Stickers on Buttons
November 18, 2026 · Edited November 19, 2026
Tomorrow morning, over 550,000 people are signed up to sit the Suneung. 1 Every one of them will have stickers put over the buttons on their clothes. Between 8:10 and 8:40, before the first subject starts, they will roll up both sleeves and have their collars checked. Every pair of glasses gets inspected as well. The Prime Minister announced it yesterday in a national address, two days before the most important exam in the country. 2
I'll say the obvious part first. On Thursday, a camera in a button is cheating. The Suneung is closed-book and it exists to rank people, so the kid who follows the rules should not lose to the kid who doesn't. If I were in that situation, I would want the check as well. But I think the stickers are a patch on something that is no longer worth patching. I want to explain why, and then say what I would do instead, plus better ways to confront them as of now. Here are the contents for this note:
- The patch: What was announced, and why it will not hold.
- The actual problem: What the Suneung measures, and why that is the real problem.
- A hotfix for tomorrow: What would actually help tomorrow.
- The sustainable fix: An open-model section of the exam.
- What schools would have to teach if the changes happen
1. The patch
Let's start with the phones. Sometime between 8:25 and 8:30, every phone in the country will go off with a top-tier emergency alert. The government is sending them to all of Korea at once, the kind you cannot opt out of, the kind built for large-scale disasters. The alarm lasts about 10 seconds. The purpose is to make any phone hidden in an exam room ring. Officials admitted there was disagreement inside the government about it, and said they could not find another way to detect a hidden phone. 3 I will suggest actual viable methods (better than this) down in Section 3.
The government's logic is that smart glasses and button cameras need a phone nearby to reach the AI, so find the phone and you find the cheater. 2 That is true for the glasses you buy from an official merchant like Meta, but it is not true for a specialized device someone built only to cheat. A camera with its own cellular module, or a pocket router, has no speaker and no screen. It does not ring. And if there is a phone, it is switched off before 8:25 and back on after. The people designing these tools are not stupid, and they heard the same announcement we did. So the alarm will catch the kid who forgot instead of the kid who planned. The whole country gets a disaster siren, and the people it was aimed at are the people it misses. In addition, the next time that sound goes off for a real emergency, some of us will think of an exam first.
Then look at how fast the hardware moved. A week ago the panic was smart glasses. This week it is buttons. I have designed my own smart glasses as a side project before, and the hardware is getting smaller, lighter, and impossible to tell apart from normal things. The newest frames already look like regular glasses. You cannot sticker your way out of that. What gets covered next year? The cheaters will come up with better options and methods. And I'm pretty sure other methods have been going on for years, with people on the other side instead of AI, and cameras plus ear devices instead of AI glasses. 4 The technological advancement made the entry cost lower though.
The ban for cheating goes from one year to ten, and the government wants to apply it retroactively. After the exam, it plans to pull AI service access records and telecom records for suspected students, with help from global AI companies. The minister admitted the details of that cooperation are not agreed yet, and nobody named the legal basis. 5 I am not saying they will read everyone's chats. I am saying a government announced access to AI logs two days before an exam, and the scope is still undefined.
2. The actual problem
Here is what bothers me most. A YouTuber's AI glasses scored 96 out of 100 on a practice Suneung math test in 18 minutes. 6 In Teenagers in the Age of Artificial Intelligence I called the Suneung a solved benchmark. The second a machine can take a test better and faster than the people it was written for, the test stops meaning what we need it to mean. It still ranks students against each other. It just ranks them on something that is now almost free. So the question is what the test was measuring.
Think about what the Suneung actually rewards: speed, accuracy, recognizing a problem type you have seen a thousand times and executing the known method without mistakes under time pressure. That is still a skill, and students give up years of their lives to build it. It is also the exact thing a language model is best at. We built an exam around the one ability that became cheap. But to be honest, the test system is just very outdated. The test was relatively understandable when it was first introduced.
We have seen this before. When Roy Lee built a tool that beat coding interviews, companies banned it first. Then some of them changed the interview, because the old one had stopped measuring anything. 7 To be fair, the government half knows this. A Blue House task force will keep running after Thursday and review national exams for a world where AI use is normal. 5 Also, "change the exam" is no longer a brave thing to say at all. With the virality of the "AI cheating glasses," everyone is now looking for better ways. Now the question is what you would rank people on when everyone has the same AI. I will talk about this in Section 4.
3. A hotfix for tomorrow
If we are going to patch, we should at least patch well. AI cheating needs three things: a way to capture the question, which is a camera, a link out, which is a radio, and a way to get the answer back, which is an earpiece, a display, or a vibration. Stickers go after the first one, the smallest and easiest to hide. The other two are where a cheater is exposed.
Let's take the link first. Anything that sends a photo to an AI has to transmit, whether it is a phone, a router, or a custom module. The government said it is not bringing in radio detectors. 8 Even a few handheld detectors per site, walked down the corridor during the exam, would cover every device type the alert will miss. The carrier check should also be live. The plan is to pull telecom records after the exam for suspected students, but carriers can already see which devices are active on the cells covering the school's area approximately. A data session starting from inside an exam building is worth flagging right when it happens, not in December. 9
Then the return path. An answer is useless unless it reaches the student, and a tiny earpiece is the classic way. Looking in both ears takes seconds and finds something a sticker never will. Glasses are already being inspected every period, and that is the one measure in the plan aimed at the right place. I would also scan everyone at the door, not only the suspicious. Metal detectors are currently reserved for suspected cases. 10 A device that is switched off does not ring for an alert, but a phone or a router still shows up on a wand. 11
After the exam, I would ask the AI companies for one narrow thing: which accounts submitted this year's actual questions during exam hours. The questions become public that day, so matching is simple, and it needs no access to anyone else's chats. That is a defined scope, which is exactly what the current plan lacks. Comparing scores against each student's mock exam history is already in the plan, and it is the right idea. 2 A jump of several grades in one subject does not prove anything, but it tells investigators where to look.
And I would drop two things. The nationwide alert, which wakes up the whole country to find a phone that a planner has already switched off. And the stickers. Thirty minutes of proctor attention is the scarcest thing in that room, and it is being spent on buttons. None of this fixes the real problem. It is a better hotfix, and it is still a hotfix.
4. The sustainable fix
My answer is the one I proposed in that essay, an open-model exam: give every student the same model and a problem too big to finish alone. Grade what they built, how they checked the AI, and what they did when it was confidently wrong. Judgment still spreads people out, and it is the skill they will actually be using.
Concretely, I would add one new section to the Suneung, sat with an AI model on the desk. Every student gets the same exam-issued device, sealed, with one fixed model snapshot, and nobody brings their own AI. The model runs on servers inside the test site or on a closed line, with no public internet, so there is no outside help and no dependence on a provider staying up that morning. Every prompt and every response is logged, and the log is part of the answer sheet.
The task is one large problem, about 100 minutes, built from a packet of materials released that morning: data, documents, a brief. Some of the materials contain planted errors, and the task has trade-offs with no single right answer, so the model cannot finish it alone. The task type changes every year, so it cannot be drilled like a problem set.
What gets graded is mainly the student's reasoning during the session, read straight from the log. Not how much they typed. What they thought. Out of 100 points, I would give 45 to that reasoning, scored on four things a grader can point to. Did they break the problem down and say what matters before asking for anything? Did they bring their own ideas, or only ask the model for its ideas? Did they question the model, catch the planted errors, and push back when it was confidently wrong? And when there was a trade-off, did they choose and give a reason? Another 30 points go to the final product, because good-sounding reasoning that ends in a wrong answer should not score well. The last 25 go to a short handwritten memo, written with no device after the device is collected: what you verified, what you are still unsure about, and why you made your main decision. That memo is the check that the work is yours. Denmark is piloting the same shape: students may use any tool including AI to prepare, and then have to defend the result in person. 12
The risk with grading reasoning is that students learn to perform it: long, impressive-sounding messages written for the grader. So the rubric scores moves, not words. A one-line message that catches an error beats a paragraph that says nothing, and length earns zero.
It still ranks people. Rubric scores convert to the same nine grades as every other section, with two graders per paper and a third for disagreements. AI can do a first pass on the logs, but a human signs every score. It also changes what cheating means. The incentive to smuggle a camera in a button mostly disappears when the AI is already on the desk and the task is new that morning. You still check for outside help, but you are no longer fighting the tool itself.
People will say everyone would just score 100. They would on today's questions, because today's questions are the kind a model finishes in 18 minutes. Make the problem big enough and open enough, and the model stops being the answer and becomes the tool. Two students with the same tool do not produce the same work. Anyone who has watched two people use the same AI already knows this.
People will also say it is unfair to students who have less access to AI. I think it is the opposite. Right now the advantage goes to whoever's family can pay for the most academy hours. If the exam itself hands every student the same model, the tool is the one thing in the room that is equal.
Then there is the money, which is always the first objection we think of. So here is the math. Each of those 551,864 applicants paid between 37,000 and 47,000 won to sit the exam. Now assume a student pushes a million tokens into the model and gets 200,000 back in a single day, which is a lot. On a solid mid-tier model at today's list prices, around 2 dollars per million tokens in and 12 out, that is under 5 dollars a student. Less than 10,000 won. Even at frontier prices it comes in below the cheapest fee, and a government buying for half a million students does not pay list price. Token prices have also fallen about 85 percent since 2023. Inference is not the obstacle at all. 13
In Korea's case, the pipes are already being laid. The government's own project, AI for Everyone, promises every citizen a free sovereign AI with no usage limit. It went into closed beta in October and is supposed to launch in December. 14 Though, I am skeptical. From the outside it has looked more like buying GPUs than building something people will want to use. But if the state can run a free model for the whole country, it can run one for an exam room, and the exam could plug straight into it. The honest hard part is running it for 550,000 people at the same moment without it going down, and that is an engineering problem, the kind this country is apparently good at.
I would roll it out slowly. In 2027, a pilot in volunteer schools as part of a mock exam, with no stakes, publishing the tasks, the rubric, and how often two graders agreed. In 2028, every mock exam nationwide, fixing what broke. After that, a live section with a small weight, raised only if the grading holds up.
There are things I have not solved. Grading 550,000 open-ended papers consistently is the hardest one. If two graders cannot agree, the section is not fair, and the pilot has to prove it can be done before anything is live. Academies will teach to it within a year, and changing the task type slows that down but does not stop it. The model has to be good enough that the exam is not testing how well students work around a weak tool. And students who have barely used AI before exam day are at a disadvantage unless schools give everyone practice time with the same model first.
5. What schools would have to teach if the changes happen
If the exam grades reasoning with a tool, the curriculum they teach has to change with it. The fundamentals stay, but only for understanding. You cannot catch a wrong answer in math you do not understand. Next to that I would teach estimation, which is knowing roughly what the answer should be before the model gives one, and verification: how models fail, how to check a source, how to read a statistic, when to distrust a confident answer.
Then the things school barely touches now. Framing problems, which means taking something vague and big and turning it into parts you can work on. Explaining your reasoning, in writing and out loud, because if the exam reads how you think, school should be where you practice thinking in front of other people. Long projects with AI, weeks instead of 50 minutes, on the same model every student will get in the exam, so nobody meets the tool for the first time on exam day. And deciding and owning it: making a call with incomplete information and putting your name on it.
What gets cut to make room is speed drills on problem types everyone has seen a thousand times, the fifth practice test of the month, and memorization past the base you need to judge whether an answer is plausible.
But until then, we will keep buying stickers, and every year there will be more things to cover.
Update (Nov 19, afternoon): At 8:30 AM today, the 10.5 second nationwide alert went off across Korea. In its midday briefing, the Ministry of Education reported 22 smartphones caught across nine provinces and zero confirmed cases involving miniature cameras or other AI hardware. Separate reporting from Maeil Business revealed that under the actual proctor guidelines, putting stickers over buttons was left voluntary for test takers, not mandatory. The siren caught the kids who forgot.