← Blog

You can't prevent cheating. But you can get better at detecting it.

August 11, 2026 · 11 min read

I'll say it straight, even if it makes some uncomfortable: you can't prevent candidates from cheating on an online exam. No camera, no lock, no monitoring AI can do it completely. We just saw this with the entrance exam of one of Mexico's largest public universities, which was conducted online for the first time this year… and ended with answers being sold and suspicion that AI was used during the test. The interesting part, and it took me time to understand, is that accepting that it can't be stopped isn't giving up.

What happened a few weeks ago

Without naming names, because this isn't about revenge: this year, one of Mexico's largest public universities administered its undergraduate entrance exam online, nationwide, for the first time. And let's put the numbers on the table, because they explain everything else: over 191,000 young people registered, about 158,000 took the test… and the seats were around 22,000. Nearly one in seven. With that pressure, what happened next shouldn't surprise anyone.

An open black market emerged. Answers were sold for about 200 pesos on social media, and "complete" packages — real-time help during the exam or someone taking it for you — reached around 4,000. The university itself acknowledged investigating whether the unusual increase in correct answers was due to a leak of answers, the use of artificial intelligence during the exam, or both. In other words: AI is no longer a future threat in entrance exams. It’s here, and it’s making a big entrance.

So big that the university ended up summoning nearly 58,000 applicants for a control exam — this time in person — of which only about 22,000 will keep their place. A process involving almost two hundred thousand people, entirely in doubt.

The exam was already a business (before the cheating)

And before talking about cheating, look at the business behind it, because it explains part of the pressure. Taking that exam cost nearly 500 pesos. Multiply that by the more than 191,000 who registered, and it amounts to around 90 million pesos — in a single process, for one university — with only one in seven getting a spot.

And that’s just the public university, with subsidized fees. Private universities charge between 500 and 3,000 pesos for the entrance exam, sometimes per subject. Applying to university in Mexico is a market in itself: all charge for the attempt, whether few or many try. With so much money and so many dreams at stake, someone selling answers for 200 pesos isn’t an anomaly — it’s basic economics.

Since when do we trust a screen with an exam?

And this isn’t new, even if it seems that way. Not long ago, an exam was paper, a classroom, and a teacher walking between the rows. Trust was placed in the proctor: if someone looked away, it was visible.

The pandemic pushed many exams online suddenly, and with that, an old question with new clothes reappeared: how do I know that the person on the other side is who they say they are, doing what they say they’re doing, and only that? The answer came in layers. First, passwords. Then, cameras that watch you while you answer. Next, programs that lock down the rest of the computer so you can’t open anything else. Each layer was created to cover the gap left by the previous one.

It’s a cat-and-mouse story, and it’s important to keep it in mind because it explains why each new solution arrives with an air of finality and, within months, finds a way to be bypassed.

What companies are selling today

Today, there’s an entire industry around this. They offer a camera that monitors you and a program that raises a hand if you look away too much, if another voice is heard, if a second screen is detected. They offer browsers that lock themselves down. It sounds robust, and some of it works.

And it’s not a small business: just the software for online exam monitoring already moves over one billion dollars a year worldwide, and it’s growing at double digits, set to multiply several times this decade. It’s an entire industry built on one promise: the infallible lock.

But there are two things that almost no brochure tells you, and they’re exactly what matter.

The gap that no camera can cover

The first: cheating no longer needs to hide. A student can read the question aloud very quietly, and an AI on their phone can answer in their ear. There’s no lock on the screen, no camera pointing at their face, that can stop that. The expert solving the exam isn’t hiding in a sleeve: they’re in everyone’s pocket, answering in seconds.

When you understand that, you stop fighting a battle that can’t be won. Perfect prevention doesn’t exist. And accepting that, far from giving up, is what finally makes the problem manageable.

So, do we give up? No.

For years, we tried to build the perfect lock, and every lock was broken. The lesson isn’t to give up: it’s to change the question. Stop asking "how do I make it impossible?" and start asking "how do I notice it?".

And it turns out that cheating leaves traces. The person dictating the question to a machine and waiting for the answer takes the same time on easy and hard questions — because they’re not thinking, they’re waiting. If, halfway through the exam, you ask "explain why you chose that," the one who knows responds, and the one just copying stays silent. None of those traces, alone, prove anything. But several together reveal a pattern.

And here’s the key difference that changes everything: a lock doesn’t improve with use; detection does. Every exam you take teaches you to recognize those patterns better. The more cases you see, the sharper your eye becomes. The wall is always the same; the eye, however, learns.

The part that should keep you awake at night

Here’s the second thing almost no one tells you, and it’s the most important. All this monitoring has a cost, and it’s not paid by the cheater: it’s paid by the honest student.

It’s documented that cameras claiming to "detect suspects" fail more often with students of darker skin — sometimes they don’t even recognize their face, and they’ve had to shine a light on their face for the program to "see" them — and that they punish nervous students: those biting their lips, looking up when thinking, as if doubting were cheating. And there’s a cruel, counterintuitive number trap: if only a few out of every hundred candidates cheat, even a "very good" detector will mostly flag innocents — not because it fails a lot, but because innocents are the vast majority, and a small percentage of a crowd is still many people.

Wrongfully accusing a student who fought for their place in university isn’t a software error: it’s real harm to a real person. That’s why, in all this, there’s a line that must not be crossed: no machine can have the final word. It can raise a hand; the decision is made by a human who sees the full case.

The truth filter was never the lock

In my experience, the truth filter was never the lock. It was the conversation. The person you truly knew was the one sitting in front of you, telling their story, not the one who filled out a form correctly.

A multiple-choice exam answered alone in front of a screen measures less and less. AI models didn’t break the exam: they broke the poorly designed exam. What no AI can do for a candidate is sit across from you, defend their reasoning, and stand by it when you question them again. There’s no voice whispering the answer in their ear.

The solution isn’t a wall: it’s layers

So, what does a working solution look like? Not like a bigger lock, but like a funnel. Think of the sieve you use in the kitchen to sift flour: if you shake it, the coarse stays on top, and only the fine passes through. Now imagine several of those sieves, one behind the other, each with smaller holes than the previous.

The first layers are cheap and filter out most — those cheating out of laziness — without bothering anyone: questions that can’t be captured at a glance, a different version for each candidate so sharing answers doesn’t work. Then, AI comes in, not to judge but to observe patterns and raise a hand. The one answering the hard question as fast as the easy one, or unable to explain why they chose their answer… AI flags them. The last, narrowest layer is a person: the conversation.

Pay attention to the order because it’s everything: AI doesn’t replace humans; it tells them who to watch. It handles the boring work of reviewing hundreds and leaves the only thing humans do well: deciding about another person.

No system catches 100%, and that’s okay

Here’s the part that’s hard to accept and that, once accepted, sets you free: no system will catch everyone. Chasing 100% is exactly what leads to the abusive monitoring we discussed. The margin of error isn’t a failure of design; it’s part of it.

Because what’s valuable isn’t perfection, but learning. It’s a cycle: each admission season, each case clarified through conversation — whether someone was cheating or just nervous — teaches the system to improve for the next time. The wall you buy is the same every year until someone figures out how to jump it. A system that learns from its own cases is better in March than it was in September. That’s the difference between spending and building.

And maybe you don’t have to buy it from outside

I’ll finish with what matters most to me, and I say it as someone who sells technology to schools: there’s money that makes no sense to spend outside.

Think about the case we discussed. That university makes its own exam, with its own people, precisely because it has the talent and size to do so. And even then, they were fooled. The easy reaction would be to run and buy the most expensive lock on the market. But the lesson is the opposite: if your institution has talent — and many do, sometimes within their own classrooms — a detection and learning system can be built and kept in-house.

And look at the full picture: an industry worth over a billion dollars a year selling cameras and locks; a black market responding with answers at 200 pesos; and in the middle, twenty-two thousand places for nearly two hundred thousand dreams. With that pressure, cheating will always exist — there’s no wall to erase it. The real question isn’t how to build the impossible lock, but who keeps your money and your judgment: an outside provider, or your own team.

Money is much better spent paying only for what truly makes sense externally — servers, infrastructure that isn’t worth building yourself — and keeping what makes you unique: the conversation, the judgment, the system that learns from your own candidates. The institutions that will succeed aren’t the ones buying the most expensive wall. They’re the ones stopping to buy walls and starting to build them. That’s not bought; it’s cultivated.