What an AI Interview Does With Your Eyes, and What It Doesn't

What Does the Score Actually Read?

For HireVue, the published answer is the transcript. Its 2025 Explainability Statement says the AI relies only on what a candidate says, uses no video analysis and no other audio characteristics, and does not assess facial expressions, body language, background and surroundings, or tone of voice ¹. The pipeline it describes runs speech to text, then a language model, then a score.

That is the company describing its own product, which is not an audit.

The document also dates itself. It calls itself a living document, updated from time to time, so it describes the 2025 configuration rather than whatever a candidate met in an earlier year. And it covers HireVue's AI-scored interviews only. It says nothing about what another vendor's system reads.

The change underneath it has a date. On 12 January 2021, HireVue's chief data scientist wrote that internal research had found visual analysis carried far less correlation to job performance than the other parts of the assessment, for the significant majority of jobs ². On that basis, the company said it would use no visual analysis in its pre-hire algorithms going forward. Wired reported the announcement the same day and quoted the chief executive on why: it was adding some value for customers, but it wasn't worth the concern ³.

No data was published with the internal research.

Other platforms are not HireVue, and the answer does not transfer between them. The candidate-side answer to what an AI interview actually scores makes the same point: no published source covers every AI interview, so the question has to be asked about the specific tool.

Why Does the Eye-Contact Warning Persist?

Because it described the product accurately once. Upturn's December 2018 report set out HireVue's assessment as it was then sold: the tool parsed recorded answers for facial expression and eye contact, vocal indications of enthusiasm, and word choice ⁴. It produced an insight score of 0 to 100, which an employer could configure to pass high scorers forward or reject low ones automatically ⁴.

That description is close to eight years old, and it is the one that still circulates.

The number attached to it comes from the same period. On 6 November 2019 the Electronic Privacy Information Center filed a complaint with the Federal Trade Commission, quoting HireVue's chief industrial-organizational psychologist saying facial action units could make up 29 percent of a candidate's score ⁵. A complaint is a set of allegations by an advocacy group, not a finding.

Wired reported in January 2021 that the agency had not acted on it ³.

Practice tools keep the advice alive in the present tense. Big Interview's VideoAI page says the tool evaluates four things in a recorded practice answer: pace of speech, vocabulary, eye contact, and body language ⁶. Eye contact there means staying focused and engaged without staring or looking away too often. Three of those four are delivery signals that HireVue states its own scorer does not use. Practicing against that feedback prepares you for a person watching the tape, which is a real audience, and not for a transcript.

Some of the writing that answers this question is published by companies selling a candidate a tool. That is not a reason to disbelieve any of it. It is a reason to check the date on a claim about a specific platform, and to check who published it.

Who Is Actually Watching the Recording?

A person can be. HireVue's statement says recruiters manually review and rate the responses of candidates who opt out of AI scoring ¹¹. Its chief executive told Wired in January 2021 that the company had helped screen more than 6 million videos over the previous year, some of them only transcribed for a human interviewer rather than scored ³. That figure is what an executive told a reporter.

Both of those describe a person, not the model, judging the answer.

Human raters do respond to what the camera shows. In an experiment with 114 participants comparing face-to-face and videoconference interviews, interviewees were rated lower in the videoconference condition ⁷. Differences in perceived social presence, perceived eye contact and impression management contributed to that gap. Ratings made from recordings were also lower than ratings made live, which is exactly the job a one-way interview hands its reviewer. It was a laboratory experiment published in 2021, and the effect sizes sit in the paper.

That study says nothing about an algorithm. A scorer reading only a transcript would be blind to the presence and eye-contact effects that moved those human ratings. What the finding supports is ordinary and small: put the camera near eye level, look toward the lens while you are speaking, and let the rest go.

It is not a reason to hold a stare.

Looking away to think is what thinking looks like.

The employer-side guidance is blunter about this than most candidate-facing pages. What a candidate's eyes moving off camera means tells hiring teams that eye movement does not track deception. It also names what off-camera glances track instead: low vision, autism, a second monitor, and notes written before the call. Acting on them, that guidance says, risks screening on a disability.

Read the Consent Screen Before You Record

The notice you click past is where a platform states what it collects. HackerRank's Candidate AI Notice says its AI features may analyze a candidate's coding, behavioral signals, webcam images and screen activity ⁸. The stated purpose is detecting anomalies, impersonation and unauthorized use during assessments and interviews, and proceeding with an AI feature is taken as acknowledgement of the notice ⁸.

The same notice lists what a candidate may request ⁸.

• an alternative selection process, or a reasonable accommodation
• human review of a decision made or substantially assisted by AI
• correction of inaccurate personal information
• an explanation of how AI was used if an adverse decision is made

Those requests go to the employer rather than to HackerRank, and the notice does not say an employer has to grant them. Outside a jurisdiction that gives you the same right by statute, the request rests on the employer's policy.

Some of this is statute rather than platform policy. Maryland has barred an employer since 1 October 2020 from using a facial recognition service to create a facial template during an applicant's interview, unless the applicant signs a waiver ¹⁰. The waiver has to name the applicant, the date of the interview, the consent itself, and whether the applicant read it.

It reaches facial templates only, not video recording, voice analysis or scoring of what you said. It names no penalty, and it does not stop an employer from declining to interview an applicant who will not sign.

In the EU, the Commission's guidelines of 29 July 2025 on prohibited AI practices read the emotion-recognition ban as reaching candidates during selection and hiring, and state that using emotion recognition during recruitment is prohibited ⁹. Commission guidelines are interpretive rather than binding, and only the Court of Justice can settle the reading.

None of that tells you what the tool in your own invitation does.

So ask before the recording opens. The invitation, the platform's candidate notice and the recruiter are the three places the answer lives, and asking for it is an ordinary request. What rights you have before an AI interview sets out which states already require notice, consent or a signed waiver, and the date each one started. Most of the country still requires none of it.

See how it works >