What is AI-proof exam proctoring?

The category didn't exist three years ago

In November 2022, when ChatGPT shipped, the online proctoring industry was a relatively settled market. Companies like Honorlock, ProctorU, Examity, and ProctorTrack had carved up the territory based on a shared assumption: that cheating leaves traces. A student looking off-screen for ten seconds. A second voice in the room. A browser tab opened to somewhere it shouldn't be.

You could review the recording later. You could flag the segment. You could escalate.

That assumption was destroyed in about ninety days.

By early 2023, students were using GPT-4 to dictate exam answers through wireless earpieces, paraphrase essay prompts in real time, and produce indistinguishably "human" written responses on demand. None of these leave a video trace. None of these are visible in a hand-flagged review.

The industry's first response was AI-text detection: bolt on a classifier, run student submissions through it, flag the suspicious ones. This worked for about six months. Then GPT-4 got better at imitating human writing, and detectors started returning a meaningful false-positive rate on essays written by actual humans — including, famously, the U.S. Constitution and several students with English as a second language.

So a new category had to exist. The question is what should be in it.

Why "detect-and-review" stopped working

The detect-and-review model assumes three things:

1. The cheating leaves evidence. A second face on camera, a mouse click in a forbidden tab, a typing pattern inconsistent with the student's known style.
2. Reviewing evidence is feasible. A human can watch flagged segments, judge them, and adjudicate.
3. The judgment will hold up. When a student appeals, the evidence and reasoning are robust enough to defend.

All three broke at the same time.

Evidence vanished. A bone-conduction earpiece is invisible to a webcam. A teleprompter app on a second device, positioned just outside the camera frame, doesn't trigger a tab-switch flag. ChatGPT-generated answers don't fail keystroke-pattern analysis if the student dictates them by hand at a natural pace.

Review became infeasible. Even when you do flag suspicious segments, the volume swamps human reviewers. UNISA, with 370,000+ enrolled students, would generate millions of "potentially suspicious" segments per exam window if every gaze break and audio spike triggered a manual review. The disciplinary backlog at major South African universities now runs twelve months in many faculties — partly from this exact problem.

Judgments stopped holding up. AI-text detectors fail in tribunal. False positives from gaze-tracking get students with disabilities flagged. The institution loses appeals; the credential loses defensibility.

The detect-and-review system was built for a world where cheating was rare, observable, and individual. It is now common, invisible, and amplified by tools more capable than the proctors.

What makes proctoring "AI-proof": five ingredients

A proctoring platform that can plausibly call itself AI-proof has to do something the previous generation didn't: prevent, not just record. Here's what that actually requires.

1. A hardened lock-screen

The student's device is the surface where AI cheating happens — that's where ChatGPT lives, that's where the second-screen teleprompter runs. If you don't control that surface, you don't have AI-proof proctoring; you have a webcam.

A real lock-screen is a native shell, not a browser extension. It runs at OS privilege level, blocks alt-tab and process-switching, prevents external display mirroring, and disables paste from outside applications. On macOS this is harder than on Windows; on Android it's trickier than on iOS. The platforms that genuinely cover all four (macOS, Windows, Android, iPadOS) are the only ones that can claim AI-proof status — anything less leaves a gap.

2. Continuous identity verification

Identity isn't a one-time check at the start of the exam. It's continuous: face match every 30–60 seconds, on every break, on every reconnect. The student in seat at minute 17 must be the same student who started.

This catches the most common high-value cheating mode in remote exams: paid impersonation. A student pays a graduate to take their final. One face check at minute one doesn't stop this; continuous verification does.

3. Dual-camera coverage

The device camera sees the student's face. It does not see the student's hands, the desk, the keyboard, the printed cheat sheet, or the second device sitting two feet to the right.

Real coverage requires a second camera — typically a phone running a side-camera app — angled to capture the workspace. The geometry alone makes most physical cheating channels infeasible. A teleprompter on a second device is now visible. Hidden notes are visible. A whisper coach in the same room is visible.

This is the layer the legacy single-webcam proctors don't have. It's also the layer that's logistically annoying — students don't love mounting a second device — but it closes more cheating vectors than any other single component.

4. On-device behavioural anomaly flagging

Not all flagging is bad. The mistake of legacy systems was uploading twelve hours of video and asking humans to review it. The fix is on-device, real-time models that surface only the moments worth reviewing — a second face appearing, a pattern of gaze breaks consistent with reading off-screen, a voice that isn't the registered student's.

The shift is from "record everything, review later" to "flag a few minutes, package the evidence, upload only what matters." This collapses server load, reviewer hours, and disciplinary backlog simultaneously.

5. Writing-pattern analysis

The last layer is for the answers themselves. Not "did GPT write this?" — that's a losing game. The right question is: is this answer consistent with this student's prior writing? Writing-pattern analysis builds a profile from each student's earlier work (assignments, drafts, prior exams) and flags submissions that diverge sharply.

This isn't perfect either, but it's the right framing. You're not detecting AI; you're detecting discontinuity — and discontinuity is what cheating produces, regardless of the method.

What AI-proof proctoring is NOT

Some honest limitations.

It's not a guarantee. A determined, well-resourced student can defeat any proctoring system — including in-person, paper-based exams. The goal isn't perfection; it's raising the cost of cheating high enough that the marginal student stops trying.

It's not a substitute for exam design. Open-book essays asking factual recall in subjects where ChatGPT performs perfectly are a design problem, not a proctoring problem. The best AI-proof systems pair with assessment redesign — more application-level questions, more scenario reasoning, more handwritten work.

It's not equivalent to in-person exams. In-person exams have geographic and physical advantages no remote system replicates. AI-proof proctoring is the right tool when remote is the requirement (distance learners, working professionals, candidates in remote regions) — not the optimal tool in absolute terms.

Evaluating an AI-proof proctoring platform: a checklist

If you're piloting a platform and want to verify it's actually AI-proof rather than rebadged, ask:

1. Does the lock-screen work at OS level on macOS, Windows, and at least one tablet OS? A browser-only solution is not AI-proof.
2. Is identity verification continuous, not one-shot? If it only checks at start, paid impersonation defeats it.
3. Does the platform support dual-camera, not just device-camera? Single webcam = workspace blind spot.
4. Does flagging happen on-device, with packaged evidence uploads — or does it stream continuous video to the vendor's servers? The latter dies under load and crushes networks. (This matters disproportionately in South Africa, where mid-exam network drops are routine.)
5. Can the exam run fully offline after a single download? Stage 6 load-shedding cancels exams that require continuous network. If your proctoring vendor needs a live connection throughout, you're one substation fault from a Senate-level incident.
6. Where does student data live? SaaS-only is fine for some institutions; on-prem or sovereign deployment is non-negotiable for others (government certifications, military, finance). Ask before you sign.
7. What's the appeal process? Every flag should be reviewable, every decision auditable, every action defensible to a tribunal. AI-flagged evidence the student can't see and the institution can't explain is a lawsuit waiting to happen.

A platform that hits all seven is probably AI-proof in the meaningful sense. A platform that hits four or fewer is using the term as marketing.

The South African angle

South Africa is one of the few markets where every constraint that breaks legacy proctoring shows up at once.

Stage 6 load-shedding removes the assumption of continuous power. An exam that halts when Eskom rotates a substation off-line is not a viable exam. True offline mode — exam fully downloaded, student writes through power loss, evidence packaged on-device — isn't a feature; it's a precondition.

Mixed device fleets mean the platform can't assume current-gen Apple hardware. A typical UNISA cohort writes on a mix of mid-range Android tablets, refurbished Windows laptops, and the occasional Mac. Native shells across all three are the requirement, not the upsell.

Disciplinary backlogs at major SA universities run twelve months or longer in some faculties. The system's choke point isn't catching cheating; it's adjudicating flags fast enough that students aren't held in academic limbo for a full year. On-device flagging that surfaces a small number of high-confidence segments — instead of thousands of low-confidence ones — is the fix.

ECSA, HPCSA and other certification bodies are actively reconsidering remote-exam policy in light of AI cheating. The market is open and the requirements are unusually exacting; a platform that lands here is robust enough for almost anywhere else.

If a proctoring system can defend a UNISA Bachelor of Education final through Stage 6 load-shedding on a R2,500 Android tablet — it can defend a Stripe engineering interview, a CFA exam in Singapore, or a Korean university entrance test. The South African constraint set is the global stress test.

Conclusion: integrity is a stack, not a feature

The AI-proof proctoring category isn't a marketing term — it's the consequence of an actual capability shift on the cheating side. ChatGPT changed the threat model permanently; the systems built before that change can't be patched into adequacy.

What replaces them is a stack: lock-screen at the bottom, identity in the middle, cameras around the workspace, on-device flagging at runtime, writing analysis at the end. No single layer is sufficient. All five layers, combined, are what now keeps a credential meaningful.

Anything less is theatre.