You limit AI cheating in interviews mostly through question design, not detection. Ask about the candidate's own past work, follow up twice on the same thread, and ask them to critique rather than produce. Use proctoring at bounded assessment stages, say so in advance, and treat any detection flag as a prompt for human review.
How much interview cheating is happening?
Reliable numbers are scarce. In a Gartner survey of 3,000 candidates in 2025, 6% admitted to interview fraud, and Gartner predicts that by 2028 one in four candidate profiles worldwide will be fake. Self-reported figures probably undercount, and many other numbers come from vendors who sell detection, so treat them with care.
What does AI cheating look like?
Three behaviours call for different responses.
Real-time answer copilots. An app listens to the question and shows a polished answer on an overlay the interviewer cannot see. This is most common in unstructured phone and video screens.
External code assistants. In a technical round, the candidate solves the problem in another tool and pastes the result back.
Identity substitution. Someone else sits the interview. This is rarer and more damaging, and it is where identity checks matter more than question design.
Why does a detection-first strategy disappoint?
Detection signals include eye-gaze patterns, response delays, keystroke rhythm and mismatches between spoken fluency and code style. Each is probabilistic, and each can misfire on candidates who are nervous, neurodivergent, non-native speakers or on a bad connection. You rarely get proof, only suspicion.
Let detection raise a flag that a person reviews alongside everything else. Do not let it reject automatically. The legal side is covered in Should AI Be Allowed to Reject a Candidate on Its Own?
Which questions are hard for a copilot?
Ask about their history, not the general topic. "How would you design a rate limiter?" is easy for a copilot. "Tell me about a time your rate limiting failed in production, what you saw first and what you changed" is not, because the model does not know their incident.
Follow up twice on the same thread. Generated answers are strongest on the first response and weaken under a second and third probe.
Ask them to critique. Hand over a flawed design or working-but-wrong code and ask what they would change and why.
When is proctoring appropriate?
At the assessment stage, where the task is bounded and the stakes are clear. It fits poorly in an exploratory first conversation, where it mainly signals distrust.
Wherever you use it, say so in advance and say what is recorded. Surprise monitoring is what hurts your employer brand. In Greenhouse's 2026 survey, AI monitoring during the interview was among the top reasons candidates gave for dropping out (26%).
How GetHirePlus supports this approach
AI resume screening with a relevancy threshold helps decide who to interview first, which frees attention for later stages.
AI phone interviews use your own question bank, so you can write history-based questions, and each interview returns a full transcript you can re-read.
Candidate assessments are proctored, and coding questions use an in-platform compiler instead of the candidate's own editor.
Live video interviews include a built-in code editor with recording, so you see how an answer was reached.
Is all AI use by candidates cheating?
No. A candidate using AI the way they would on the job may be showing the skill you want. What you object to is undisclosed substitution of someone or something else for the person you are evaluating. State which stages allow tools and which do not, and much of the ambiguity goes away.
No setup makes an interview cheat-proof. The realistic goal is to make honest preparation cheaper than cheating, and to collect enough signal for a person to make a defensible call.
Sources
Gartner, candidate trust and interview fraud survey findings (31 July 2025)