An AI Interview Won't Verify Your Resume. It Will Ask.

What does an AI interview actually check?

Nothing outside the conversation, on the accounts the vendors publish. A live AI interviewer reads what you submitted and uses it to decide what to ask and how hard to push, which is a different job from confirming that any of it happened. Mercor's engineering blog says its interviewer is personalised before the session starts by processing the candidate's resume, and that this shapes both the questions and their depth ¹.

That post describes around 10,000 such conversations a day, each fifteen minutes long, with coding problems generated from the profile and the live exchange ¹. The volumes are self-reported, and they describe Mercor's own contractor marketplace rather than a typical employer's pipeline. The design claim is the part that transfers: the resume sets the agenda.

Alex, a live AI interviewer sold to recruiters, describes the other half of that loop: follow-up questions built from the answers you give ².

Alex assembled that page from selected testimonials, so read the mechanics as the vendor's account of its own product. HireVue draws a boundary from a third direction. Its 2025 explainability statement says the AI relies only on what the candidate says, using no video analysis and no assessment of facial expression, body language or tone of voice ³.

None of the three descriptions contains a verification step.

Mercor, Alex and HireVue are each describing their own systems, which is not an audit, and no published description covers every employer's configuration. What is consistent across the three is the shape, and the shape is the answer. These systems elicit.

So the honest answer to whether the screen catches an exaggeration is that it mostly will not. That is the least interesting part of the question. A screen that cannot verify a claim can still make you say more about it than you meant to, on a record that outlasts the conversation. What happens to an inflated line is not that a machine flags it. The line gets elaborated, in your own words, into a paragraph a person can check later.

Why does a follow-up question make faking more likely?

The published abstract does not say why, but a plausible reason is that each follow-up tells you more about what the interviewer is looking for. Levashina and Campion ran an experiment on ways to reduce faking in structured interviews, and one finding runs against the intuition: follow-up questioning increased faking rather than suppressing it ⁴.

The same experiment found past-behaviour questions harder to fake than situational ones ⁴. Take the whole result with its limits. This was 2007, in person, with undergraduate participants, and the published abstract carries no effect sizes. Whether it transfers to a transcript where an AI generated the questions is not something that study tested. But the proposed mechanism is not exotic.

Each probe can narrow the target, and the claim can grow to fit.

The same paper measured how common this already is. Its faking scale was built and validated across six studies of 1,346 people, and the authors report that over 90% of undergraduate job candidates fake during employment interviews ⁵. The share admitting behaviour closer to lying was lower, running from 28% to 75% depending on the behaviour asked about ⁵. Those are self-reports from students, on a scale whose mild end covers tailoring and embellishing.

Most people shade, and fewer admit to the behaviours closer to lying.

So a resume line reading "led a team of six" where the honest version is "coordinated three contractors" is a familiar kind of stretch, though that paper measured interviews and not resumes. Familiar does not mean harmless. It means the unwinding is a routine move, not a confession.

The screen is not where that gets settled, which is why the next question matters more.

Who reads the transcript after the screen ends?

A person does, after the session ends, on the accounts vendors publish. A candidate testimonial on Alex's own page describes the answers going to a human recruiter who reaches out if there is interest ². Whatever the AI checked or did not check, what it produced is a record, and that record can sit beside everything else an employer learns about you.

Two beliefs about that person both lack support, and they point in opposite directions. Roulin, Bangerter and Levashina ran five studies with 246 professional and 270 novice interviewers, coding impression management from video in real time. Attempts to detect it were often unsuccessful. Interviewers read honest impression management better than deceptive impression management, and experience made no difference to detection ⁶. The fear that a seasoned interviewer will see straight through you has no support in that work.

Arseneault and Roulin close the other door. They had 582 people across ten countries complete an eight-question asynchronous video interview for a mock role. Honest tactics such as self-promotion were positively associated with rated performance. Deceptive ones such as extensive image creation were negatively associated with it ⁷.

That was a mock interview with human raters and self-reported tactics, so it shows an association and not a scoring rule.

The checks that reach outside the conversation come later, and none of them are secret. Employers are told to ask for an artifact the work already produced, and to run three follow-up questions that expose whether someone understood the answer they gave. Reading how a hiring manager verifies 'uses AI daily' on a resume in ten minutes shows the kind of check a later round can bring.

Neither belief has the evidence behind it.

Nobody is reading you like a polygraph, and nothing about that makes an invented claim a safe bet.

Say the accurate version inside the answer

Give the true scope, number or title in the first sentence of the answer, then spend the rest of it on detail the true version can carry. If the resume says you led a team of six and you coordinated three contractors, say you coordinated three contractors on a named project, and describe what coordinating them took.

There is no need to announce the correction or apologise for the resume. The accurate answer is simply the answer. If someone asks why it differs from the resume, the plain reply is that the resume line overstated it.

A corrected claim costs a sentence, and an inconsistency found later can cost the process.

Two habits make this harder than it needs to be. Rehearsing a defence of the inflated line commits you to it before anyone has asked. Adding specifics to prop it up gives the transcript more surface for a later round to contradict, which is the opposite of what the extra detail was for.

• Fix the line on the resume before the next application, because the file is what the next screen reads.
• Where a number was never measured, describe the scope instead, the same advice given on whether to let AI add numbers and percentages to your resume.
• Where a title overstates the level, use the accurate title and let the described work carry the weight.

Doing this in the AI screen rather than in the human round has a practical advantage. The screen is short and often the first conversation, so a correction there is a small edit to a record no person has read yet. The same correction three rounds later, after a hiring manager has repeated your line to a panel, arrives as news that changes what they already decided. The cost is not the correction. It is how much has been built on top of the version you did not correct.

The claim gets cheaper to fix the earlier you touch it.

The point here is timing, and where records go. The resume line is a promise about what you can talk about for twenty minutes, and the fix is to make the promise one you can keep.

See how it works >