Your Face Is the Ticket. What Happens When the Computer Says No?

Imagine studying for months, driving hours to an exam center, and then being stopped at the door — not because you forgot your ID, but because a computer looked at your face and decided it wasn't sure about you. No human override. No explanation. Just a flag on a screen and a question mark where your exam seat should be.

That's the risk hiding inside a very real headline: on June 28, more than 600,000 people sat down to take the MahaTET — a teaching eligibility exam in Maharashtra, India — under a system that used AI monitoring and biometric screening (think: face-matching technology that compares your live face to a photo on file, to confirm you're really you) to verify every single candidate.

Six hundred thousand people. One biometric gate. And almost certainly, zero public disclosure about error rates.

Why Exams? Why Now?

Exam fraud in India has become a genuine crisis. Proxy testing — where someone pays another person to sit the exam in their place — has derailed careers and corrupted credential systems for years. Then deepfakes arrived. Now you can fabricate a convincing photo ID. You can create forged documents that look real enough to pass a tired human examiner at 7am. Manual checks, alone, are no longer enough.

So officials responded at scale. For the NEET-UG retest in 2026 — India's national medical entrance exam — India.com reported that authorities deployed over 1.38 lakh (138,000) CCTV cameras, 51,000 signal jammers, and specifically assigned 48,448 personnel to handle biometric verification across 95,000 exam rooms. That's not a pilot program. That's a full national infrastructure bet on this technology.

The MahaTET on June 28 follows the same logic. Biometrics catch imposters. AI cameras watch for suspicious behavior during the test itself. The intention is genuinely good.

But good intentions and reliable execution are two very different things. This article is part of a series — start with Deepfake As A Service Fake Boss Scams Workplace Risk.

The Number Nobody Puts on the Flyer

Here's the thing that should be on every exam notice but never is.

Facial recognition algorithms — the software doing the face-matching — don't give you a yes or a no. They give you a confidence score. Basically: "I'm 97% sure this is the same person." The humans running the system then set a threshold: how sure does the computer need to be before it waves you through?

Set the bar low, and more fraud gets through. Set the bar high, and more real people get blocked. iMEdD Lab's analysis of real-world facial recognition data found that without confidence thresholds, algorithms miss about 4.7% of valid matches on real-world photos. Tighten that threshold to accept only 99% certainty matches, and the rejection rate for legitimate people jumps to 35%.

Thirty-five percent. At an exam with 600,000 candidates, even a fraction of that rate means tens of thousands of honest people hitting a wall.

And it doesn't hit everyone equally. The Bipartisan Policy Center's review of NIST's facial recognition vendor testing found that these algorithms consistently produce higher error rates for people with darker skin tones and for women. That's not a fringe finding. It's a documented, repeatable pattern across multiple vendor systems. When you run a sweeping biometric check on a demographically diverse population — which every large Indian exam certainly is — that bias doesn't average out. It concentrates.

"Algorithms tested under the NIST Facial Recognition Vendor Test showed a 4.7% miss rate on real-world photos without confidence thresholds — jumping to 35% when vendors tighten thresholds to only accept 99% certainty matches. The exact tradeoff exam operators face: reject false positives, but lock out legitimate test-takers." — Expert analysis, CSIS Strategic Technologies Blog

There's also a pure physics problem. CaraComp's education resource on accuracy measurement explains that facial recognition accuracy drops 30 to 40 percentage points when moving from clean studio-quality photos to real-world surveillance images — the kind taken under fluorescent exam-room lights, with candidates who haven't slept, who are nervous, who might have a new haircut since their registration photo was taken six months ago. Lab benchmarks don't tell you how the system performs at 8am in a crowded hall. And vendors almost never publish that data.

So What Actually Happens to the Person Who Gets Flagged?

This is the question that keeps the whole story honest.

When a bank's face-match system has doubts, a human steps in. You talk to someone. You show additional documents. The stakes are a delayed transaction. When an exam system flags you, the stakes are your career path — a teaching credential, a medical license, a government job. One flag at the wrong moment and months of preparation are gone. Previously in this series: Your Bank Wants To Scan Your Face Heres The One Rule That St.

Proctortrack's analysis of exam security makes a useful point: modern biometric systems for exams go beyond a single photo match. They check for "liveness" — meaning they verify you're a real person physically present, not someone holding up a photo — and they analyze whether your facial movements match what a real, awake human looks like. That's genuinely useful against deepfake spoofing, which is a real and growing threat. The same research on AI-assisted cheating notes that layered biometric systems create much harder targets for fraudsters than simple photo ID checks.

The technology, in other words, is not the villain here. Fraud is real. Deepfakes are real. Proxy testing ruins the value of legitimate credentials for everyone. Nobody reasonable wants a system with no checks at all.

The problem is deployment without disclosure.

What You Have the Right to Know

If your face is going to be the deciding factor in whether you access a career credential — a teaching license, a medical exam seat, a government job — there are five things you should be able to find out before you walk through that door. Not after you've been flagged. Before.

One: What is this system's documented error rate for people with my demographic profile? Not the lab benchmark. The real-world number, for people who look like me.

Two: What happens if I'm flagged? Is there a human in the room with authority to override the system? Is there an appeal process that doesn't require me to miss the exam?

Three: How long will my biometric data (my face scan, fingerprint, or any body-based identifier they collect) be stored, and who can access it?

Four: Was this system independently tested before it was deployed on 600,000 people? Up next: Your Boss Just Called It Wasnt Him And It Cost 25 Million.

Five: Who is responsible if the system is wrong — the exam board, the vendor, or me?

If you've ever looked at a photo of yourself on an old driver's license and thought "that barely looks like me," you already understand exactly why these questions matter. That's the exact moment facial recognition systems can fail you. The Springer Nature research journal AI & Society documents precisely this gap: error rates in operational settings vary dramatically from controlled benchmarks, and that variation is not distributed evenly across populations.

One thing you can do right now, if you're facing any kind of identity check for an exam, a job application, or a license: before the day of, contact the exam body in writing and ask for their appeal procedure in the event of a biometric mismatch. Get it in writing. Not because you expect to be flagged — but because having a documented answer puts you in a much stronger position if you are. Organizations that have a clear, fair appeal process will tell you. Organizations that don't have one will suddenly feel pressure to create one.

That one question, asked by enough candidates, does more than any policy paper.

Here's what I keep coming back to. India deployed 48,448 dedicated biometric verification staff for a single exam retest. That's a small city of people whose entire job, on that day, was to make sure the technology got it right. Someone in that room clearly understood that the algorithm alone wasn't enough — that you need humans in the loop, ready to catch the mistakes a computer makes.

The real question isn't whether 600,000 people should have been screened. It's this: if the exam board already knows the system needs 48,000 human backups to function fairly, why isn't that number — and what triggers those overrides — printed right there on the admit card?