AI Interviews and One-Way Video Screens

The format drives away good candidates, and the AI doing the scoring appears to be fairer than the humans it replaced. Both findings come from the same experiment, they are uncomfortable together, and almost every article on this topic reports one while ignoring the other.

If you have been sent a link asking you to record answers to prompts on your own time, this is what is known about what you are walking into.

The study worth knowing about

Most writing on AI interviews cites vendor surveys. There is a better source: A Brave New World of Hiring, a 2026 natural field experiment by Mallory Avery, Edwin Ip, Andreas Leibbrandt, and Joseph Vecci, published as a CESifo working paper.

The design is unusually strong for this subject. Over 3,000 real job applicants were randomised into asynchronous audio interviews, asynchronous video interviews, live online interviews, and a control. Responses were assessed by professional recruiters and by a commercial AI recruitment tool of the kind used across large employers. Because it ran inside a real hiring process, it measures what applicants actually did rather than what they told a survey.

Three findings.

Asynchronous interviews cut application continuation by more than 50%. Not marginal candidates: the drop-off included the most qualified applicants. Adding a recording step to a funnel halves it.

The decline was largest for women. A companion vignette study attributed the effect to perceptions about the fairness and competitiveness of the process rather than to discomfort with technology.

The AI scored women and underrepresented racial minorities higher than the human recruiters did, and scored men, white, and Asian applicants lower. The human assessments displayed multiple cognitive biases; the AI system did not. The AI's scores also predicted subsequent employment success substantially better than the recruiters' did.

That third finding is the one nobody quotes, and it is the one that should change your mental model. The problem with asynchronous AI screening, in the best evidence available, is not primarily that the machine judges you unfairly. It is that the format is off-putting enough to make people quit before the machine sees them at all.

Two caveats worth keeping. This is one experiment, in one hiring context, with one commercial tool, and it is a working paper rather than a peer-reviewed publication. And "less biased than human recruiters" is a comparison, not a clean bill of health.

What this means when you get the invite

The rational response to that evidence is roughly the opposite of the common advice.

The common advice treats the AI as an adversary to be gamed with keyword-stuffed answers and forced eye contact. If the AI is the better-calibrated evaluator in the room, gaming it is both harder and less necessary than gaming a tired human.

The genuine risk is the drop-off. Over half of applicants are removing themselves, disproportionately the ones who suspect the process is unfair. Which means completing the recording is itself a meaningful edge, because a large fraction of your competition will not.

That is not an argument that the format is good. It is inconvenient, it strips out the two-way conversation, and there is real evidence it damages the applicant pool. But if you have decided to apply, quitting at the recording stage is the one choice that guarantees nothing.

How to actually do one well

The format has different mechanics from a live interview, and most preparation advice ignores the differences.

No one redirects you. In a live interview, a rambling answer gets interrupted and rescued. Here it runs until the timer stops. Answer the question asked, in structure, and stop when you are done.

Prepare four or five stories, not scripts. Situation, what you decided, what happened. Ninety seconds each. Most prompts in these screens are variations on a small set of behavioural questions, and five well-chosen stories cover the majority.

Record yourself in advance. Not to memorise, but to remove the strangeness of talking to a lens. The first take is always worse than the third, and many platforms give you fewer takes than you would like.

Check the retake policy before you start. Some allow unlimited retakes, some one, some none. This is usually stated on the instructions screen and rarely read.

Look at the lens, not at yourself. Everyone watches their own preview. The result is a recording of someone looking slightly down and away.

Handle the environment. Front-facing light, a plain background, a closed door. Research on asynchronous interviews finds background and setting affect ratings, which is a bias in the format rather than a reflection on you, and it is cheap to neutralise.

Do it early in the window. Deadline recordings are worse, and technical problems at 11pm on the last day have no recovery path.

Say the concrete thing. Whether a human or a model is scoring, specificity is what registers. "Cut onboarding time from three weeks to nine days for a 40-person team" carries information; "passionate about operational excellence" carries none. The same principle that makes a resume bullet work applies to a spoken answer.

What not to do

Do not use real-time AI assistance. Reportedly around a fifth of job seekers have tried this. It is visible in the pauses and in where the eyes go, a growing number of employers prohibit it explicitly, and discovery ends the process. Preparation with AI beforehand is ordinary and carries none of this exposure, which is the same editor-versus-ghostwriter line that applies to resumes.

Do not perform for an imagined algorithm. Advice circulates about smiling at fixed intervals or inserting job description keywords verbatim. It is guesswork about systems the writer has not seen, and it produces recordings that are strange to watch if a human reviews them, which one usually does at the next stage.

Do not assume nobody watches. These recordings are frequently reviewed by the hiring manager later. Record for a person.

Your rights, where you have them

This is the practical part almost no career article covers, and in three jurisdictions it is enforceable.

Illinois. The Artificial Intelligence Video Interview Act requires employers to tell you AI will analyse the interview, explain what characteristics it evaluates, and obtain your consent before proceeding. You may also request deletion of the video.

New York City. Local Law 144 covers automated employment decision tools. Employers must give notice before use, including what the tool assesses and what data it uses, must publish an annual bias audit summary, and must offer an alternative process where a reasonable one exists.

Elsewhere in the US and internationally, coverage is patchy and expanding, with Colorado, Connecticut, and the EU AI Act at various stages.

Two things follow. If a posting is in one of these jurisdictions and you received no disclosure, that is a compliance failure and you may ask about it. And where an alternative assessment is available, asking for one is a legitimate request rather than a difficult one.

Outside those places, declining a one-way interview usually ends the application. Reportedly about one in three candidates has walked away from a role rather than do one. That is a real choice with a real cost, and it is yours to make with the arithmetic visible.

Where this fits

One-way screening is now a normal stage in the funnel rather than an anomaly, and it sits at the front of it. Treat it as an extension of the screening problem rather than as an interview: it is the stage that decides whether a human ever engages with your application.

That reframes the preparation. A polished recording does not rescue a weak match. If only 2–3% of cold applications reach an interview, the recording is a gate you clear, not the thing that gets you hired. The leverage is still upstream, in applying to roles you genuinely fit and tailoring the application to them.

The short version

Asynchronous interviews cut applicant continuation by more than half in a controlled field experiment, and the effect was strongest for women and for strong candidates. The AI scoring those interviews was, in the same study, less biased than the human recruiters and better at predicting job performance. Prepare five structured stories, record early, look at the lens, skip the real-time AI assistance, and know that in Illinois and New York City you have disclosure rights worth using.