Two different behaviors hide behind this question, and they need different answers. Live AI assistance — hidden tools feeding answers during a video interview — is cheating. Prepared AI-generated stories — polished STAR answers rehearsed before the call — is at worst lazy interviewing. The tells, and the fixes, are separate. Here's how to run the detection honestly and redesign the interview so assistance stops mattering.
The live-assist tells
A candidate with an AI copilot whispering in real time produces a distinctive pattern. Any single sign is weak; three or more stacked is a strong signal:
- The answer latency beat. A consistent small delay after each question — a breath, then fluent, structured delivery starting at the same word count every time. Humans fumble openers; streamed answers don't.
- Gaze geometry. Eyes fixed on a point beside the camera (where a prompt window lives) rather than distributed naturally; micro-movements consistent with reading; rarely blinking through a full answer.
- Read-aloud artifacts. Occasional mid-sentence restarts as the text stream corrects, answers that end abruptly mid-thought when generation stops, a faint "uh" rhythm matching display scroll.
- Polish-difficulty mismatch. Answers that are consultantly smooth and substantively hollow — frameworks with no numbers, zero specifics about trade-offs, no war stories about the boring parts. Expertise is granular; generated answers are structural.
- In-ear proxies. Small earbuds worn at "the wrong" angle, a second screen visible via reflection, off-camera second voice caught at the edges.
- Environment inconsistency. Lighting or framing that changes between rounds in ways that suggest a different person or setup than the screening call.
What NOT to do about it
Don't accuse on a stack of tells. Don't run "gotcha" AI-detection tools on interview transcripts — they produce false positives on rehearsed humans and are barely better than chance on polished text. Both create incidents out of suspicion and miss the better move: change the questions so assistance stops helping.
Redesign the interview so live AI fails
Live assistance is just an interview skill: it can produce a story, it cannot produce ownership under unscripted depth.
- Walk me through it live. Ask for a screen share of real work — the pull request, the spreadsheet, the campaign dashboard — and narrate choices. Assistive tools stall when the artifact, not the description, is the subject.
- The rejected alternative. "What did you consider and not do? Why?" Generated answers avoid trade-off detail because it requires having decided something.
- Interrogate the numbers. What was the baseline, the result, the cost of the mistake? A person who did it has numbers; a reader has ranges.
- Two levels down. Take their best answer and go one technical or operational level deeper than any prompt would cover. Live assistance survives breadth; it collapses on depth.
- The failure with receipts. Describe the specific moment it went wrong — what did you say to your manager? When did it resolve? Stolen stories compress the middle; lived ones have texture there.
- For take-homes and assessments, the proctoring mechanics exist and are documented across the assessment tools; but the interview-side fix is sequencing: practicals with the candidate present reduce the value of external help entirely.
The structural version of this is a screening pipeline where assessment signals, live practicals, and documented scoring standards feed one evaluation record instead of vibes-per-panelist. AI candidate evaluation covers that design — and it's the part that protects honest candidates most, because consistent evidence beats suspicion.
If you're confident, then what
You will occasionally have a strong read with no proof. The defensible response is not accusation, it's process: say plainly that you'll be doing a live practical portion in the next conversation. Honest candidates shrug; assisted ones negotiate or disappear. The design is its own verdict.
Bottom line
Tells for live AI assistance: answer latency, fixed off-camera gaze, read-aloud artifacts, polish without specifics, earbuds and environment drift. Don't rely on detector tools or accusations — redesign so assistance is useless: live walkthroughs of real artifacts, rejected-alternative depth, granular numbers, and two-levels-down follow-ups. The questions that expose cheating are the questions that never reward it.