reaches no scorer anywhere. What it does not show is how often real applicants are affected, or in which direction, or whether that tool still behaves the same way.

One message to the recruiter can ask all of this: which platform runs the round, what it scores, and whether a person reviews the recording. A vendor that has published a statement has answered before you ask. A vendor that has not is the case where the plain wall is worth the two minutes.

Set the Audio Before You Buy Anything

Order the spending by what reaches the transcript. A wired headset, or any microphone close to your mouth, sits in the one channel two vendors say their scorers read. Everything visual comes after it, and most of what matters there is free.

• Record 30 seconds on the exact setup you plan to use, then play it back at low volume. A hum you stopped hearing an hour ago is still in the file.
• Shut the door, switch off the fan, and close the window on the traffic side.
• Sit closer to the microphone than feels natural, and slow down on names, tools and numbers.
• Face a window or a lamp instead of sitting in front of one.
• Leave the wall behind you plain. A virtual background is one more piece of software that can fail in the middle of an answer.

None of that promises anything about how the round turns out. It is the part of the setup you can check yourself, in the ten minutes before you start, by listening to your own recording instead of trusting a checklist.

The rest of the hour belongs to the answers.

Your last hour before recording is better spent elsewhere. How to prepare for a one-way video interview in an hour opens by telling you to skip the lighting advice, then spends the hour on competencies and worked examples. That page reaches the same verdict on lighting from the preparation side.

See how it works >