AI proctoring for university admissions: how AdmissionGuard works

Admissions has a question hiring mostly doesn't: is this really them, and did they really do this? AdmissionGuard is the product we built for that question — Lantern's interview carried into university admissions, with the proctoring that stakes like these demand: AI-proctored aptitude tests, written assessments checked for generated text, and structured video interviews with behavioural analysis and ML anomaly detection.

It runs at 99.7% cheating-detection accuracy with a false-positive rate under 0.1%, processes in under 100ms, and holds 99.9% uptime. The University of Hong Kong uses it to screen 10,000+ applicants a year.

Why does admissions need proctoring when hiring doesn't?

Because the threat model inverts. In hiring, the risk you're managing is a candidate who can't do the job; the employer finds out within a quarter and the correction is expensive but available. In admissions, thousands of applicants compete for a few hundred places against a scored artefact, and the correction arrives years later or not at all.

That changes what has to be verified. When a place is zero-sum and awarded on a test result, the incentive to have someone else sit the test is structural rather than exceptional. Proxy test-takers, AI-written essays and outright impersonation are the specific failures the format invites, and an assessment that can't detect them is scoring something other than the applicant.

What are the three modules?

Aptitude tests, written assessments and video interviews — each proctored differently, because each is gamed differently.

• Aptitude tests. Cognitive assessments watched in real time, sat inside a secure browser lockdown. The lockdown handles the mechanical routes — a second window, a second screen, a lookup mid-question.
• Written assessments. AI-proctored essays and subject papers, checked for generated text and plagiarism. This is the module that changed most in the last two years, for reasons nobody needs explained.
• Video interviews. Structured interviews with behavioural analysis and ML anomaly detection, which is where impersonation and coaching surface.

The modules are separable. An institution that only needs proctored written work doesn't have to adopt an interview stage to get it.

What does 99.7% accuracy with under 0.1% false positives mean in practice?

The second number is the one that matters more. Detection accuracy tells you how much cheating the system catches; the false-positive rate tells you how often an honest applicant is accused, and in admissions that error is the one with a person attached to it.

Any detection system trades between the two, and a vendor quoting only the accuracy figure is quoting the half that flatters them. A system tuned to catch everything will flag nervous candidates, poor connections, and anyone whose room has a door in it. Under 0.1% is the constraint we hold the accuracy number against, and it's the number we'd want to interrogate if we were buying rather than selling.

What happens when something is flagged?

It goes to a person. This is the design principle AdmissionGuard inherits from Lantern, and we treat it as non-negotiable: a flag is evidence with a timestamp attached, not a verdict, and the institution's admissions team decides what it means.

The reasoning is the same in both products. An automated system that rejected applicants on a detection signal would be making a consequential decision about a person's access to education on the basis of a probability, with no route to explanation and no way for the applicant to answer. Under 0.1% of a large applicant pool is still real people. Evidence to a human, human decides. The equivalent logic on the hiring side is in deepfake candidates and interview fraud.

What does this look like at 10,000+ applicants a year?

Under-100ms real-time processing and 99.9% uptime stop being specifications and start being the whole product. At admissions scale the assessments do not arrive evenly — they arrive in a spike against a deadline.

That's the load the numbers describe. Real-time processing means a flag is raised during the assessment rather than in a report generated next week, which is what makes intervention possible at all. Uptime means the deadline-eve surge doesn't turn into a queue of applicants who couldn't sit the test. HKU screens 10,000+ applicants a year through it, which is the volume claim we can make.

Which standards apply to student records?

FERPA is the one that's specific to this context — it governs student education records, and it's the reason an admissions product can't simply reuse a hiring product's compliance posture. AdmissionGuard holds SOC 2 Type II, ISO 27001 and GDPR as Lantern does, plus FERPA.

For institutions running procurement, the useful questions are the same ones we'd give an employer: ask for the classification in writing, a sample audit export for a single applicant, the false-positive methodology, the retention and deletion terms, and a clear statement of who holds the decision when something is flagged. We go through that checklist for the hiring side in AI hiring compliance in Hong Kong, the US and the EU.

Common questions

Is AdmissionGuard the same product as Lantern?

It's an extension of it. The interview engine is shared; the proctoring modules, the FERPA obligation and the threat model are specific to admissions. An employer doesn't need AdmissionGuard, and an admissions office usually needs more than Lantern alone.

Can an essay be flagged as AI-written just because it's well structured?

That's precisely the failure the false-positive constraint exists to bound. Written assessments are checked for generated text and plagiarism, the result is evidence rather than a determination, and an admissions officer reads it before anything happens to the applicant.

Does an applicant know they're being proctored?

Yes — proctoring that the assessed person doesn't know about is a different and much worse product. Transparency to the person being assessed is a principle we carry across both products.

If you run an admissions office deciding how to verify at scale, we'll walk you through it.