How to Detect AI Cheating in a Cloud Architect Case Study Interview

Cloud architect case studies ask candidates to design a system live — multi-region failover, cost tradeoffs, security boundaries. AI cheating here looks like a candidate narrating an architecture diagram that a hidden LLM window generated seconds earlier: too complete, too fast, using textbook terminology without justifying tradeoffs when pressed. Neuroxa's AI Meeting Proctor watches gaze and tab activity through the whole call to catch it.

Threat Model	Observable Tell	Confidence
ChatGPT open in second window generating the architecture answer in real time	Candidate reads left-to-right in bursts, pauses match typical LLM streaming-response cadence	High
Pre-generated diagram/answer memorized before the call	Answer is unnaturally polished but candidate can't adapt when you change one constraint (e.g., swap AWS for GCP)	High
AI-solved cost/tradeoff calculation surfaced via screen share	Numbers appear instantly without visible work; cursor idle then jumps to a "clean" answer	Medium-High
Remote proxy/second engineer feeding answers	Verbal answer doesn't match candidate's own resume-stated experience level	Medium

Interviewer script: "Walk me through your design as you build it — I'll be changing a constraint halfway through, so keep it flexible." Mid-answer, swap a requirement (e.g., "actually this needs to run in a region with data residency laws") and watch whether the reasoning updates live or the candidate stalls waiting for a new AI output.

Evidence to capture:

Recording of the full design-narration segment
Tab/window-focus change log during silences
Gaze pattern during the constraint-swap moment
Time-to-first-word after each new constraint
Screen-share metadata if diagramming tool was used

Neuroxa product: AI Meeting Proctor — live call monitoring for Zoom/Teams architecture walkthroughs, flagging tab switches and gaze anomalies tied to your constraint-swap moments.

FAQs

Architects often think with their eyes closed or looking up — won't that trigger false positives? No — Neuroxa's model is trained on natural "thinking gaze" (unfocused, wandering) versus reading gaze (fixed, horizontal scan pattern), and only flags the latter.

What if the candidate uses a whiteboard tool, not just talking? Combine this with screen-share tab-focus logging; drawing tools are expected and not flagged, but a second browser tab with an LLM open is.

Should we ban all AI tool use for architects, since they'll use AI daily on the job? Most teams allow documented AI use post-hire but require unassisted reasoning in the interview itself — set that expectation up front.

Does a fast, confident answer always mean cheating? No — some candidates are simply fast. The signal is the combination of speed AND inability to adapt to a changed constraint.

Verify architecture answers are genuine with Neuroxa.ai AI Meeting Proctor.