Are one-way and AI interviews fair? What the research says

A one-way interview, also called an asynchronous or on-demand interview, sends a candidate a fixed set of questions to answer on camera with no interviewer present. An AI interview adds software that scores or summarizes the recording. Are they fair? It depends on how they are built.

That is a less satisfying answer than yes or no, but it is the honest one. The format has a real fairness advantage: everyone gets the same questions under the same conditions, and the selection-science research on structured interviews is broadly positive on that. Fairness is decided at the scoring step, not the recording step. A well-built structured interview can be fairer than a casual chat. A loosely scored one, or one run by a model nobody has tested, can be less fair.

The fairness question runs through both the recording and the software, so this page treats them together and flags where they differ. It works through what the research actually says, where the format helps fairness, where it can hurt it, and what separates a fair process from an unfair one.

What “fair” actually means here

Fairness is not one thing, and a lot of arguments talk past each other because of it. In selection research it breaks into three questions worth keeping separate. Consistency: does every candidate get the same questions, prep, and shot? Validity: does the thing being measured actually relate to the job, or is it scoring something irrelevant like how someone looks on camera? Adverse impact: do qualified people from a protected group pass at a meaningfully lower rate, for reasons unrelated to the job? That last one is the legal and ethical core, and the question regulators ask of any hiring tool, human or automated. A one-way interview can score well on the first two and still need watching on the third. Holding them apart is what lets you answer the fairness question without hand-waving.

The part of the case that rests on solid ground

The strongest fairness argument for one-way interviews is also the oldest and best-supported one in the field, and it has nothing to do with AI. It is about structure.

Decades of selection research, summarized across major meta-analyses, point the same direction: structured interviews predict job performance better than unstructured ones, and they tend to show smaller differences between demographic groups. A structured interview is one where every candidate gets the same questions and answers are rated against defined criteria, instead of a free-flowing conversation that wanders wherever rapport takes it. The unstructured chat feels fairer because it feels human, but it is the format most exposed to first-impression bias, similarity bias, and an interviewer simply remembering the last person more vividly.

A one-way interview is structured by construction. The questions are fixed, the order is fixed, and the conditions are close to identical for everyone. So the careful version of the claim, the one the literature supports, is that a structured format removes some of the inconsistency that unstructured interviews introduce. It does not make a process bias-free. It removes one specific, well-documented source of unfairness, the kind that comes from improvisation.

This is the same logic behind the effectiveness case for the format. Consistency is the quiet strength, and consistency and fairness are close cousins.

Where the fairness can break: the scoring step

Here is the part the structure argument leaves out. Structure earns you a fairer front door, but says nothing about what happens once the recording exists. Fairness is mostly won or lost at the scoring step, whether a human or a model does the scoring.

Humans bring their own bias to a recording. A reviewer can still mark down an accent, a home that looks unlike theirs, an older face, or a speech difference, exactly as they might in person. One recruiter, arguing against the format, put the worry directly: it is “extremely easy to discriminate based on what you see in the video, if they have a lisp or maybe a visible disability.” That risk is real. The defense is not the camera, it is a written rubric that scores the substance of the answer, plus reviewers who rate content over polish.

Algorithms can inherit the preferences in their training data. A model taught to predict “good hire” from a company’s past hires can learn that company’s past patterns, including the biased ones. This is the substance behind the concern people raise about AI interview scoring, and why both researchers and regulators treat hiring algorithms as something to be tested for adverse impact, never assumed neutral. New York City now requires bias audits of automated hiring tools. Illinois requires notice and consent before AI analyzes a video interview, plus deletion within 30 days on request. Colorado has a broader AI-hiring law phasing in. An algorithm in hiring is a tested instrument, not a trusted one.

The genuinely reassuring shift is in what the mainstream tools score. The original fear was a model grading facial expressions and eye contact. That practice peaked around 2019 to 2021, and HireVue, the largest vendor and the source of most of the alarm, removed facial analysis from its assessments in 2021. Most AI scoring today works from a transcript of the words you said, not your face, which we cover in full on do AI interviews use facial recognition. Scoring content rather than appearance is not a complete fix, but it removes the single most indefensible input.

The accessibility question, answered honestly

The fairness argument that deserves the most care is access, because a fixed-format recording lands differently on different people, and pretending otherwise is the fastest way to lose trust.

It can disadvantage some candidates. Someone who needs an accommodation, has a speech difference, is interviewing in a second language, or simply has no quiet, well-lit place to record after work is carrying a load the format does not see. One candidate described the hidden cost bluntly: a “one minute” video can mean cleaning the house, finding the right outfit, and recording many takes, “incredibly unfair to people who have families and children where they might not have a quiet place to record.” Unevenly distributed friction is a fairness problem, not just an inconvenience.

It can also help some candidates. People who freeze in live conversation, many neurodivergent candidates, and anyone who interviews better with time to think often do better when they can prepare and record on their own schedule. One person who offers it as an option said it lets “people with certain disabilities or neurodivergences to perform better.” Both things are true at once. The format removes one barrier and can raise another.

What makes the difference is whether accommodations are real. A fair process states up front that adjustments are available, offers a live alternative on request, and does not penalize the ask. If you need an adjustment you are entitled to request one, and our accommodation request email template gives you the words. More on your options is on video interview accommodations, and the legal backdrop on is it legal to record a job interview.

So, fairer or less fair than the alternative?

The right comparison is rarely the format versus perfect fairness. It is the format versus the thing it replaces, usually a quick, unstructured phone screen. Against that, a well-built one-way interview is plausibly fairer on consistency, which is the part the research supports. Against a well-structured live interview with trained reviewers, it trades away the live follow-up and adds the access friction above, and there it is not obviously fairer. So the defensible summary is narrow on purpose: the format has a built-in consistency advantage that can reduce some biases, the scoring step decides whether that advantage survives, and access has to be handled deliberately or the format quietly creates a new unfairness while removing an old one. Anyone who tells you it is simply fair, or simply biased, is giving you the easy version of a real question.

The questions that separate fair from unfair

Fairness in a recorded interview comes down to a short list of yes-or-no questions. They work whether you are a candidate sizing up a process or an employer building one.

1. Same questions for everyone? A fixed, structured set is the foundation. If the questions vary, the consistency advantage is gone before you start.
2. Scored on substance, not polish? Is there a written rubric that rates what the answer says, with reviewers told to weigh content over delivery, accent, and background? This is the single biggest lever on fairness.
3. Does a human make the decision? AI can transcribe, organize, and surface. A person should decide. Tools that surface and humans who decide is the defensible arrangement, and the one regulators increasingly expect.
4. Has the scoring been tested for adverse impact? If a model is involved, has anyone checked that it does not pass one group at a meaningfully lower rate for reasons unrelated to the job? Untested is not the same as unbiased.
5. Are accommodations and a live alternative genuinely available? Stated up front, offered on request, and with no penalty for asking.
6. Is the candidate told what is happening? What the tool does, whether anything visual is analyzed, and what happens to the recording. In a growing list of places this is a legal requirement, not a courtesy.

A process that can answer yes to those is a fair one, regardless of whether AI touches it. A process that cannot is where the bad experiences come from, and the failures are in execution, not in the format itself, which means they are fixable.

If you are facing one of these and want to know specifically what is being measured, read do AI interviews use facial recognition and can AI detect cheating in a video interview. If you are weighing whether the format earns its place at all, are one-way video interviews effective takes the employer’s side of the same question.