Interviewing
Say Yes, Then Repair the Round You Just Changed
When a candidate asks mid-interview whether they can use AI and no policy exists, say yes out loud. Then add the sentence that keeps the round readable: use it if you want, and talk through what you accept from it and what you throw out. A refusal with no reason behind it reads as a trap and buys a worse sample. The repair happens afterwards. Write down what you said, tell the recruiter that day, and decide what the earlier candidates get.
The takeThe improvisation is not the failure here. The silence in front of it is. Somebody set this role's AI rule in one sentence, under time pressure, with a stranger waiting, because nobody with the authority to write it down had done so. Interviewers absorb that cost, and it lands in no document. So take the yes as settled, and treat the question itself as the finding: the loop had a hole in it that a candidate found before the hiring team did.
Where Olive fits
Open a role and see what the work shows
An interview holds a candidate's account of how they decide what to accept from a model, and it holds the deciding itself in comparable form only when somebody planned the round for it. Olive puts the deciding in front of them as work: a role-grounded assignment, an assistant that will do the whole thing if nobody stops it, and a human reviewer who reads what actually happened and writes or edits every finding.
Assess your shortlistWhat Do You Say in the Next Ten Seconds?
Say yes, and name what you are listening for in the same breath. Something close to this: the tool is fine, and what interests me is how you decide what to keep from it and what to throw out. Two sentences do the work. The first gives the permission, the second sets the standard, and the candidate learns that the round has not turned into a test of nerve.
A flat no costs more than it looks like it costs.
You have refused something without giving a reason, thirty minutes into a conversation the candidate is reading for signals about the job. The rest of the hour gets worse. And the refusal buys no information about the thing you were curious about, which was never whether the person can operate a chatbot.
Half of employers in one 2025 survey had no objection to candidates using AI in recruitment.
The Institute of Student Employers, a UK membership body for graduate and apprentice recruiters, reported that 49% of its surveyed employers had no problem with candidate AI use during recruitment, and that 10% had banned it or added technical measures 1. An attitude reported in a survey is not a published policy, and the use those employers described sat largely at the covering letter and CV stage.
None of that measured the live round, though only one employer in ten had banned candidate AI use or blocked it technically.
The no is defensible when the reason survives being said out loud. A round built to watch unaided reasoning, a client contract barring third-party processing, a data rule no approved model satisfies yet: each of those is a condition of the work, and each fits in one sentence.
Say the reason in the same breath as the refusal, or the candidate leaves with the version they invent themselves.
Why Does One Off-Script Minute Change the Whole Round?
Because a comparable round is built out of things held constant, and one of them just moved for a single candidate. The US Office of Personnel Management's guide defines a structured interview by three properties: the same questions in the same order, a common rating scale, and agreement in advance on what an acceptable answer looks like 2. A mid-round decision breaks the third: nobody agreed in advance what an acceptable answer looks like with the tool open.
The guide dates from September 2008 and was written for federal hiring, so it binds no private employer. The three properties are still the cheapest description of what a comparable round is made of.
An unscored stretch of an interview does not stay unscored.
In a study of 189 accounting students put through structured mock interviews, the impression an interviewer formed during the rapport-building small talk, before a single structured question, correlated .42 with that same interviewer's structured score. When a different interviewer supplied the score, the correlation fell to .25 and .24, depending on which one did the scoring 3.
Read the number small. A student sample in mock interviews is not a hiring loop, and much of the .42 is one person's impression predicting their own later rating.
What the study supports is narrow, and enough: an impression formed off-script is not sealed off from the part that gets scored, though the authors trace part of that link to competence the candidate showed early.
The US federal rules on selection procedures also reach the unscripted parts of a round. The Uniform Guidelines on Employee Selection Procedures, in force since 1978, define a selection procedure as any measure used as a basis for an employment decision, and say the term covers the full range of techniques through "informal or casual interviews and unscored application forms" 4.
The improvised minute sits inside the procedure rather than beside it.
Those Guidelines impose a validation burden only where adverse impact shows up, and they name no software at all. What they settle is the category question, which runs the opposite way from the usual assumption: a conversation is a selection procedure too.
Holding a round comparable has never meant reading from a card. The workable rule is to fix the questions and the anchors and let the follow-ups differ, and an unplanned grant of permission is not a follow-up.
Write Down What You Said Before the Debrief
Write the note before the next meeting starts, in four lines: the question as the candidate asked it, the words you gave back, what they did with the permission, and the time on the clock. A note written later from memory records the phrasing you meant, which may not be the phrasing you used. The recruiter needs all four the same day, because the question is coming at the rest of the loop next.
What goes in the note is the part with evidence behind it.
Researchers text-mined the post-interview notes on 7,650 candidates hired at one large Chinese internet technology company. The number of job-related capabilities an interviewer named in the notes tracked later performance ratings and promotions, at roughly 2 percent of a performance measure per standard deviation of match to the job analysis 5.
Three limits ride with that number: only candidates who passed and joined could be observed, the effect is small, and the setting is one firm in technology roles. What was measured is whether the notes name the capabilities the job calls for, which leaves note-taking as such untested.
Name the capability you saw.
A line saying the candidate accepted the model's first framing and moved on is something a second reader can weigh six weeks later. A line saying the candidate seemed comfortable with AI is not.
The scorecard is the other half of the problem, because it has no box for a round that changed shape at minute thirty.
Rating the AI line as though the round ran to plan hides the change, and leaving it blank loses the observation. Rate what happened and put the change in the note. Adding not observed to the scale, and counting how often it fires, is the repair for a line the round never reached.
Do the Candidates Already Interviewed Get the Same Offer?
The test for the earlier candidates is whether the exchange moved anything you scored. If the permission changed what the candidate produced, and this round is deciding, the people still live in the loop get the same option in writing before their next stage. If it changed nothing you wrote down, record that the round is not comparable and lean on the stages that are.
Rejected candidates are a different question, and the honest answer is usually no.
Reopening a closed decision on this alone implies the decision turned on the tool grant, so check the scorecards before assuming it did not.
Adding a round after some candidates have already finished asks the same question at a larger scale, and the same test answers it.
An improvised yes is still an improvisation, and mailing three earlier candidates a late option does not make the loop consistent. The loop is now consistently late.
Where the decision is close, rerunning the AI portion for everyone, or striking it from the scorecard, is the cleaner call.
Put the Line in the Loop Brief, Not in Your Head
Write one sentence per stage into the loop brief before the next candidate books, and say the AI rule out loud in the first two minutes rather than waiting to be asked. Saying it costs about fifteen seconds and removes the whole class of mid-round improvisation. The interviewer who improvised should not be the person who owns the rule, and by the end of the week they should not have to be.
A stage-by-stage answer beats a single yes or no, because the stages are asking different things.
One employer has published exactly that. Anthropic's candidate guidance, dated 10 July 2025, asks applicants to draft the written application themselves and use Claude to refine it, and encourages Claude for interview preparation. The same page rules out AI on take-home assessments and during live interviews unless the company says otherwise 6.
One company's page is not a norm.
The page names no detection method and no penalty, and it says it will be reviewed.
What the page demonstrates is narrower and more useful: the answer can be written down per stage, and shown to candidates before anybody has to ask for it.
Deciding which stages allow AI and which do not is the real exercise, and the job decides each one. A take-home built on a public dataset and a live reasoning round are not the same question.
In the same 2025 survey, 45% of employers had given applicants no guidance at all on when AI use is appropriate 1. In that UK cohort silence was common, and silence is how the question ends up with whoever happens to be in the chair.
The next candidate may ask at some ordinary minute of some ordinary round. By then the answer is a line in the brief, said in the first two minutes, identical for everyone who sits in that chair from then on.
Nobody has to improvise it twice.
Common questions
What exactly do you say when a candidate asks mid-interview whether AI is allowed?
Give the permission and the standard in one breath: the tool is fine, and what interests me is how you decide what to keep from it and what to throw out. Then run the rest of the round as planned. The phrasing matters because it moves the round off the question of whether the tool was touched. What the candidate does with what the tool hands back is the thing you wanted to see anyway. Write the two sentences down afterwards, word for word, before the debrief.
Is saying no mid-interview actually worse?
It is worse when no reason comes with it. An unexplained refusal thirty minutes into a conversation reads as a trap, and the candidate spends the remaining time managing you instead of working. A refusal with a stated reason is fine and sometimes correct: a round designed to watch unaided reasoning, a client contract barring third-party processing, a data rule no approved model meets yet. The test is whether the reason survives being said out loud in one sentence.
Does one unplanned exchange really affect how the candidate gets scored?
An unplanned exchange can reach the score, on modest evidence. In a study of 189 accounting students in structured mock interviews, the impression an interviewer formed during the rapport-building small talk that opened the interview correlated .42 with that same interviewer's later structured score. The correlation was .25 and .24 when a different interviewer did the scoring. That is a student sample in mock conditions, and it is a correlation rather than evidence about when a decision gets made. What it supports is the modest point: an off-script stretch of an interview does not stay sealed off from the scored part.
Do you have to go back to the candidates you already interviewed?
Offer the same option, in writing, to anyone still live in the loop if the permission changed what this candidate produced and the round is deciding. Rejected candidates are not reopened on this alone unless their decision turned on the tool grant, so check the scorecards before assuming it did not. Record either call rather than leaving it in a hallway conversation. If the decision between finalists is close, rerunning that portion for everyone or dropping it from the scorecard is the cleaner option.
What goes in the note afterwards?
Four lines: the question as the candidate asked it, the words you gave back, what they did with the permission, and when in the round it happened. Then name the capabilities you saw. A note saying the candidate accepted the model's first framing and moved on can be weighed by a second reader; a note saying they seemed comfortable with AI cannot. Text-mining of interviewer notes on 7,650 hired candidates at one company found that notes naming job-relevant capabilities tracked later performance, though only among people who were hired, and the effect was small.
Should the AI rule be the same at every stage?
Usually not, and at least one large employer publishes it stage by stage. Anthropic's candidate guidance, dated 10 July 2025, asks for a self-drafted application refined with Claude, encourages Claude for interview preparation, and rules out AI on take-homes and in live interviews unless it says otherwise. One company's page is not an industry norm and it names no penalty. What it shows is the format worth copying: one line per stage, written before the loop opens and read out in the first two minutes.
References
- 1. 5 trends you need to know from ISE's Recruitment Survey 2025 ✓ ise.org.uk Supports the survey figures on employer attitudes to candidate AI use in recruitment, and the share giving no guidance: 49% of these UK early-careers employers reported no problem with candidate AI use (largely at the covering letter and CV stage), 10% had banned it or added technical measures, and 45% had given applicants no guidance at all. Nothing in it measures permission in a live interview.
- 2. Structured Interviews: A Practical Guide ✓ opm.gov Supports the three properties named, the third of which a mid-round decision breaks: the same questions in the same order, a common rating scale, and agreement in advance on what an acceptable answer looks like. Cited as a September 2008 federal practice guide that binds no private employer.
- 3. Initial Evaluations in the Interview: Relationships with Subsequent Interviewer Evaluations and Employment Offers ✓ homepages.se.edu Supports the claim that an unscored stretch of an interview is not sealed off from the scored part: 189 accounting students in structured mock interviews, an impression formed during rapport building correlating .42 with the same interviewer's structured score and .25 across interviewers. Cited for direction only, not as evidence about when a decision is made.
- 4. 29 CFR Part 1607 - Uniform Guidelines on Employee Selection Procedures (1978), sections 1607.16(Q) and 1607.3(A) ✓ govinfo.gov Supports the claim that an unscripted interview exchange sits inside a selection procedure rather than outside it: the definition covers the full range of assessment techniques through informal or casual interviews and unscored application forms. Cited for definitional scope, with the validation burden attaching only where adverse impact appears.
- 5. Predictive Validity of Interviewer Post-interview Notes on Candidates' Job Outcomes: Evidence Using Text Data From a Leading Chinese IT Company ✓ frontiersin.org Supports the instruction to name capabilities in the note rather than impressions: across 7,650 hired candidates at one company, the number of job-related capabilities named in interviewer notes tracked later performance and promotions, at roughly 2 percent per standard deviation of match to the job analysis. Cited with its limits, since only hired candidates were observable.
- 6. Guidance on Candidates' AI Usage - How to collaborate with Claude during our hiring process ✓ anthropic.com Supports the stage-by-stage format recommended for the loop brief: draft the application first and refine with Claude, Claude encouraged for interview preparation, and no AI on take-home assessments or during live interviews unless indicated. Cited as one employer's own published policy, dated 10 July 2025.
6 sources, numbered by first appearance. Every one was opened and checked against the claim it carries. 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.