No Published Source Says an AI Scores You for Sounding Spontaneous

Is Sounding Spontaneous Anywhere in the Score?

No published source shows a vendor scoring how spontaneous an answer sounds.HireVue's 2025 explainability statement spells out its scoring in detail, and what it describes is something else. The document sets out three stages: speech-to-text transcription, then a natural-language model, then a score benchmarked against expert human rater evaluations of answers to similar competency-based questions ¹.

The same document says the AI relies only on what the candidate says. It uses no video analysis and no other audio characteristics, so facial expression, body language, background and tone of voice sit outside the score ¹. A transcript carries the words that were said. Nothing in it records whether those words arrived on the first attempt or the fourth, or whether they were written out the night before and read back.

The statement is HireVue's own account of its product, not an independent audit.

It is also a description of one company's assessments rather than of every system you might be sent. That same company stopped using visual analysis in its scoring years ago and said why. A post from its chief data scientist in January 2021 reported that internal research found visual analysis had far less correlation to job performance than other parts of the assessment, for the significant majority of jobs ⁵. On that basis it said it would use no visual analysis in pre-hire algorithms going forward.

No data came with that claim, and no outside auditor has examined it.

Finding out what the system in front of you reads is its own job, and what an AI interview scores, and what it does not covers it.

Where Does the First-Take Rule Come From?

From advice pages, not from research. The claim that first takes score better, and that an over-rehearsed answer gets marked down, turns up in advice guides with no study, no patent and no vendor page attached to it. The evidence that does exist points the other way.

Lukacik and Bourdage ran three mock asynchronous video interview experiments testing six design features, with 206, 230 and 297 participants ². Two samples came from an online panel and the third was university students. Longer preparation time and the chance to review and re-record answers increased interview performance ratings. Warnings about faking changed nothing, and neither did telling people that a human or an automated system would be evaluating them ².

If a second take cost anything with these raters, that is where it would have shown.

The hedges are doing real work here. These were mock interviews with no job at stake, and the answers were rated by the researchers' own raters rather than by a commercial scoring model. The automated-evaluation condition was what participants were told, not an algorithm that ran. So this is not proof that an AI rewards a second take. It is the closest measured test of the idea, and it moved in the opposite direction from the advice.

A retake is still not free. The employer sets how many takes you get, and that can be just one.

What Does a Script Actually Cost You?

A memorised script costs you at the follow-up rather than in any published score. It fits one phrasing of one question. Reword the question, or ask what got rejected along the way, or ask for the part that went badly, and the prepared sentences stop answering what was asked.

That gap is exactly where the advice for the other chair now aims. The piece on why every candidate gives the same polished STAR answer tells hiring teams to stop asking for more of the story. It tells them to ask for what the story left out: the option rejected, the person who disagreed, the baseline under the number. A script survives the first question and comes apart on the second.

The employer-side advice on the round itself goes further, and not toward a spontaneity detector.

Hiring teams asking whether the async video round is worth keeping when the answers are read off a second screen are weighing a different move. They can change what the round asks for, or drop the round. That is the employer-side response to scripted answers, and none of it depends on hearing that an answer was rehearsed. A rehearsal detector would not answer the question they are actually asking, which is whether the round still measures anything.

Writing an answer out in advance also drifts it general, because you write for the question you imagined rather than the one that arrives.

Prepare the Story, Not the Sentences

Prepare the material and leave the wording to the moment. Take three or four examples and work each one until you can start it anywhere: at the decision, at the mistake, at what it cost. Then answer the question that actually arrives, in whatever words you have that day.

What has been measured is adjacent rather than identical, and the useful part is who gained. Buehl and colleagues had 111 students complete the same 20-question structured interview twice, once under honest instructions and once instructed to answer as an applicant would, a condition the authors studied as faking. Mean ratings rose from 3.20 to 3.62 ³. The size of each person's gain tracked how well they could identify what each question was assessing ³.

That last part is the usable one, and it needs nothing invented: knowing what a question is assessing tells you which of your real examples to reach for.

The study is a long way from a hiring decision.

Students in a research interview, no job at stake, human raters sitting across a table, and a design in which everyone did the honest condition too. It says nothing about automated scoring. Sackett and colleagues rank structured interviews first among selection methods. They note that as interview administration and scoring move toward technology and automation, validity needs to be evaluated under those changed circumstances ⁴.

So the validity estimate people quote for a structured interview came from interviews humans conducted and humans rated. Carrying it across to a machine-scored version is an assumption rather than a finding, and the caution runs both ways: nothing there supports the spontaneity claim either.

When a guide tells you the AI can hear a rehearsed answer, ask which source says so.

Rehearsed answers do land flat, and that has a real cause on the human side of the table. Why your STAR answers sound like everyone else's now covers what changed and what still works.

See how it works >