Interviewing
The AI Interview Can Be the Friendlier Gate for a Thin Resume
If your resume is the weak part of your application, the interview stage can be the friendlier gate, because it asks what you can do rather than where you did it. In the one large randomised test on record, a preprint with vendor co-authors, 54% of the AI-interviewed track passed the same blind final human interview against 34% of the resume-screened track, on wide error bars at a single employer. The catch is completion: most invited candidates never finished, and there not finishing meant not advancing.
The takeRead from the chair of someone whose resume is the weak part, the sharper risk is not the model doing the reading. It is attrition. A stage you never start cannot help you, and in the experiment that measured it, declining and withdrawing amounted to the same act. That is a different worry, and it is the one you can do something about this week.
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 openly with an AI assistant, think aloud or type, and a person writes or edits the six findings in words. The candidate is granted the same report the employer reads, free.
Assess your shortlistWhy the Interview Stage Suits a Thin Resume
The interview stage suits a thin resume because it reads an answer you gave, rather than a proxy for one. A screen reads the school, the job title and the unbroken run of dates. The experience part of that record carries little signal: across 81 independent samples, prehire work experience correlates .06 with later job performance 1.
That is a corrected correlation, so it is not an artefact of under-correction, and it describes experience measured at the point of hire, which is exactly what a screen reads 1. The same analysis puts prehire experience at .11 with training performance and .00 with turnover 1. The authors name real exceptions: experience predicts a little better in the first weeks on a job, and task-level experience does predict training performance, both from small subsets of studies. It says nothing about roles where a licence or a legally required minimum applies.
Employers say much the same about their own filters.
In a 2021 report from Harvard Business School and Accenture, 88% of surveyed employers that use a recruiting management system said it screens out qualified high-skills candidates who could do the job 2. The stated reason is that those candidates do not match the exact criteria in the job description.
For middle-skills roles that figure rose to 94% 2.
That number needs its hedge attached. The 88% is employers answering yes to any degree, with sometimes, often and always added into one number, and it is an opinion senior leaders hold about their own process, collected in an online survey in early 2020 2. It is not a rejection rate for candidates.
Set next to why people get rejected while matching every requirement, it says a rejection at the screen is weak evidence about you. Hiring teams ask the same question from their chair: how would you know if the resume screen is throwing away the wrong people.
The paper record was weak evidence before any machine read it.
What the Randomised Test Actually Measured
One randomised field experiment carries most of the direct evidence: it measured who passed a blind final human interview, and on that measure it points in the format's favour. Researchers assigned 37,000 applicants for a junior developer role to one of two pipelines: a resume screen followed by human selection, or a structured video interview scored by AI placed in front of the human stage 3.
Both tracks then met the same final human interview, and the interviewers were blind to how each candidate had been selected. 54% of the AI-assisted track passed it, against 34% of the traditional track 3.
The gap is 20 percentage points with a standard error of 12, which makes it informative about direction and imprecise about size 3. It is a preprint, not yet peer reviewed. It is one employer, one country and one job family, on a single recruitment platform.
The authors also report that the system tended to select younger applicants with less experience and fewer advanced credentials 3. That is a description of who it advanced in this experiment, not a recommendation. In the United States a selection skew by age is a legal exposure rather than a design goal.
A revised version of the same preprint adds a second, smaller experiment on the same platform: 4,323 applicants across four engineering roles 4. It reports the two effects together, 17.5 percentage points with a standard error of 8.5, and 20 with a standard error of 11.8 4. The second of those is the headline result above, recounted, and it is barely distinguishable from zero at conventional levels 4.
Two of its four authors are affiliated with the platform that ran the experiments and sells the interview, which is a direct commercial interest in the result.
What both experiments measure is passing a human interview, not job performance and not how long anyone stayed.
One platform is not a literature.
Take the Invitation Seriously, Because Most Invitations Go Unfinished
Take the invitation seriously, because in the same experiment most invitations went unfinished. Of the 23,000 candidates invited to complete the AI interview, 24% did 5. Inside that pipeline, not completing meant not advancing, so for the other three quarters the stage that might have read their work never read anything.
A 24% completion rate is not evidence that three quarters of candidates object to being interviewed by a machine. Non-completion mixes refusal with a dead link, a lost email, a job found somewhere else and a camera that would not turn on, and the study separates none of them 5.
The candidates who declined skewed slightly older and more experienced, and their later labour-market outcomes, read off LinkedIn profiles, did not differ statistically from the outcomes of those who finished 5. An absence of a detected difference in one sample is not proof that declining costs nothing.
Employers are running the same arithmetic from the other chair.
Whether it is cheaper to assess every applicant than to screen resumes is a live budget question for hiring teams, and the answer turns on what share of invited applicants actually finish. Their cost problem and your invitation are the same event, described from two sides.
So decide about the interview on purpose, rather than by letting it expire.
What One Vendor Says It Does Not Look At
Start with what HireVue, a different vendor from the platform in that experiment, says it ignores. HireVue's candidate FAQ says its assessments never look at age, race, gender, or things on your resume such as GPA or the college you attended, and that only skills and competencies are measured 6. That is a direct answer to the question a candidate with an unranked degree actually asks.
It is candidate-facing copy on the vendor's own site rather than an audit. A claim about what is not an input cannot be checked from outside. Saying age and race are not inputs is also not a claim about disparate impact, which is a separate measurement the page does not report 6.
Elsewhere on that page, the interview and assessment are described as intended first steps that, for the vast majority of jobs, help recruiters make the first cut, after which the top group moves on to person-to-person interviews before people make the hiring decisions 6. Intended is carrying weight in that sentence. It describes a design, not any particular employer's configuration of it.
Read what a platform publishes for candidates before the interview, not after the rejection.
None of this deletes the resume screen.
An employer can run both, and a stage that reads an unusual background more kindly can still sit behind an application form that may never forward you to it. The case for applying without meeting every requirement holds either way, and the interviews you are offered are the ones worth finishing.
Common questions
Does an AI interview see where I went to school?
One platform says no. HireVue's candidate FAQ says its assessments never look at age, race, gender, or resume details such as GPA or the college attended, and that only skills and competencies are measured. HireVue is describing its own design on its own site, not reporting an audit, and a statement about what is not an input cannot be verified from outside. Other platforms publish their own answers, and they are worth reading before the interview rather than after a rejection.
Is there real evidence that an AI interview helps a weak resume?
There is one large randomised field experiment. It assigned 37,000 applicants for a junior developer role either to a resume screen or to a pipeline with a structured video interview scored by AI in front of the human stage, and 54% of the AI-assisted track passed the same blind final interview against 34% of the other. The effect is 20 percentage points with a standard error of 12, from one employer on one platform, in a preprint that has not been peer reviewed, and a later version lists two authors affiliated with the platform that ran it. The authors report the system tended to advance younger applicants with fewer advanced credentials.
What happens if I don't complete the AI interview?
The employer's own process decides that, and in the one experiment on record it ended the application. Of 23,000 candidates invited to the AI interview, 24% completed it, and inside that pipeline not completing meant not advancing. The people who declined skewed slightly older and more experienced, and their later outcomes read off LinkedIn did not differ statistically from those who finished, which is an absence of a detected difference rather than proof that declining is free. Non-completion also mixes refusal with a dead link, a lost email and a job found somewhere else.
Does experience on a resume predict how well I would do the job?
Barely, on the best available evidence. A meta-analysis of 81 independent samples found prehire work experience correlates .06 with later job performance, .11 with training performance and .00 with turnover, and experience with relevant tasks or occupations did little better. Those are corrected correlations, so the near-zero result is not an artefact of under-correction. It is about experience measured at the point of hire, which is what a screen reads, and it says nothing about roles where a licence or a legal minimum qualification applies.
How should I handle a career gap or a career change in the interview?
Answer the question that was asked, name what you actually did in the period, and stop. A structured interview puts the same questions to everyone and judges the answer you give, so a short factual account of a gap costs less airtime than an apology for it and leaves room for the work you want described. You do not owe anyone a narrative arc. If the format is a recorded one-way interview, write those two or three sentences down before you record, because no interviewer is there to steer you back.
References
- 1. A meta-analysis of the criterion-related validity of prehire work experience digitalcommons.unf.edu The near-zero criterion validity of prehire work experience across 81 samples, which this article uses to say the paper record carries little signal about later performance.
- 2. Hidden Workers: Untapped Talent - How leaders can improve hiring practices to uncover missed talent pools, close skills gaps, and improve diversity hbs.edu The share of surveyed employers who said their own hiring system screens out qualified candidates not matching the exact criteria, and the response options that make up the figure.
- 3. Better Together: Quantifying the Benefits of AI-Assisted Recruitment arxiv.org The randomised field experiment's design and pass rates, its effect size and standard error, and its report that the system advanced younger applicants with fewer advanced credentials.
- 4. Quantifying the Benefits of AI-Assisted Recruitment arxiv.org The revised preprint's second, smaller randomised experiment on the same platform, both effects reported with wide standard errors, and two of the four authors' affiliation with the platform that sells the interview.
- 5. Better Together: Quantifying the Benefits of AI-Assisted Recruitment arxiv.org The completion rate among invited candidates, who the decliners were, and the fact that inside this pipeline not completing meant not advancing.
- 6. FAQ | Frequently Asked Questions for Candidates hirevue.com HireVue's candidate-facing statements that its assessments do not look at GPA or the college attended, and that the stage is intended as a first cut before person-to-person interviews.
6 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.