What Reading Off a Second Screen Does to an AI Interview Answer

Does the AI Scorer See Your Second Screen?

No, not on the platform this question is usually about. HireVue's 2025 Explainability Statement says its AI relies only on what the candidate says and uses no video analysis, so it does not assess facial expressions, body language, your background and surroundings, or tone of voice ¹. A model reading a transcript has no way to know where your eyes were.

That statement is the vendor's own description of its product, not an independent audit.

The pipeline the same document describes is speech-to-text transcription, then a language model, then a score benchmarked against expert human ratings of answers to similar competency questions ¹. Nothing in it reaches your other windows. It also calls itself a living document, so it describes a 2025 configuration rather than whatever ran in an earlier year.

Vervoe publishes the same shape of answer and adds the part people skip past. Its candidate help article says that for audio or video questions the AI reviews only the transcript, and that your video and audio are transcribed and provided to the employer along with the recording ².

Two things follow from that split, and they pull in opposite directions. A transcript keeps no record of where you were looking, so a model built on one has nothing there to mark you down for. The video keeps everything: the pauses, the delivery, the small stall while you find the next line, all of it open to whatever a person watching reads into it. Which of the two a given employer weighs is a question about that employer, and no vendor document answers it.

The scorer stops at the words, and the recording does not.

Who Notices a Read Answer?

A person does, if anyone does. Reporting in March 2025 on Amazon's recruiter guidelines, relaying internal documents, listed the tells interviewers were given: candidates typing while being asked questions, candidates appearing to read rather than respond naturally, and eyes wandering ³. Those are heuristics, not findings, and they describe a nervous candidate as easily as a scripted one.

Amazon's on-record position, given through a spokesperson the same month, is narrower. Candidates must acknowledge they will not use unauthorized tools, such as generative AI, during the interview or assessment process ³. The reported guidance adds that failing to follow it may result in disqualification, which is discretionary wording, and nothing in the reporting says how often it has happened.

Notes on a desk and a copilot feeding you answers are not the same thing.

Recruiters reading this site get a much less confident answer than the folklore gives. Whether a candidate's eyes moving off-camera is cheating or nerves sets out the evidence that gaze does not track deception, and the reasons a glance sideways also tracks low vision, autism, a second monitor, or notes written before the call.

The evidence is thinner than the confidence people bring to it.

Assessment platforms are a different product class and some advertise more. HackerRank's Screen page lists multiple-monitor detection among the signals in its integrity stack, alongside tab proctoring, copy and paste tracking, and full session replay ⁴. Which of those a candidate meets depends on what the employer switched on, and the page publishes no accuracy figure. What a signal produces there is a flag shown to an employer, not a verdict about you.

A detection claim is an advertisement until somebody audits it.

What Does Reading Cost the Answer?

Specificity, mostly, which is the thing a script read aloud tends to trade away. In an eight-question mock video interview completed by 582 participants across ten countries, honest impression management such as self-promotion was associated with higher rated performance, while deceptive tactics such as extensive image creation were associated with lower ⁵.

That study was a mock interview with no job at stake, rated by people rather than by a scoring model, and tactic use was self-reported. It shows association, not that any platform penalizes invention. Reading a true answer off a page is not deception either.

But it points at something usable. What went with higher ratings there was honest self-promotion, which comes down to your own material, and a script read aloud can sand that down. The names come out, the numbers round off, and the sentence that would have been yours turns into one anybody could have said.

Written prose read aloud has a sound, and it is the sound of nobody in particular.

The follow-up is where a written answer meets its limit, and nobody has to catch you first. A recruiter who liked the recording asks about that project in the live round, and the question lands where the page did not go: the option you rejected, the number underneath the number, the part that went badly. An answer built out of your own memory has somewhere to move. One assembled the night before, in sentences written to sound finished, tends to stop at the edge of what was written.

Hiring teams are being told the same thing from their side of the table. Whether the async video round is worth keeping when answers are scripted argues that the round still separates people on delivery and hardly at all on substance, and tells employers to move the deciding into twenty minutes somebody watches live.

Which is the point where a script runs out. Why rehearsed STAR answers sound generic is the same problem from your chair: depth survives a follow-up question and a script does not, so an answer you read your way through is one you still have to defend later.

Keep the Notes, Lose the Script

Where notes are allowed, keep names, dates and numbers where you can see them, and speak the rest. A page of full prose pulls your eyes down to it. Four words on a card ask far less of you, and they hold up in a live round, where nothing you wrote is in front of you and the question arrives without warning.

That is preparation rather than a trick.

Timing is the mechanical part worth protecting. HireVue's statement gives two conditions for an answer to be scored by AI at all: the candidate consented, and the system detected enough content. It names the ways that fails, including an answer too short, an answer not given in the allotted time, speech it cannot understand clearly, or a microphone problem. When either condition is unmet, the response goes to human review instead ⁶.

Scrolling for a paragraph you cannot find is one way a timer runs out.

Read the invitation for the rule that applies to you. Two questions get collapsed into one here, and they are different: whether notes are allowed, and whether a generative AI tool is. Employers vary on both, some put it in writing, and if the instructions say nothing, the recruiter is the person to ask.

There is an hour of preparation here worth doing whoever is watching. How to prepare for a one-way video interview pulls the competencies out of the posting, puts one real example against each with a number and a decision in it, and records one answer to watch back.

Whatever you write down, write it so you can put it down.

See how it works >