AI interview anti-cheat that gives you evidence, not verdicts

Remote interviews removed the one control an in-person round had for free: someone in the room could see what the candidate was looking at. Most cheating now falls into a few familiar patterns.

Help off camera. A second person sits out of frame, or joins on another call, and feeds answers. Sometimes the person in the interview is not the person who applied.

Answers on a second screen. Notes, a prepared script, or an AI assistant running on another device or in another window, read back as if recalled.

Generated code. On a coding problem, the question is pasted into an assistant and the answer is pasted back, often correct and arriving all at once.

None of these is exotic. They are cheap, they work against a fixed list of questions, and a one-way recording cannot ask the follow-up that exposes them.

A recorded answer shows you what was submitted. A live session shows you how it was produced: the pause before an answer, a voice that is not the candidate, eyes moving to a second screen, a block of code that appears with no typing before it.

Because the AI interviewer is present for the full 45 minutes, it can also probe. A rehearsed answer gets a follow-up question that depends on what the candidate just said, and a script rarely has the next line. That follow-up is often more telling than any monitoring signal.

This is why anti-cheat belongs to the AI video interview and not to the 15-minute phone screen. A phone screen is a short first filter on fit and logistics; the video round is where depth is tested, so that is where the evidence matters.

A signal tells you that something happened at a specific moment: a second voice at 12:40, a device on screen at 31:05. It does not tell you why.

Ordinary things trigger signals too. A family member walks past. A candidate looks away to think. Someone reads from their own notes because they are nervous, not dishonest. Treating every flag as proof of cheating would reject good people, and treating flags as noise would waste the evidence.

That is the reason QuillHire never turns signals into a pass or fail. The finding sits in the interview record with its timestamp, next to the transcript and the scorecard, so a person can look at the moment and decide.

Watch the moment, not the summary. Jump to the timestamp and see what was actually happening on screen and in the room.

Compare it with the answer. A flag next to a strong, specific, well-reasoned answer means something different from a flag next to a generic one that falls apart under the follow-up.

Ask about it in the next round. If the candidate moves forward, have a person revisit the flagged topic live. Understanding survives a second conversation; a borrowed answer usually does not.

Tell candidates up front. Saying that the session is monitored, and what is monitored, discourages cheating before it starts and keeps the process fair to the candidates who were never going to cheat.

Anti-cheat is not a coding-test feature. A product manager, a marketing analyst, or an operations lead takes the same 45-minute interview with the same monitoring as a backend engineer, with no coding involved. The coding-specific checks simply switch on when the round includes a coding module.

Remote hiring for non-engineering roles is exactly where AI-assisted answers are hardest to spot by ear, so leaving those roles unmonitored would leave the biggest gap open.