Candidates Are Using AI in Live Interviews. Now What?

First the applications went synthetic - we covered that in screening AI-written applications. Now the interview itself is contested ground: candidates arrive with a second screen running a model that listens to your question and drafts their answer in real time. The employer response so far has mostly been indignation. Indignation is not a process.

What is actually happening out there?

The 2026 numbers sketch it clearly. Roughly one in five candidates admits to using AI during real-time interviews, and among those, over a third use it for suggested responses as the conversation happens. On the other side of the table, interviewer suspicion in remote-heavy fields has gone from paranoia to default: in technical hiring, large majorities of interviewers believe they have already interviewed AI-assisted candidates, and platforms that analyse remote interviews flag a striking share of sessions.

Two things follow. This is not a fringe behaviour you can shame out of existence. And it is not going away: the tools are cheap, invisible and improving monthly.

Why does detection lose?

Because every tell has an innocent twin. Glancing sideways is a second monitor or a nervous habit. A pause before answering is thoughtfulness. Polished phrasing is preparation - the thing we used to reward. Detection tools score probabilities, and probabilities wrongly accuse real people: reject a genuine candidate as a cheat and you have created the worst candidate-experience story possible, one they will tell publicly. Meanwhile the truly prepared cheater - earpiece, teleprompter cadence training - sails past. An arms race where false positives cost you more than false negatives is a race to lose gracefully.

What actually works: redesign, not surveillance

• Move judgement testing before the interview. Short, timed, situational questions at application stage - where response time is a scoring signal - separate instinct from lookup. A candidate answering ten scenario questions in ninety seconds is demonstrably not typing each into a chatbot.
• Anchor everything in their specifics. "Tell me about the Tuesday your section was double-booked" defeats generic assistance, because the model does not know their Tuesday. Follow-ups into detail - names, numbers, what they said, what happened next - expose borrowed stories in two exchanges.
• Make part of the interview doing, not describing. A five-minute role-play, a pricing decision on a real menu, a schedule to fix. Performance under observation is the one format assistance cannot fake - the logic of skills-based hiring applied to the AI era.
• Decide your AI policy per role and state it. Where the job includes AI tools, assess candidates using them well. Where it cannot, say the interview is a no-tools conversation - most candidates respect a clear rule, and the ones who do not have told you something.

The uncomfortable mirror

Employers earned some of this. Candidates adopted AI assistance in part because hiring processes became automated gauntlets - one-way videos scored by software, applications answered by nobody. If your process respects candidates - fast responses, human interviews, transparent criteria - you get more honesty back. If it is robotic, candidates feel licensed to respond in kind. Fixing your side of the arms race, as laid out in our one-way video guide, is part of fixing theirs.

The takeaway

Assume AI in the room, and design so it does not matter: timed situational screening before the interview, specifics-anchored conversation inside it, observed doing where the stakes justify it, and a stated policy instead of quiet suspicion. The employers who adapt their process will quietly out-hire the ones still trying to catch people.