Hybrid Proctoring: Why AI Flags Plus a Human Reviewer Beat Either Alone

What the AI catches that a human misses

A single human proctor watching a live video feed cannot reliably catch a hidden phone, a second screen, a coached answer, or a subtle handoff, and cannot review what happened after the fact unless the session was recorded. AI systems track keystroke and typing rhythm, keep the same person confirmed in the seat, watch for a second person entering the room, and monitor application or browser activity for patterns associated with outside help or content that was memorized or copied rather than produced live. The 2025 numbers show the gap: hybrid proctoring detects 35.6 percent violations versus 25.9 percent human-only, and AI flags 35 percent more than humans alone.

The scale of the problem is what makes the AI necessary. By February 2026, CodeSignal documented that cheating on technical hiring assessments had doubled in a single year, from 16 percent to 35 percent of test-takers. HirePro's analysis of platform data across entry-level assessments put the figure between 30 and 50 percent. A 2025 survey found 59 percent of hiring managers suspected candidates of faking abilities with AI during live assessments, while only 38 percent had structured proctoring in place. The gap between suspicion and structure is the market for proctoring, and it is growing: the AI proctoring market was valued at $1.42 billion in 2025, growing 20 percent year over year. The second device problem is the one humans miss most: a phone off camera is invisible to a live proctor and obvious to a system watching gaze direction and audio.

What the human catches that the AI gets wrong

Automated cheating detection is not reliable enough to stand alone, and the evidence for that is in the institutions that turned it off. Yale, Vanderbilt, and Johns Hopkins Universities disabled Turnitin's AI writing detection feature, citing concerns about false positives. The same failure mode shows up in proctoring: glasses adjustments flagged as gaze shifts, a family member walking through the background flagged as a second person, a nervous candidate's typing rhythm flagged as copy-paste. False positives persist, and expulsion rates for caught cheaters are up 15 percent, which means the cost of a wrong flag is higher than it was.

The human reviewer's job is to look at the flag in context, with the full session recording available to pause, rewind, and examine. The AI signals are handed to a trained human reviewer as a diagnosis to investigate, rather than triggering an automatic finding of misconduct. Proctoring produces evidence, people make the decision. The software can tell you a second face appeared on camera or a candidate switched tabs eleven times. It cannot tell you why, and it should never auto-reject anyone. Treating a flag as proof of cheating is the fastest way to lose a good candidate and invite a fairness complaint. Proctoring flags honest students too, and the false positive numbers are the reason the human has to stay in the loop.

How the hybrid actually runs

Most systems run in three stages: a check before the test, live monitoring during it, and a scored review after. Before the test, the candidate confirms identity with a webcam photo matched to a photo ID and does a quick scan of the room and desk. During the test, the system reads face presence and count, gaze direction, voices or coaching in the audio, tab switches, copy-paste, full-screen exits, and connections to a second monitor. After the test, the AI records and analyzes the session, then highlights suspicious events for a reviewer, with timestamps, severity ratings, and the relevant clip or screenshot, rather than hours of raw footage.

Some systems look across sessions as well as within one. The Duolingo English Test analyzes patterns across thousands of test sessions at once, for example near identical answers from test takers in the same region, which a single proctor watching one session could never see. That cross-session view is the strongest argument for the AI layer, and it is also the one that needs the human most, because a pattern across sessions is a statistical claim, not an observation, and a statistical claim needs a person to decide what it means. Contract cheating in 2026 runs on proxies and remote access, and the hybrid model is the only one that catches both the proxy and the person behind it.