Policy
Write the Retake Rule Before the First Person Asks
Allow retakes on an AI exercise, but decide the rule before anyone asks for one and print it in the invite. One retake on request, open to everyone who sat that exercise, is the version most teams can defend. The second attempt uses a different form of the same task, the later result stands on its own, and the fact of a retake never enters the file as a mark. Granting them case by case is what creates the disparity, because who asks is not random.
The takeThe expensive thing here is not the second attempt. The discretion is. A team deciding retakes one email at a time has built a benefit that flows to candidates who assume hiring rules bend, and those candidates are not spread evenly across a pipeline. Publishing a rule costs one sentence in the invite and removes the judgment call entirely. Any team that cannot imagine granting a second attempt to everyone should ask what the first sitting was really worth, given how far two sittings of the same questions can drift apart.
Where Olive fits
Open a role and see what the work shows
A second sitting creates a second record, and the question a month later is what each one rests on. Olive returns six findings per session, each written or edited by a person and anchored to a timestamped excerpt, and every released report exports with its rubric, scorer and bank versions attached.
Assess your shortlistWhat do you tell the candidate who already asked?
Tell them yes or no today, and name the rule you are applying. If no rule exists yet, the honest sentence is that the team is settling one now and will apply it to everyone in this round, including them. What you cannot do cheaply is treat a single request as a judgment about one person, because the answer becomes the precedent and it will be quoted back.
Ask what failed before you answer, because two of the three causes worth checking belong to the employer. An exercise nobody had sat before it started deciding careers. A clock set to a length the job never asks for, which leaves you with what a timer actually measures rather than judgment.
The third cause is the candidate, and only that one is a finding about the person.
The first two are catchable in advance, and cost nothing once caught. Piloting the exercise before it becomes a hiring gate is the version of this conversation that happens once, with people whose job is not on the line, instead of once per rejected finalist.
Which of the three retake rules can you defend?
Three rules are defensible here. No retakes for anyone. One retake on request, open to everyone who sat the exercise. Or a retake only where something outside the candidate's control spoiled the session, with the cause written down at the time.
The middle rule is right for most teams, and it costs less than the argument about it suggests.
Some video-interview vendors already treat attempts as a setting, at a smaller grain: a take there is one recorded answer, not a whole second sitting. On Spark Hire, the number of times a candidate may re-record an answer is set per question and chosen while the question set is being built, from a menu running 1 to 5 attempts or unlimited 2.
Karat publishes a four-step procedure for a whole second interview. A redo is requested within 24 hours of the original interview, the second sitting uses a new set of questions, and both recommendations go to the client, which decides 1.
The redo is a rule with a deadline, written down before anyone needs it, and offered to candidates of Karat's clients rather than the market at large.
Willo's guidance to employers is blunter. Give at least one retake, since a zero-retake setting can sometimes produce rushed or low-quality answers, say so in the assessment description if the number is limited, and weigh what a limit does to the many neurodiverse candidates who will need additional retakes to finish 3.
Price the middle rule before adopting it: one more sitting to grade for every candidate who fails and asks. If most of the round is failing, retakes are the wrong instrument anyway, and a pass rate that low is a question about the bar, which moves for evidence and never for yield.
One person has to own the answer for the role. A retake granted by one hiring manager and refused by another for the same job is one role carrying two different bars, and the candidate on the wrong side of it has a fair question.
Why does deciding one request at a time create the disparity?
Because the people who ask are not a random sample of the people who could have. A retake granted on request, with no published rule, is a benefit that reaches whoever assumed the process was negotiable. Confidence, coaching and having once been told that hiring rules bend are spread unevenly through any pipeline, and none of the three is what the exercise was built to measure.
One field experiment measured this for a different ask. Between November 2011 and February 2012, the economists Andreas Leibbrandt and John List posted 18 real administrative-assistant job ads across nine US cities, then randomly told half of the interested job seekers that the wage was negotiable 4.
Where nothing was said about negotiating, 9.4% of applicants opened a negotiation. Where the sentence appeared, 23.1% did 4.
The gap between men and women moved with it. Under the silent version women negotiated less often than men, 8.2% against 10.6%. Under the explicit version women asked slightly more often, 23.9% against 22% 4.
The authors decline to claim that reversal, and say so in the paper: across the 863 job seekers who went on to apply, the flip is not firm enough to assert.
They do say the data rules out the opposite: that women ask less often than men once the rule has been stated 4.
A wage is not a retake, and no study cited here ran this experiment on a hiring exercise. What carries over is the mechanism the study isolated: one sentence, saying the thing was available, changed who asked for it.
Silence is not a neutral policy: silence selects for the candidates who assume a process bends.
Design the second attempt so the result means something
Give the second attempt a different form of the same task, and let the later result stand on its own. Reusing the identical exercise measures recall of the first one. Averaging the two drags the second result back toward a sitting the rule already agreed to set aside.
A retake in the file should record what happened, never that the person needed one.
Different form, same task, same bar. Swapping the dataset, the seeded defect or the source document keeps the exercise comparable, which is the general rule that says vary the data, never the task or the bar.
How much a single sitting is worth has been measured for automated personality scores from recorded interviews rather than in AI work. Ninety-nine Purdue students recorded the same six interview questions twice, on average 15.6 days apart, and research models scored both recordings 5.
The two sittings lined up at about .46 to .50, which leaves roughly a quarter of the movement between people holding steady across the fortnight 5.
The rest was the occasion, the performance and noise.
Students in a lab are not applicants with a job at stake, and those were academic models rather than any vendor's, so the figure does not transfer to your exercise. It is a reason to ask how stable yours is, not an answer.
Practice is the standard objection. The study cited here measured coaching rather than a second sitting, and it points the other way. Maurer, Solamon and Lippstreu compared coached and uncoached candidates on a situational panel interview taken by 146 public safety employees, and found that scores from the coached sample predicted later performance better rather than worse 6.
The coaching studied was content-focused, aimed at helping people convey what the interview was asking about, and the two samples were not randomly assigned.
A second sitting costs a reviewer time that no other candidate's file costs, and the case the first attempt burned may have no second form written yet. Capacity is the objection that holds. A team with one exercise and no time to author another is choosing between no retakes and an unfair one.
When a retake request is really an accommodation request
A candidate does not have to say the word accommodation for the request to be one. Under the guidance the EEOC issued in 2022 on algorithmic assessments, an applicant who says a medical condition made a test hard to take has already requested a reasonable accommodation, without mentioning the ADA and without using the phrase 7.
At that point the email leaves the retake policy and enters a different process.
The same document names an alternative version of the test, alongside extended time, among the accommodations that may work 7.
A second attempt on a different form is close to that shape, so a published retake rule may meet some needs before anyone has to disclose anything. It does not replace the accommodation process once a request has been made.
Naming an accommodation does not make it automatic, and the guidance now reads only in a web archive: the agency took its two AI documents off its own site in January 2025 7.
Employment counsel is the right reader for what binds you, in your jurisdiction, this quarter.
The operational failure is smaller and more common than the legal one. A request counts whether it arrives to a recruiter, a coordinator or the vendor running the test 7, so a process that only recognises a request when it comes through one form has already missed some.
Extra time is the more familiar version of this ask. How much of it is reasonable, and whether it makes results incomparable, is a question about judging the work rather than the rate. The retake version is easier to run.
A different form of the same task, graded against the same bar, with nothing in the record about why the second sitting happened.
Answer the email in front of you today, with the rule rather than a verdict on the person who sent it. Then put that sentence in the invite going out next week, where every candidate reads it at once.
The next person to bomb the exercise should not have to guess whether asking is allowed.
Common questions
Does a retake have to be offered to every candidate?
To everyone who sat the same exercise in the same round, yes, if the rule says retakes exist. The exposure is not the second attempt but the inconsistency: a benefit given to the candidate who asked and withheld from the one who did not know to ask is two different processes run under one job title. If a rule is adopted mid-round, extend it backwards to anyone already rejected on that exercise, or accept that the round ran on two standards and record which candidates got which.
Should the second attempt use the same exercise?
No. A repeat of the identical task measures memory of the first sitting, and the result stops being about the job. Change the data, the seeded error or the source material, and hold the task and the bar constant. If no second form exists, that is the real constraint on the policy, and it is cheaper to write one alternate case now than to argue about fairness per candidate later. Reusing the same case with a different clock is not a second form.
Doesn't a second attempt just measure practice?
Partly, and one study of preparation suggests that need not be the problem it sounds like. In a study of a situational panel interview taken by 146 public safety employees, scores from candidates who had been coached on the interview's content predicted later job performance better than scores from uncoached candidates. Preparation improved what the number was worth. The design compared two samples rather than assigning people at random, the coaching was content-focused, and coaching is not a second sitting, so it says nothing about tactics aimed at polish or about retakes directly.
What if the tool broke or the session was cut off?
Grant the retake and log the cause the same day. A spoiled session is the one carve-out that even a no-retakes policy should carry, because the first result measured the outage rather than the person. Write down what failed, who reported it and when, and treat that log as the record of why this candidate sat twice. Keep the standard the same as any other retake: a different form, the later result standing on its own.
Should the hiring manager be told a retake happened?
Decide that when the rule is written, not when the file lands. At least one vendor publishes both results: Karat sends the client both recommendations and leaves the advancing decision to the employer. That is honest and it carries a cost, because a reader who sees two results may weight the first one unless the rule says otherwise. The workable version states which result stands, gives the reviewer only that one, and keeps the reason for the second sitting out of the scorecard.
What if a candidate asks for a third attempt?
The published number answers it, which is the main reason to publish a number. One retake means one. A request beyond the stated limit that names a disability or a medical condition is not a policy question at all; it moves into the accommodation process, where the answer depends on what the exercise is measuring and what alternative would still measure it. Route those to whoever owns accommodations, and keep the retake rule for everything else.
References
- 1. Redo Interviews ✓ connect.karat.com Supports the published four-step redo procedure cited in the article: a request within 24 hours of the original interview, a second sitting with a new set of questions, and both recommendations published to the client, which decides whether to advance the candidate.
- 2. Add a limit for how many times a candidate can re-record ✓ support.sparkhire.com Supports the claim that the number of attempts is a setting chosen when the question set is built, from a menu of 1 to 5 takes or unlimited. The page states the range of the setting, not a default or what employers typically choose.
- 3. Number of retakes ✓ support.willo.video Supports the vendor advice quoted in the article: give at least one retake, tell candidates upfront in the assessment description if retakes are limited, and the statement that many neurodiverse candidates require additional retakes to complete an assessment.
- 4. Do Women Avoid Salary Negotiations? Evidence from a Large Scale Natural Field Experiment ✓ nber.org Primary source opened for this article. Supports the 18 administrative-assistant ads across nine US metropolitan areas posted November 2011 to February 2012, the 2,422 job seekers, the 9.4% and 23.1% negotiation rates, and the 8.2% against 10.6% and 23.9% against 22% gender splits. The abstract states: 'when there is no explicit statement that wages are negotiable, men are more likely to negotiate than women. However, when we explicitly mention the possibility that wages are negotiable, this difference disappears, and even tends to reverse.' On the reversal the paper says: 'we cannot say with confidence that women are more likely than men to initiate wage negotiations but we can clearly reject the opposite.'
- 5. Automated Video Interview Personality Assessments: Reliability, Validity, and Generalizability Investigations ✓ par.nsf.gov Supports the retest measurement in the article: 99 Purdue students recorded the same six interview questions twice, on average 15.6 days apart, with mean test-retest correlations of .46 for self-report-trained models and .50 for interviewer-report-trained models. Academic models on students, not any vendor's scoring.
- 6. How does coaching interviewees affect the validity of a structured interview? ✓ doi.org Supports the claim that preparation did not degrade what the interview was worth: in a combined sample of 146 public safety job incumbents sitting a situational panel interview, predictive validity and reliability were higher in the coached sample than in the uncoached sample. Content-focused coaching, two samples rather than random assignment.
- 7. The Americans with Disabilities Act and the Use of Software, Algorithms, and Artificial Intelligence to Assess Job Applicants and Employees ✓ web.archive.org Supports three claims in the article: that saying a medical condition may make a test difficult already is a request for reasonable accommodation with no need to mention the ADA, that extended time and an alternative version of the test are named as accommodations that may be effective, and that the document is archived after the EEOC removed its AI guidance in January 2025.
7 sources, numbered by first appearance. Every one was opened and checked against the claim it carries. How Olive sources claims
General guidance, not legal advice. Hiring rules differ by state and country and change often; check anything here against your own counsel before you act on it.
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.