AI Video Interviews: What They Actually Score, and How to Prepare

You get an email inviting you to a “video assessment.” No live interviewer — just a webcam, a set of questions on screen, and a countdown timer. This is a one-way or asynchronous video interview, and some form of it is now common in high-volume hiring, especially early-career and hourly roles. Understanding what it grades takes most of the anxiety out of it.

What changed: from faces to words

Early versions of these tools (circa 2017–2019) claimed to analyze facial expressions, “micro-expressions,” and tone of voice to infer traits. That approach drew heavy criticism from researchers and regulators — there is little scientific basis for reading personality or competence off a face, and it created obvious bias and disability-discrimination risks.

By the early 2020s the leading vendors had stepped back from facial analysis. HireVue, the best-known platform, announced in 2021 it would stop using visual analysis in its assessments, citing both an independent audit and its own research. Others followed. So in 2026, what most of these tools score is:

• The transcript of what you said — your speech is transcribed and analyzed as text, the same kind of language modeling used elsewhere in hiring.
• Structured competency signals — whether your answer covered the elements a good response to that question should contain.
• Sometimes basic delivery metrics — pace and filler-word frequency, framed as communication signals, not personality reads.

The mental model to hold: it’s grading your answer, not your face. Treat it like a written screening question you happen to be speaking.

What a good answer looks like to the scorer

Because the substance is transcript-driven, the same things that make an answer good to a human make it score well:

• Use a clear structure. For behavioral questions, the STAR shape — Situation, Task, Action, Result — maps almost exactly onto the competencies these systems look for. Name the situation, what you did, and the outcome.
• Be concrete and specific. Numbers, named tools, actual decisions. Vague answers score thin because there’s little substance in the transcript to credit.
• Answer the question that was asked. Rambling off-topic hurts you; the model is checking whether your answer contains the relevant content.
• Speak clearly enough to transcribe. Accent is fine — modern speech-to-text handles a wide range — but mumbling, heavy background noise, and a bad mic degrade the transcript, and a garbled transcript scores worse.

Practical prep

• Do the practice question for real. Most platforms give an untimed practice round. Use it to fix your framing, lighting, and audio, not just to click through.
• Look at the camera, not the screen. It reads as engaged and, more importantly, keeps you from staring at your own preview and freezing.
• Prepare stories, not scripts. Have four or five concrete examples ready that you can bend to common competencies — leadership, conflict, failure, initiative. Reciting a memorized script sounds like reciting a memorized script.
• Mind the timer, but don’t race. You usually get 30–60 seconds to think and 1–3 minutes to answer, sometimes with limited retakes. Know the limits before you start.

Your rights are expanding

Because these tools sit at the intersection of hiring, AI, and disability law, candidates increasingly have rights around them. Illinois’ AI Video Interview Act requires employers to disclose that AI is being used, explain what it evaluates, and get consent. NYC’s Local Law 144 requires bias audits and notice for automated employment decision tools. If a video interview is scored by AI, you can usually ask what it evaluates and, where a disability affects your delivery, request an accommodation such as extra time or an alternative format. Neither of those laws forces an employer to offer an alternative process, but disability law can independently require a reasonable accommodation — so it’s worth asking.

The short version: it’s a spoken answer to a structured question, graded mostly on what you say. Prepare like it’s an interview, deliver like it’s a recording, and the “AI” part stops being the scary part.