Interviewing
Say the AI Rule Out Loud in the First Two Minutes
An AI question in an interview reads as a trap unless the rule comes first. Two sentences at the top of the round: using AI on this work is allowed, and what I want to talk about is how you used it rather than whether you did. Add that nothing they say about using it counts against them, and that a person makes the final decision. Then ask about choices instead of compliance. A candidate who believes disclosure is disqualifying will describe a careful workflow they do not have.
The takeInterviewers rehearse the question and skip the frame, which is backwards, because the frame decides what the question is able to collect. An AI question asked cold reads as an integrity check, since that is the genre the candidate has been reading about all year, and no amount of clever phrasing gets you out of it. Two sentences of permission change what the room is about. It is the cheapest improvement available anywhere in an interview loop, and it costs thirty seconds.
Where Olive fits
Open a role and see what the work shows
Olive says the same thing to the candidate before the session starts: the assistant is there, using it is expected, and what gets written up is what they did with it. The report that comes back is granted to the candidate as well, free, on every tier.
Rank your shortlistSay the rule and the reason in the first two minutes
Two sentences, before the first question, in your own words. Using AI on this kind of work is allowed here, and what I want to talk about is how you used it rather than whether you did. Then the clause that does the actual work: nothing you tell me about using it counts against you. Say it once, plainly, and move straight into the question.
The reason is worth attaching because an unexplained decision is the version candidates rate worst. In a four-scenario vignette experiment with 921 working-age Austrians, a rejection from an AI with no explanation scored lowest on all four outcomes measured, from 1.49 for recommendation intention to 1.86 for outcome fairness on a five-point scale, while an AI rejection that included an explanation drew the same ratings as a human rejection without one 1. Hypothetical rejections imagined by an online panel, not real applicants who were actually rejected. Every condition scored below the midpoint, so this is a ranking among unhappy outcomes. The direction still transfers: explaining is cheap and its absence is expensive.
A short version that works out loud, said once at the top of the round:
> Before we start on the work: you can use AI here, and I am going to ask how you used it. That is not a test of whether you should have. Nothing you tell me about using it counts against you, and a person makes the decision at the end of this process.
The last clause is not decoration. It answers the question the candidate is actually holding.
Why do candidates hear an AI question as a trap?
Because the loudest employer-facing writing on the subject is about screening out AI-written answers, and candidates have read it too. By the time you ask, the genre is already set, so the question sounds like an integrity check and the answer arrives shaped as a defense. None of your intent is visible from the other chair unless somebody says it out loud.
The format itself can produce what it was meant to prevent. A two-phase study of asynchronous AI interviewers analysed 11 subreddit discussion threads, interviewed 17 applicants, then evaluated a redesigned interface with 180 participants across six groups, and found that applicants' expectations, shaped by their own familiarity with these models, went unmet against employers' framing, damaging their sense of agency and trust and pushing them toward workarounds 2. A January 2026 preprint resting on 17 interviews and self-selected posts, testing a research prototype, not a commercial product. Read the workarounds finding as a description of what an unaccountable format provokes, not as evidence that candidates are dishonest.
Public opinion points at the same place, which is who decides. In a Pew survey of 11,004 US adults fielded in December 2022, 71% opposed AI making a final hiring decision against 7% in favor, while opinion on AI merely reviewing applications was far softer at 41% opposed, 28% in favor and 30% unsure 3. General-public attitudes before mainstream model use, toward AI in hiring as respondents imagined it. The useful part for a script is the shape rather than the size: the objection concentrates on the decision, so saying a person makes the call is worth the four words it takes. Whether the candidate should be using AI in the round at all is a separate decision, and letting candidates use AI during the interview or banning it sets out both versions.
Ask about choices rather than compliance
A compliance question asks what the candidate used and whether they were allowed to use it. A choice question asks what they decided and what they would do differently. Only the second kind produces evidence you can still read in six months, because the tooling turns over every quarter and the decision does not. Changing the verb removes most of the defensiveness on its own.
Three swaps that carry almost every round:
- Did you use AI on this? becomes *which part of this did you hand over, and which part did you keep?*
- How do you make sure the output is accurate? becomes *tell me about a time the draft was confidently wrong. How did you find out?*
- Are you comfortable with AI tools? becomes *what is the last thing you refused to let the assistant do, and why?*
Each one asks for a moment rather than a policy, and a moment is checkable. It also hands the candidate something to control, which matters more than it sounds. A review of applicants' procedural-fairness perceptions of algorithmic recruitment tools, drawing on seven studies with more than 1,300 participants between them, found overall fairness perceptions mixed but perceptions of behavioral control and social presence mostly negative: candidates feel less able to influence the outcome and feel the human element is missing 4. A narrative review of scenario-based studies with panel and student samples, by authors affiliated with the assessment industry, so it supports a direction, not a magnitude. Asking someone to narrate their own choices is the cheapest way to give some of that control back.
If you are not sure what a strong answer to these sounds like before you hear one, what being good at using AI actually looks like in an interview is worth ten minutes of panel prep.
What do you say when the answer is defensive anyway?
Repeat the permission once and narrow to a single artifact. Something like: that is genuinely fine either way, so take the last thing you shipped and tell me which part the assistant wrote first. A defensive answer is nearly always a general answer, and asking for one specific piece of work gives the person somewhere concrete to stand and gives you something checkable.
Three recoveries cover most of what goes wrong:
1. The denial. Someone says they do not really use it. Ask what they use instead and how they check it. A candidate working under a tool ban or on air-gapped systems can have excellent verification habits and no model in the story, and a strong candidate who says they do not use AI is a judgment call rather than a disqualification. 2. The compliance answer. Someone says they always check everything. Ask what they found the last time they checked. If nothing comes back, the routine is aspirational, and knowing that is worth the question. 3. The over-disclosure. Someone says they used it for all of it. Ask what they kept, what they rewrote and what they would not have sent. Volume of use is not the thing being assessed, and saying so out loud usually settles the answer down.
Keep the wording identical for every candidate in the role, and put the script in the interview guide so it does not depend on whoever runs the round. Two interviewers improvising two different framings produce answers that cannot be compared, which is the same failure as asking two different questions. What you are allowed to ask has its own boundaries, and whether you can legally ask candidates how they use AI covers the parts that vary by jurisdiction and by whether the round is recorded. Check the specifics with counsel before you record anything.
Common questions
Does the script change for a phone screen?
Shorten it, keep the permission clause. On a fifteen-minute screen there is no room for the reason and the reassurance and the question, so say the essential half: you can use AI in this role, I am going to ask how, and it will not count against you. The question that follows should ask for one moment rather than a process, since a screen has time for a single specific example and nothing else. The full version belongs in the round where the work is discussed.
What if the company has not decided its own AI rule yet?
Say what is true, which is that the rule is being worked out and the interview is not the enforcement of one. That sentence is more credible than a policy invented on the spot, and candidates handle honesty about an unsettled rule better than a confident answer that turns out to be wrong in week two. Then ask the choice questions anyway, since what someone decided in their last job does not depend on your policy existing yet.
Should the same script be used for junior and senior candidates?
The permission clause stays identical for everyone, and the question after it moves. Early-career candidates are often being told two contradictory things about disclosure, so the reassurance matters more, not less. What changes is what a passing answer contains: a first-job candidate can reasonably describe checking a draft against a source, where someone senior should be able to describe what they refused to delegate and what it would have cost. Same frame, different bar.
Does saying AI is allowed invite candidates to over-rely on it during the interview?
Sometimes, and that is informative rather than a problem. Someone who hands the whole question to a tool in front of you has answered a question about delegation boundaries. What matters is that the rule was stated first, so nothing here is a misunderstanding and what is in front of you is a choice. If the round involves live work, say plainly what is in bounds for this specific exercise, since a general permission and an exercise-specific rule are different statements and candidates cannot guess which one applies.
Do candidates believe the promise that disclosure carries no penalty?
More often when it is specific and attached to a reason. A bare reassurance sounds like something an interviewer is required to say, while naming what you will do with the answer makes it checkable: I am writing down what you decided, a person reads it, and the notes go into the debrief. Behaviour is the stronger signal anyway. If the follow-up questions are about choices rather than about permission, the promise gets confirmed within two minutes of being made.
References
- 1. Rejected by an AI? Comparing job applicants' fairness perceptions of artificial intelligence and humans in personnel selection frontiersin.org Supports the claim that an unexplained process is the worst available version, and that attaching a reason closes most of the gap.
- 2. Expecting Too Much, Getting Too Little: Exploring the Challenges and Design Opportunities of Asynchronous AI Interviewers arxiv.org Supports the claim that a framing mismatch damages agency and trust and pushes candidates toward workarounds.
- 3. Americans' views on use of AI in hiring (chapter of 'AI in Hiring and Evaluating Workers: What Americans Think') pewresearch.org Supports the claim that public objection concentrates on who makes the decision rather than on AI appearing anywhere in the process.
- 4. Robots are judging me: Perceived fairness of algorithmic recruitment tools pmc.ncbi.nlm.nih.gov Supports the claim that candidates object to losing influence over the outcome, which is why a question about their own choices helps.
4 sources, numbered by first appearance. How Olive sources claims
General guidance for hiring teams. What works at one company and one volume may not transfer to yours.
Olive assesses how a person works with AI. It does not detect AI-written documents, and it never produces a score, a ranking, or a match percentage for a person. Candidates read the same report the employer reads.