Why AI Interviews are creating an arms race between candidates and hiring companies

By Sumit Rai, Co-founder and CEO at Skima.ai

12 October 2026 - In the past two years, job interviews have become a technical problem, thanks to the widespread adoption of AI. Candidates can now sit in the comfort of their couches, speaking to AI interviewers, in their own time. While AI interviews have massive pros in terms of managing applicant volume and employee convenience, it’s becoming increasingly difficult to avoid the cons.

An analysis of more than 19,000 AI-run interviews between July 2025 and January 2026 found that about two out five candidates were guilty of AI-assisted cheating. For technical roles specifically, that number spiked up to 48% - which means nearly half of the people interviewing for technical roles are not who they pretend to be. More recently, the North Korean hiring fraud came into light, where the accused planted fake IT workers inside Western companies. This scam was exposed from the inside, after one of its operatives tried to infiltrate the very firm that tracks the fraud.

How do AI interviews work?

An AI interview is a job screening where a computer program asks you questions, listens to your answers, and scores how well you fit the job instead of a human. To understand the root causes of frauds, it’s imperative that we first understand how AI interviews work from a technical standpoint as well. Most interviewing tools in the market run on one of a few formats: a) a conversational interview designed to cut through noise from a very large applicant database, or b) a mixture of skills and conversational interview where the applicants get access to AI-assistants who can guide them if they’re stuck during the process.

For conversational interviews, recruiters normally have a choice between asking a preset list of questions, or opt for a more adaptive method where they can just upload the job description and ask their tool to come up with questions itself. There’s another conversational method of interviewing where the AI can ask follow up questions based on the candidate's answer. The method recruiters choose has no impact on the number of fraudulent cases - they vastly depend on how exactly the candidate decides to cheat.

To avoid that, a proctoring system, or in layman's terms, a security system is set pre and post the interview which detects abnormalities in candidates’ behaviour or technical set-up. At the end of the interview, a candidate is assigned a score - positive ones for each ‘right’ answer, and negative ones, which are attributed to each abnormality. The ones that clear the threshold move to the next round.

The need for a robust proctoring system

Fraud in AI interviews rarely comes down to one trick. It is a collection of small workarounds, each aimed at a different layer of the proctoring system, which is precisely why the system has to work across all of them at once.

Take eye contact. Certain tools now simulate a natural, sustained gaze at the camera, which is why proctoring tracks gaze duration and flags anything that holds a beat too long to be genuine. A dual monitor setup, where a candidate reads answers off a second screen while appearing to look straight ahead, gets caught the same way tab and window switches do: logged the instant they happen, regardless of how brief.

Some candidates have someone else sit the interview altogether, so verification now runs on two fronts: an identity document uploaded before the call, and facial recognition checked against it both at the start and the end, ensuring the person who began the interview is the one who finished it. A VPN, used to mask location or route around geographic restrictions, shows up as an inconsistency in the connection itself.

Then there is what happens just out of frame. A phone or a book kept off camera to feed answers gets picked up by video detection built to spot exactly that. A second person in the room, whether prompting quietly or simply present, registers as multiple faces or an extra voice, and audio monitoring adds unusual sound levels to that same picture. Camera tampering, whether through lighting, angle or a strategically placed obstruction, gets caught by quality checks run every couple of minutes, watching for the kind of drop in clarity that tends to mean something is being deliberately obscured.

Collateral damage caused to candidates and businesses

The part of this story that gets the least attention is what happens to the candidates who are not cheating. As detection tightens, it becomes more suspicious of everyone; a candidate who glances away to think, loses a beat to a slow connection, or simply answers in an unusual rhythm can trip the same flags as someone running a script, and genuine nervousness starts to look a lot like evasion. This tends to fall hardest on candidates with the least room to absorb it, such as first-generation graduates unfamiliar with the format, people interviewing in a second language, or those who simply don't perform confidence well on camera. It's little surprise, then, that only a quarter of candidates trust AI interviews to judge them fairly, despite two-thirds having sat one.

Businesses carry a cost too. A candidate wrongly flagged and rejected doesn't come back to explain why they left, but they do talk about it, on Glassdoor, on LinkedIn, to their own network, and a system that feels invasive or unfair leaves a mark on employer brand long before it shows up in a hiring report. That matters more than most companies realise: a weak employer brand alone can push cost per hire up by as much as 50%. Every additional layer of proctoring adds cost, but weighed against a human reviewing every candidate at round one, it remains the cheaper option by a wide margin. The real expense lies in what companies never see: the good candidates lost to a trigger-happy system, and the reputation cost of a bad experience compounding in the background.

At the end of the day, even the most robust system will get some calls wrong. That is precisely why the threshold for escalation matters as much as the detection itself. High-volume roles need AI to manage sheer scale, but borderline scores, and aspects of recruiting that should involve critical judgement, deserve a human recruiter looking closely before a rejection goes out. Companies that get this balance right will spend less time chasing an arms race they cannot fully win, and more time actually finding out who they are hiring.