Interviewing
The Number You Can Explain Beats the Number You Cannot
Estimate when you can say out loud how you got there, and leave the number out when you can't. "About forty a week, because I took two a shift across five shifts" is an estimate with its method attached. "30% faster" with nothing behind it is a guess wearing a percent sign. Where the improvement was never measured, give the size of the work instead: how many, how often, over how long.
The takeMost pages answering this question are lists of ways to manufacture a figure, which is coaching toward a claim you can't defend in the follow-up. An unmeasured number is the part of an interview answer most likely to be tested and fail. Most of the rest is judgment, and judgment gets probed rather than checked. So spend your precision where it is real, and stay honestly vague where it isn't.
Where Olive fits
Open a role and see what the work shows
An Olive assessment is not an interview, and nothing drawn from how you sound or look enters the findings: you work a real assignment with an AI assistant, think aloud or type, and a person writes or edits the findings on what happened, in words rather than a number. You are granted the same report the employer reads, free.
Assess your shortlistShould you estimate, or leave the number out?
Estimate when you can reconstruct the arithmetic out loud, and leave the number out when you can't. That is the whole test, and it is quick to run. Ask what you would say if the next question is "how did you get to that." If the answer is made of things you actually know, the estimate is yours to give.
A defensible estimate carries its own method. "Roughly forty tickets a week, because I took about two a shift and worked five shifts" is a number and a receipt in one sentence.
An invented percentage carries nothing.
The distinction is structural rather than moral. An estimate is built out of facts you can still name a year later: the unit you counted, the period, the rate. A manufactured figure got chosen because it sounded like the kind of number the sentence needed, and there is nothing underneath it to name when someone asks.
Precision is itself a claim.
Say "about" and "roughly" out loud when you mean them. The hedge is not a weakness in the answer. It is the word that tells a listener which kind of number this is.
The same test applies before anyone interviews you, including to whether you should let AI add numbers and percentages to your resume, where a percentage arrives inside the bullet without anyone having measured anything.
What separates an estimate from an invention?
The method separates them, and a follow-up question is what reaches for it. An estimate comes apart and goes back together: here is the unit, here is the period, here is how you knew the rate. An invention has no parts. Asked a second time it either stays perfectly round or quietly becomes a different number.
Among the undergraduates one study asked, some shading was close to universal. Levashina and Campion built and validated a scale for interview faking behavior. They report that over 90% of undergraduate job candidates fake during employment interviews, with fewer engaging in behavior semantically closer to lying, a range of 28% to 75% depending on the behavior 1. Their "faking" runs from tailoring what you emphasize all the way to inventing, the numbers are self-reported, and the interviews were in person in 2007.
Tailoring is the norm; inventing is less common.
Choosing which of your true examples to lead with is tailoring. Attaching a figure nobody measured to one of them is a different act, and it is the part of the answer most likely to be checked and fail. Much of the rest is judgment, and judgment gets probed rather than verified.
Arseneault and Roulin had 582 participants across ten countries answer eight questions in a mock asynchronous video interview for a bank manager role 2. Self-reported honest tactics such as self-promotion were positively associated with rated interview performance, and deceptive tactics such as extensive image creation were negatively associated with it 2. Nothing was at stake in that study, the tactic use was self-reported, and human raters did the rating, so it is an association rather than proof that any scoring model punishes invention.
Real strengths, clearly stated, is the whole tactic.
Say the size of the work, not the improvement to it
When the improvement was never measured, describe the work instead: how many, how often, over how long. Those three are almost always true, and you can usually check them against an old calendar or a mailbox before the interview. "Two hundred accounts, reviewed monthly, for three years" tells a listener more about the scale of a job than a percentage nobody recorded.
Three things survive a follow-up because you can rebuild them from memory:
- Scope. How many of the thing, and how big each one was.
- Frequency. Daily, per sprint, every month-end close.
- Duration. How long you did it, and what changed while you did.
An interviewer who wants a number will ask for one, and "I never had a dashboard for that, but it was around two hundred accounts" is a complete answer. It is specific about the thing you can be specific about, and it puts no claim in the room that you cannot back.
A ladder written for interviewers shows where that kind of figure gets reached. What follow-up questions expose whether someone understands the answer they gave is written for the person on the other side of the table, and its first rung asks which alternative you ruled out and what number or constraint ruled it out. A figure with no method behind it is what that question lands on.
Scope is checkable, and an invented percentage is only repeatable.
What happens when someone asks where the number came from?
You get asked to rebuild it, usually in one short follow-up, and the answer either has parts or it does not. That is the moment the two kinds of number stop looking alike. A reconstructable estimate turns into a small story about shifts and counts. An invented one turns into a pause, and then a softer version of the same claim.
What a reviewer writes down about you is not nothing. Text-mining the post-interview notes on 7,650 candidates hired at one large Chinese internet technology company, Liu and colleagues looked at what interviewers had written 4. The number of job-related capabilities an interviewer named in the notes was positively related to later performance and promotions, and negatively related to turnover 4.
The effect is small, roughly 2 percent of a performance measure per standard deviation of how closely the notes matched the job analysis 4. It is correlational, at one firm, and only candidates who passed the interview could be observed at all, so it says nothing about how the notes would have read for anyone rejected.
Take the mechanism from it and leave the effect size.
What carried signal there was a named capability rather than a rating. The study did not look at whether a figure in the notes carried anything. A number you cannot source, though, gives a reviewer nothing useful to write down, because the moment it gets tested the note becomes about the number instead of about you. A true specific gives them a sentence.
On HireVue's AI-scored one-way interview, by the vendor's own account, a number is one more word in a transcript. HireVue's 2025 explainability statement says its AI relies only on what a candidate says, with no video analysis and no other audio characteristics, so facial expressions and tone of voice are not assessed 3. That is a vendor describing its own product rather than an audit of it, it covers that product alone, and the recording still exists for a person to watch.
If you want the same ground from the hiring side, how an employer verifies "uses AI daily" in ten minutes shows what a short check actually asks for: an artifact the work already produced, and one output that got thrown away. Neither of those needs a percentage attached.
Common questions
Is estimating in an interview the same as lying?
Not if you can reconstruct it and you say it is an estimate. "About forty a week, because I took roughly two a shift over five shifts" is arithmetic with the working shown. What crosses the line is a specific figure with no method behind it, offered as though something measured it. The tell is the follow-up: an estimate has parts you can name again a year later, and an invented number has none.
What do I say when the job never produced any metrics?
Give the scope of the work instead of an improvement to it. How many of the thing, how often, over how long. "Two hundred accounts, reviewed monthly, for three years" answers "how big was this" and needs no dashboard to defend. Then spend the rest of the answer on the decision you made and why, which is usually what the question was reaching for anyway.
Should I round a real number up?
Round to the precision you can defend, and no further. If you counted 38 and say "about forty," the method still holds and the word "about" is doing honest work. If you counted 38 and say 52 because it sounds stronger, there is nothing left to reconstruct when someone asks. A figure with a decimal point tells a listener it was measured, so use one only where something actually measured it.
Does an AI interview read a hedge differently from a hard number?
For at least one vendor, both are words in the transcript. HireVue's 2025 explainability statement says its AI relies only on what a candidate says and uses no video analysis or other audio characteristics, so facial expressions and tone of voice are not assessed. That is the vendor's account of its own system rather than an independent audit, and it covers that product only. The statement says nothing about whether the model weighs "about" differently from an exact figure. The recording still exists, and a recruiter may still watch it.
Doesn't almost everyone embellish anyway?
Most people shade, fewer invent, and the difference is where the risk sits. Levashina and Campion report that over 90% of undergraduate job candidates fake during employment interviews on their scale, with between 28% and 75% admitting the behaviors closer to lying, depending on the behavior 1. Their "faking" includes mild things like choosing which strength to lead with. The self-reports came from undergraduates in 2007, in person, before any AI scoring.
References
- 1. Measuring faking in the employment interview: development and validation of an interview faking behavior scale. pubmed.ncbi.nlm.nih.gov Supports the article's claim that mild shading is close to universal while inventing is far rarer, with the self-report and undergraduate-sample hedges stated in the same paragraph.
- 2. Investigating impression management use in asynchronous video interviews across 10 countries doi.org Supports the claim that honest self-presentation was associated with higher ratings in a mock one-way video interview and deceptive tactics with lower ones, hedged as an association in a study with no job at stake.
- 3. 2025 Explainability Statement hirevue.com Supports the claim that on one vendor's AI-scored interview the score comes from the transcribed words alone, cited as the vendor's own published description rather than an audit.
- 4. 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 claim that named job capabilities in an interviewer's notes carried signal about later outcomes, with the small effect size and range-restricted sample stated alongside it.
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.