Is the Async Video Round Worth Keeping When Answers Are Scripted?

Should you keep the async video round?

Keep it as a scheduling instrument and stop using it as the stage where candidates are compared. Those are two different rounds wearing one name. The first solves time zones, volume, and a hiring manager's full calendar. The second asks a recording to tell you whose answer was better, and that is the job a scripted answer stopped doing.

The logistics savings are real, and they are the honest reason to keep it. A scoping review of 43 studies of video-based interviewing in medical selection, 26 peer-reviewed manuscripts and 17 conference abstracts, found 13 reporting reduced financial costs for applicants, programs or both, and 11 reporting that applicants spent less time on the process ¹. Most participants still preferred face-to-face interviews, and the authors noted that the reasons for that preference are not clear from the available literature ¹.

The same review is careful about what the format can carry: one-way interviews may hold some promise as an initial screen, with no evidence found that they should replace a bidirectional interview ¹. That was written in 2022, before a polished answer cost a candidate five minutes and a browser tab. It has aged in one direction only.

What does a scripted answer still separate?

Delivery, and not much else. In a preregistered experiment, 245 Prolific respondents completed an asynchronous video interview after being randomly assigned to answer unassisted, to read ChatGPT's answers word for word, or to personalize them with a resume first. Both ChatGPT conditions scored considerably higher on overall performance and content. Response delivery ratings did not differ between conditions ².

Read that split slowly, because it is the whole decision. Content is the dimension the assistant moved, and content is what your rubric weights and what reviewers write their notes about. Delivery is the dimension it left alone: pace, warmth, whether a stranger can follow two minutes of explanation without a diagram.

Delivery is a genuine requirement in a narrow set of jobs. Inside sales, support, clinical intake and client coordination all turn on whether somebody is intelligible and steady with a person they have never met, and a recording tests that directly. For the analyst who will spend the year writing, or the engineer who will spend it reading other people's code, two minutes of talking to a webcam measures something the job asks for at the margins.

So the round did not stop working. It narrowed, and most teams still point it at everything. The same collapse is visible in live rounds, where every candidate gives the same polished STAR answer, and the cause is identical: the format asks for a rehearsed paragraph, and rehearsal got cheap.

Don't screen for the reading

There is no reliable tell, and the experiment that measured suspicion found it changed nothing. Raters gave the ChatGPT conditions lower honesty ratings and still scored them considerably higher on content and overall performance ². A hunch that does not move the score is not a screen. It is a rejection reason you would have to write down and could not support if anybody asked.

The behaviors offered as proof are the ones a recording exaggerates for everybody. Eyes moving off camera is the most popular, and gaze is not evidence in either direction. Neither is flat affect, a long pause before the first word, or a sentence that lands too neatly. Those describe a nervous person, somebody composing in a second language, and a candidate who prepared exactly as your invitation told them to.

There is also less to object to than the framing suggests. Preparing with an assistant sits in the same category as rehearsing with a friend or reading the company's last three product posts. If some uses are out of bounds for your round, name them in the invitation before anyone records: what is allowed, what has to be disclosed, and what is being judged. A rule that appears after you watch one video gets applied to one person.

Hold yourself to one line. No rejection reason may be about how an answer sounded. Write the sentence you would be willing to send, because the difference between a defensible screen and a bad afternoon is whether it names something the candidate did, which is also most of what you owe a candidate an automated step screened out.

What does the round cost before anyone gets hired?

Three things, and only one of them shows up on your calendar. Reviewing time scales with the applicant count rather than the shortlist: three questions at two minutes each, across 200 applicants, is twenty hours of watching before a single conversation happens. The other two costs are the candidates who never finish, and a format that cannot ask a second question.

Drop-off is the cost nobody measures, because it happens where you cannot see it. In a qualitative study of 15 HR professionals in Türkiye, practitioners described candidates who keep postponing the recording and fail to complete it on time, technical trouble that creates panic when a camera fails or a connection drops, and the right candidates being lost at that stage to the stress ³. Fifteen interviews is not a prevalence estimate. It is a description of a failure mode from the people who watch it happen.

The structural cost is the one better questions cannot fix. Those same practitioners named it directly: the format leaves no way to ask an extra question or give feedback when an answer half-lands ³. MIT Sloan Management Review puts the consequence plainly, warning that polished, contextualized responses candidates merely parrot may lead hiring managers to mistakenly attribute knowledge, skills and abilities to them ⁴.

Then weigh the round against the one it displaced. A 2022 re-analysis of meta-analytic validity in personnel selection concluded that range-restriction corrections had substantially overestimated many procedures: most of the ones that ranked high before stayed high, with mean validity estimates reduced by .10 to .20 points, and structured interviews emerged as the top-ranked selection procedure ⁵. Feeding a structured conversation is a fine use of an async round. Replacing one trades the best-evidenced instrument for the most convenient, and whether structured interviews survive AI coaching is a separate question with a better answer.

Swap the round for twenty minutes you watch

Spend the same candidate-minutes on work somebody watches happen. A one-way round costs an applicant the setup, the retakes and the recording, and costs you the review afterward, for an artifact that was rehearsed by design. Twenty live minutes on a small piece of the real job buys the two things a recording structurally cannot: the decision while it is being made, and the follow-up that lands on it.

Hand over material that settles nothing until somebody opens it, and let them use whatever they would use on a Tuesday.

• Support and inside sales. One live ticket and a knowledge base with one out-of-date article in it. Ask what they would send, and what they would escalate instead of sending.
• Financial analysis. One page of a filing and a claim about it. Ask which number the claim rests on, and whether that page supports it.
• Marketing. A research packet whose most quotable statistic comes from a vendor's own 40-person customer survey. Ask which claim goes in the deck.
• Software engineering. A failing test and an assistant that will propose a fix in seconds. Ask what gets checked before it merges.
• Operations and coordination. A scheduling conflict with three stated constraints and one nobody wrote down. Ask what moves, and who gets told.

Twenty minutes is enough because the deliverable is not the thing being graded. You are watching four moves: what they asked before producing anything, what they demanded a source for, what they refused, and what they checked against something outside the conversation. Take-home or live working session is the same trade at a longer horizon, and the shorter the task, the more the watching has to carry. See how Olive measures this.

Rewrite the questions if the round stays

Change what a recorded answer has to contain. "Tell me about a time you handled a difficult stakeholder" is a template, and a template is what a language model fills best. A question about the candidate's own artifact, a decision under a constraint you supply, or something they could not settle has no generic answer sitting ready to be pasted in.

Three shapes that survive rehearsal:

• Their own artifact. Ask them to open something they built or wrote in the last year, name the part they would do differently, and say what changing it now would cost.
• A constraint they did not expect. Put one paragraph of context in the invitation, then set a limit inside the question: half the budget, a week less, a stakeholder who will not agree.
• The thing that stayed unsettled. Ask for a decision on that project they were not confident about, and what would have changed their mind.

MIT SMR's probes work on a recording too. Ask whether the candidate can explain how to do something rather than what to do, why it works, when and for whom it is more effective, and what alternatives they considered, all of which push past rehearsed or surface-level answers ⁴. Those are the same rungs that make follow-up questions expose real understanding in a live round.

Then cap the format so it costs what it is worth. Three questions, ninety seconds each, retakes allowed and said so up front, one short paragraph naming what is being judged and who watches it, and a line saying that preparing with an assistant is fine. Be honest about the ceiling while you are there: a recording still cannot ask the second question, so the best a rewritten round becomes is a better first filter, never the place the decision gets made.

See a sample report >