Can AI detect lying in interviews?

Lie detection by machine has been promised for a century, from polygraphs to voice-stress analyzers to micro-expression scanners, and the scientific record is consistent: none of these methods reliably separates deception from nervousness. Stress signals are real, but interviews make honest people stressed, which is exactly the population being scanned.

The screening industry has quietly conceded the point. Video-interview vendors that once advertised facial analysis publicly removed those features after sustained scrutiny from researchers and regulators, and the assessments that remain score what you say rather than how your face moves while saying it. Where legislation applies, several jurisdictions now require disclosure or consent before AI analysis of interview video at all.

So the direct answer to the anxious version of this question is no: the AI on the other side of your screening call is not reading your pulse through the webcam. What it does instead is more mundane and worth understanding precisely.

You are worried the machine will flag you for rounding your job title up, and it is worth knowing what the machine actually grades. This section covers what automated screening scores, and where embellishment genuinely gets caught.

Automated screeners evaluate transcribed content: whether your answer addresses the competency the question targets, whether it carries concrete detail, structure such as situation, action, and result, and keyword overlap with the role profile. Some score delivery basics like pace and filler-word rate. None of that is truth verification; an articulate fabrication scores well on every one of those axes.

Where exaggeration fails is later and lower-tech. A claimed skill collapses in the technical deep-dive, a story that shifts between the recruiter screen and the onsite gets noticed by humans comparing notes, and reference checks surface title inflation in minutes. Consider a candidate who upgrades a support role into a team lead: the screening AI passes her, and the hiring manager's third follow-up question about headcount decisions does not.

The detectability cluster covers the adjacent question, what interviewers and platforms can see about the tools you run, in the detectability and privacy answers.

A real-time interview assistant transcribes the conversation and suggests talking points grounded in the documents you gave it. It does not verify your claims, and it cannot make an invented story true; if the resume you load is embellished, the suggestions inherit the embellishment. Aaron Cao, founder of SubcueAI, frames the design boundary plainly: the product amplifies preparation, and honesty stays the candidate's job, which is why the suggestions are grounded in your own resume and the job description rather than generated from thin air.

The practical move this enables is the opposite of lying. Your real experience almost always contains better material than an invented version, and the gap is retrieval under pressure, not substance. A mock interview against an AI interviewer surfaces those true stories in advance, with follow-up questions that pressure-test them the way a human would, so the live answer comes from rehearsed memory instead of improvisation.

For permitted live contexts, the desktop app then keeps those rehearsed points within reach during the real call; how the capture and suggestion pipeline works is documented in the how-it-works answers.