How do I proctor coding assessments and technical tests?
TL;DR: Coding tests are the most AI-compromised assessment format in hiring: standard challenges are solved by AI assistants instantly, and real-time copilots feed answers during live coding calls. Proctor the take-home stage with a locked, monitored browser session around your coding platform's URL, and proctor the live round with an AI agent in the call. Decide explicitly which stages allow AI — then enforce the boundary you chose.
The claim: an unmonitored coding screen measures AI fluency, not engineering
Evidence: Popular coding-challenge formats are exactly the pattern generative models excel at: well-specified problems with known solution shapes. A candidate with an assistant in a second tab completes screens in a fraction of the expected time with idiomatic solutions — indistinguishable from a strong candidate by output alone. In live rounds, real-time interview copilots overlay suggested answers during the call. Hiring teams that don't monitor either stage are ranking candidates by tooling and nerve, then discovering actual skill levels after the start date.
Coding assessment stages and controls
| Stage | AI threat | Control |
|---|---|---|
| Take-home / async screen | Full AI solution in another tab or device | Wrap platform URL in monitored, locked browser session; tab-switch and multi-monitor detection; gaze and audio monitoring |
| Live coding call | Real-time copilot overlays, coached answers | AI Meeting Proctor in the Teams/Zoom call: gaze anomalies, reading cadence, second voices |
| Any stage | Proxy candidate | ID + selfie match, continuous face verification, virtual-camera flags |
| AI-allowed stage (by design) | None — it's permitted | Identity verification only; observe how they use AI |
Step-by-step: an integrity-sound technical funnel
- Declare the AI policy per stage. "Screen 1: no AI, monitored. Stage 2: AI-allowed pairing exercise." Ambiguity is where both cheating and candidate resentment live.
- Proctor the screen via URL. Your coding platform stays; Neuroxa.ai wraps its link with lockdown, monitoring, and identity verification. Zero installs keeps completion rates healthy.
- Watch the tells in review. Solution pasted in bursts after off-screen gaze; typing rhythm mismatched with thinking pauses; a second voice. The trust report timelines all of it.
- Put the AI Meeting Proctor in live rounds. It monitors for copilot-usage patterns and identity swaps while your engineer runs the interview.
- Ask follow-ups that punish borrowed code. "Why this data structure?" Unscripted probing plus monitoring is the strongest combination available.
- Keep evidence for disputes. Rejections tied to integrity flags should rest on the exported report, not a reviewer's hunch.
FAQ
Shouldn't we just allow AI, since engineers use it on the job? Allow it where you measure AI-assisted productivity — deliberately. Keep at least one monitored stage that measures unaided reasoning, because debugging, review, and design still require it.
Do candidates accept proctored coding tests? Serious candidates do when it's disclosed, quick to join, and install-free. The drop-offs skew heavily toward those depending on assistance.
Can copilot use in live calls really be detected? The pattern is detectable: gaze locked to a fixed off-camera region, reading cadence in answers, latency mismatches — flagged for your interviewer to probe.
Does this work with our existing coding platform? Yes — URL-based proctoring wraps whatever platform you already use; no integration required.
By Pinal Dave · Last updated: 2026-07-23