AI Video Interviews: How the Screen Actually Scores You (2026)

Most advice about the AI video interview is aimed at the wrong target.

Candidates rehearse eye contact, adjust their lighting, practise smiling at the webcam — preparing to be watched. But the systems doing the scoring stopped watching years ago. HireVue, the market leader, retired its facial analysis features in 2021 after external bias audits. What today's models evaluate is much closer to a transcript than a video: the structure, relevance and specificity of what you actually said.

That changes what preparation means. If the camera is effectively a microphone, your answers are the entire exam.

What is an AI video interview?

The format is standardised almost everywhere. You log in at a time you choose. A written or recorded question appears. A preparation timer runs — usually 30 to 90 seconds. Then a recording window opens, typically two to three minutes, and what you say in it becomes your answer. Most systems allow limited or no re-records.

The scale is easy to underestimate. HireVue alone processed more than 20 million assessments in the first quarter of 2024, and by 2026 roughly 21 to 23% of organisations use AI to run at least the initial interview. If you apply to a major bank, consultancy or graduate programme — HireVue's partners include JPMorgan, Goldman Sachs, Citi, Bain, BCG, IBM, Capital One, Microsoft and Amazon — this screen is very likely your first interview.

One structural fact matters more than everything else in this article: the AI does not reject you directly, but it decides who gets seen. It produces a scored shortlist for a human to review. Answers below the threshold never reach human eyes at all. The machine is the gatekeeper, not the decision-maker — which, from the candidate's side of the glass, is a distinction without much comfort.

Does the AI analyse your face and body language?

This is the most persistent myth in the category, and it survives because it feels plausible. A camera is pointed at you; surely the machine is reading your face.

It largely is not. The audit pressure that pushed facial analysis out of HireVue in 2021 moved the whole mainstream market in the same direction, and what current models evaluate is the language: whether the answer addresses the question asked, whether it is structured or rambling, whether it contains specifics or filler.

Think of it as a transcript being analysed, not a lie detector.

Two honest caveats. Audio quality still matters, because a transcript can only be scored if the speech recognition can produce one — a bad microphone hurts you mechanically, not aesthetically. And presentation is not entirely irrelevant, because a human reviews the shortlist afterwards. But the machine gate, the one that decides whether any human watches at all, is reading your words.

What kind of answer scores well?

The scoring dimensions reward exactly what a tired human screener would reward, applied with mechanical consistency:

On-question relevance. The model checks whether the content of your answer matches the question asked. A polished answer to a slightly different question — the classic politician's pivot — scores badly here, because the mismatch is precisely what the system measures.

Structure. Situation, action, result, in that order, with most of the time on action. The two-to-three-minute window fits this shape almost exactly: fifteen seconds of context, ninety seconds of what you did, fifteen seconds of outcome.

Specificity. Named tools, real numbers, actual situations. "I handled a difficult stakeholder" is unscoreable filler; "I inherited a client who had escalated twice in a quarter, and after three weeks of weekly calls the account renewed" is content.

Coherence. Rambling is penalised even when fluent. The model has no ear for charm.

Rambles: "So, um, teamwork is really important to me, I've always been a team player, like in my last role we had a lot of projects where collaboration was key and I always tried to make sure everyone was heard..."
Scores: "In my last role, two departments owned one launch and weren't speaking. I set up a shared tracker and a fifteen-minute daily sync. We shipped nine days early, and the same structure was reused for the next two launches."
Both took about twenty seconds to say. Only one of them contains anything a model u2014 or a person u2014 can evaluate.

How should you prepare for an AI video interview?

The preparation is unglamorous and it works.

Collect the raw material first. Six to eight real situations: a conflict, a failure, a deadline, a thing you built, a decision you got wrong, a result you can quantify. This is the same raw-material discipline that makes ChatGPT resume prompts produce usable output instead of filler — the specifics are the part no tool and no rehearsal can invent for you.

Map them to the values in the posting. Companies using these screens usually publish what they assess. KPMG scores against its stated values; Amazon maps questions to Leadership Principles, and serious candidates prepare two stories per principle. The job description tells you which stories to lead with.

Practise aloud, timed. Thirty seconds of prep, two minutes of answer, no second take. The first few attempts are reliably worse than people expect, which is the entire argument for doing them before the real one. A voice-based mock interview compresses this loop — question, timed spoken answer, feedback, again.

Fix the mechanical layer once. Quiet room, decent microphone, camera at eye level, face lit from the front. Not because the model scores it, but because the speech recognition and the eventual human reviewer both depend on it.

Can you refuse an AI video interview?

The candidate-side numbers here deserve to be better known. In Enhancv's April 2026 survey of 1,066 US job seekers, 50.5% had been rejected at least once in the past year without a single word from a human, and 63.8% of that group believed a machine made the call. Only 9.7% said an employer had ever clearly told them AI was involved.

And 31.4% said they had abandoned a role entirely rather than sit through a one-way screening. Notably, 79.1% of those abandoned roles paid under $100k — the people with the least leverage are opting out the most.

Whether that opt-out is principled or self-defeating depends on your situation, and it is genuinely both. Some jurisdictions now require disclosure of AI-based screening, and you can ask whether human review is available. But at companies where the screen is the process — and at HireVue's enterprise partners it usually is — declining the interview is declining the job. The pragmatic reading of the data: the screen rewards preparation heavily, most candidates prepare for the wrong thing, and that mismatch is an advantage available to anyone who prepares for the right one.

Half of job seekers (50.5%) had been rejected at least once in the past year without a single word from a human. Only 9.7% said an employer had ever clearly told them AI was involved.

Where does the AI interview fit in the rest of the screen?

It helps to see the whole pipeline, because each stage is scored on a different thing and preparation for one does not transfer automatically to the next.

The first gate reads your file. An AI resume checker shows you what that parser sees and whether your wording matches the posting — and whether employers can detect an AI resume is, as covered elsewhere in this series, the wrong worry at that stage too. The second gate, this one, reads your transcript. The third is finally human, and everything you claimed in the first two has to survive being asked about out loud.

The pattern across all three is the same and worth stating once: every gate scores specificity that could only be yours. The file, the recorded answer and the live conversation all fail on the same generic filler and all pass on the same named, numbered, real experience. Prepare that once, and you have prepared for all three.