What to Do When Candidates Bring AI Into the Interview

A candidate on a video interview pauses for a second or two before each answer, glances slightly to one side, and then says something considered and well organized. It sounds human. It might even sound good. What the interviewer can't see is the phone just below the camera, running a transcription tool that listens to the questions and puts talking points on the screen while the interviewer is still speaking.

This is more common than most hiring teams assume. One 2025 survey found that 52% of job seekers admitted to using AI help during at least one interview in the previous twelve months. That's the number who admitted it, so the real share is probably higher.

The natural reaction is to call it cheating and go looking for a way to catch it. I think that's the wrong frame, and it leads to the wrong fix.

Start with what the interview is for. You're trying to find out how someone thinks when they have to answer on their own. A candidate reading suggestions off a second screen will often do better in the interview than they would without help, and the improvement tells you nothing about how they'll do the job. What you end up measuring is how well someone can read an AI's suggestion and deliver it convincingly. That's a real skill, but it's rarely the one you're hiring for.

This matters more than it used to, because the interview had become the last honest step. The resume stage was already largely gamed by AI-generated applications. Teams could live with that as long as the interview sorted things out. If the interview is compromised too, then every stage of the funnel is working from information the candidate didn't produce on their own.

Why detection won't hold up

The industry's first answer has been detection software: tools that watch for unusual pauses, eye movements, and phrasing that looks machine-generated. A few enterprise ATS vendors have started building these features in.

I don't think they'll work at scale. The models that generate the answers improve faster than the tools built to spot them, so detectors trained on 2024's models were already behind by 2025. And there's a worse problem. A tool that flags odd pauses and unusual phrasing can easily end up flagging non-native speakers and neurodiverse candidates more often than everyone else, which is a legal risk no hiring team should want.

Even a perfect detector would be attacking the problem from the wrong end. It tries to catch the behavior after it happens. Candidates use AI in interviews because the format makes it easy. Nobody can see it, it rarely has consequences, and it often helps. As long as that's true, people will keep doing it, and you won't shame or police your way out of it. You have to change the format.

What makes a format hard to game

Why does AI help work so well in some interviews? It works best in text-based screens and asynchronous video, and it works in ordinary video calls too. What these have in common is a gap between hearing the question and giving the answer. That gap is where the tool does its work.

A live spoken conversation mostly closes the gap, as long as it's truly adaptive. When the interviewer picks up on something specific in your last sentence and asks you to say more about it, there's nothing queued up. The candidate who had a polished answer ready for the first question doesn't have one ready for a follow-up they didn't expect, and the lag between the tool suggesting something and the candidate saying it becomes noticeable.

That's why I think agentic recruiting matters most at the screening stage. An AI that runs a spoken conversation and actually listens, choosing its next question based on the last answer, creates the conditions under which a candidate's performance is their own.

This is how we designed Asendia. It phones candidates within hours of their applying and has a live spoken conversation with each of them: a question specific to the role, then a follow-up on something in the answer, the way a good recruiter would. Candidates who can explain themselves clearly in that setting come out as qualified. Candidates who can't produce coherent answers in real time don't make the shortlist, however polished their written application was.

What the recruiter finds in the ATS afterward is a structured qualification summary for each candidate, with their key quotes, from a conversation that actually happened. Because everyone for the role was asked the same core questions, those summaries can be read side by side. So instead of a few hundred applications of unknown quality, the recruiter starts with a short list of people who have already shown they can communicate when it counts. That helps the later stages too, since the panel is meeting candidates who were evaluated in a format that was hard to fake.

The interview was supposed to be where you find out who someone actually is. For many teams it has stopped doing that, and mostly not because candidates are dishonest. The format left an opening and the tools walked through it. Detection will always be a step behind. What works is going back to a format that makes authentic answers necessary: a real conversation, in real time, with an interviewer who adapts. At the top of the funnel you need that to happen for every applicant, without waiting on a recruiter's calendar.

Ready to transform your hiring strategy? Schedule a Demo with our founders today!