AI-scored video interviews — HireVue, Modum, Spark Hire, and a handful of competitors — are now a routine first-round screen at large employers in finance, retail, healthcare, and some consulting. They're talked about with a mix of dread and conspiracy. The dread is sometimes warranted; the conspiracy mostly isn't. The systems do real, consequential work, but they don't do the work the popular narrative says they do.

This post is what the AI is actually scoring, what it isn't, and how to prepare without overcorrecting.

What's actually being scored

The current generation of AI video-interview scoring is dominated by NLP — natural language processing on the transcript of your answer. The composite signal at major vendors breaks down roughly:

Speech content (~40% of weight). The transcript is analyzed for keyword presence, answer structure, length, and topic relevance to the question. This is the largest single bucket. The AI is essentially reading what you said and grading it against a rubric the employer built for the question.

Speech delivery (~25% of weight). Pace, clarity, hesitation patterns, filler-word frequency. This isn't tone-of-voice in the human sense — it's measurable speech characteristics. Speaking too fast, too slow, with excessive filler words ("um," "like," "you know") affects the score modestly.

Rubric matching (~20% of weight). For behavioral questions, the AI looks for STAR-like patterns — situation, task, action, result. Answers that follow the expected structure score better. Answers that ramble or skip the structure score worse.

Visual scoring (~15% of weight, declining). This is the part that gets the most press. At several major vendors, this category has been substantially reduced or removed entirely (more on that below). What remains is mostly about presence — were you looking at the camera, was the lighting okay, did you appear engaged.

The composite gets compared against a model trained on the employer's previous hiring decisions. The result is a score and a rank within the candidate pool. The top X candidates by that ranking are forwarded to a human recruiter, who makes the actual interview decision.

What it isn't doing (popular myths)

The popular narrative about AI video interviews exaggerates several things:

It doesn't reliably read personality from your face. The research on facial-expression analysis predicting job performance is weak and was the subject of a sustained controversy that led most major vendors to remove or significantly reduce that scoring component. HireVue removed facial-expression analysis from its scoring algorithm in January 2021 following an independent audit. Modum, Spark Hire, and Yobs followed. The current AI is heavily weighted toward what you said, not how your face looked when you said it.

It doesn't detect lying. The "AI lie-detector" framing is marketing copy, not capability. No deployed video-interview AI has demonstrated reliable lie-detection in peer-reviewed evaluation. Don't optimize for trying to "look honest."

It doesn't make the hiring decision. Even at companies that use AI scoring heavily, the final interview and hiring decision is made by humans. The AI determines the short list. The recruiter and hiring manager determine the offer.

It doesn't score attractiveness or accent (at major vendors). Following EEOC scrutiny and audit pressure, the major vendors removed or reduced features that scored protected-class signals. Smaller and offshore vendors may differ — but if you're interviewing at a US Fortune 500, the platform almost certainly does not score accent or attractiveness as primary features. Accent can still affect transcription quality, which indirectly affects scoring.

What the regulatory picture looks like

The regulatory environment matters because it changed what the systems do. Three landmarks:

• 2021: HireVue removes facial-expression analysis from its scoring algorithm after an independent algorithmic audit. The change was prompted in part by an EPIC complaint to the FTC and growing scrutiny from civil-rights groups.

• 2023: New York City Local Law 144 required bias audits and candidate notification for automated employment decision tools. This significantly increased compliance pressure on vendors operating in NYC.

• 2024: EEOC guidance on AI in hiring clarified that employers using AI tools are responsible for discriminatory outcomes, regardless of vendor claims. This pushed most major employers toward audited, conservative scoring approaches.

The cumulative effect is that the AI video interview in 2026 is a more conservative, transcript-heavy system than the version that ran in 2019. The face-reading component has shrunk; the language-analysis component has grown. The preparation calculus has shifted along with it: the right prep looks more like preparing for a structured behavioral interview than for a personality test.

How to prepare

Two clear implications follow from what's actually being scored:

Treat the answer like a structured behavioral interview. The AI is looking for STAR-like patterns (situation, task, action, result). It's looking for relevant keywords from the JD. It's looking for answers in the 60-120 second range. Practice answering the predictable behavioral questions ("tell me about a time you led a team," "describe a conflict") in that structure, with role-relevant vocabulary, in that length.

Optimize speech delivery to a reasonable middle. Don't speak unusually fast (penalized as rushed) or unusually slow (penalized as uncertain). Reduce filler words. Don't try to over-control your tone — natural cadence at a normal pace is what the system scores best. The point isn't to sound like a podcast host; it's to not sound like you're reading or panicking.

A few practical setup tips that aren't AI-specific but help:

• Use a wired headset for better audio. Poor audio degrades transcription, which degrades scoring.
• Light your face from the front, not from behind. Backlit candidates sometimes register as low-presence in visual scoring.
• Look at the camera, not at your own image. The visual scoring rewards camera-attention.
• Practice with the platform's practice mode if available. HireVue has one; use it.

For the broader behavioral-interview prep this overlaps with, see behavioral-interview-star-framework. For the related question of how AI is being applied to other parts of hiring (cover letters, resume screening), see ai-cover-letter-detection-truth.

What this isn't

A few clarifications:

• It's not the same for every vendor or employer. Smaller platforms and offshore vendors sometimes do things the major ones have stopped doing. If you're interviewing through an unfamiliar tool, the assumptions in this post may not fully apply.
• It's not always a hard gate. Many employers use AI scoring as one input among several in early-round screens, not as a strict cutoff. A middling score with a strong human-recruiter touchpoint can still advance.
• It's not bias-free even when scored well. The audits caught major issues but not all of them. If you suspect a specific platform is mistreating your demographic, EEOC complaint procedures exist and are increasingly enforced.

The short version: the AI scores transcripts more than faces, structured behavioral answers more than personality, and pace and clarity more than enthusiasm. Prepare like it's a structured behavioral interview, use a wired headset, look at the camera, and don't try to game the visual scoring — there's less of it than the headlines suggest.