When China’s AI Platforms Go Quiet for the College Entrance Exam: A Textbook Case of Security Theater

Every June, more than ten million Chinese students sit the gaokao, the national college entrance exam. It is the single most important exam in the country. Taken on the same days nationwide, it is a one-shot test, and the score decides almost everything that follows: which university a student enters, and for many, the trajectory of the rest of their life. There is no holistic review, no essays or interviews, no meaningful second chance. One exam, one number. It is hard to overstate what rides on it, which is why anything that touches its fairness becomes a national flashpoint.

This year, as it approaches, Doubao, Yuanbao, Kimi, DeepSeek, and other major AI platforms announce that during exam hours they will temporarily disable photo-based question solving, problem analysis, and essay generation. Everyday chat and general queries are unaffected.

The public applauds, reading it as corporate responsibility, a gesture in defense of the exam’s fairness.

I want to offer a contrarian view. As a gesture, the move is flawless. As an anti-cheating measure, it is close to meaningless. And what the platforms are so visibly guarding against is not student cheating at all.

---

They closed an entrance that does not physically exist

Start with what the move actually prevents.

The gaokao examination room has long been a sealed physical environment. Candidates may not bring in any electronic device. Even the proctors’ phones are collected and held. The rooms have full radio-signal jamming. Which means that during the exam, in that time and that space, a phone cannot get in, and even if it did, it would have no signal.

So what real obstruction does disabling an AI’s photo-solving function add, inside that room? None. Because the entrance of “using AI to photograph and solve a question” does not exist there in the first place. For a tool that cannot be brought in and cannot connect, whether or not you switch off one of its functions makes zero difference to cheating inside the room.

Some will object: the truly high-tech cheating is not a student pulling out a phone in the room, but using a concealed micro-device to transmit the paper outside, where an off-site team feeds it to an AI, gets an instant answer, and relays it back. The chain sounds sophisticated, but it has one physical precondition it cannot get past: the questions must first get out of the room. And the room’s signal jamming cuts off not just cellular networks but all wireless transmission in that band. Micro-cameras and hidden earpieces depend on exactly that wireless signal. If the questions cannot get out, the AI outside, however powerful, has nothing to work with.

Suppose further that some ring really had the technology to break through state-grade jamming. That would mean a large criminal organization with professional equipment and serious counter-surveillance capability. Against cheating at that level, switching off the AI function is the worst possible move. Because during the gaokao, if someone uploads real exam questions to a platform for answers, the real-name data, timestamps, activity logs, and question images left in the back end are ready-made hard evidence. Following that data trail to make arrests is far more precise than scanning a sea of candidates by eye. To actually counter this kind of cheating, what a platform should do is retain the data and assist the investigation, not sever, with its own hands, a line that would have left perfect evidence.

In either case, what actually secures the exam room is physical isolation and entry screening, not a temporary blackout of every AI on the network.

---

A textbook case of Security Theater

The security expert Bruce Schneier coined a precise term for this: security theater.

It refers to measures that appear to improve security, that let the public visibly feel someone is in charge, but that add nothing to actual safety. Their real output is not security but a feeling of security, and a posture of “I have done my duty.” Certain elaborate airport-screening steps that stop no real threat are the classic example.

AI platforms going quiet during the gaokao is a textbook case of security theater. It gives the public a response, and a sense that the platforms are responsible and the exam is safe. But what it shuts off is a useless channel that no cheater would, or could, actually rely on.

---

What the platforms guard against is not cheating, but a PR crisis

If it is useless, why do the big firms execute it in lockstep?

Because what they are doing is, in essence, defensive self-protection. In the current regulatory climate, the absence of an explicit legal obligation does not mean the absence of real exposure. If the law actually required platforms to prevent cheating, it would be a written, hard compliance rule, not a soft press release carrying the words “voluntary industry action.”

If, during the gaokao, it came out that some ring had used a particular company’s AI to generate a top-scoring essay, then even if the cheaters bore full legal liability, that platform would be drowned in public outrage and could face severe regulatory correction. So when a platform disables a function, what it is guarding against is not cheating in the exam room, but its own entanglement in a cheating scandal. What it is isolating is its own commercial risk, not the exam’s risk of fraud.

This motive shows most clearly in the platforms’ selective action.

If a platform were truly committed to exam fairness, then cheating is cheating wherever it happens. The civil service exams, the provincial exams, the postgraduate entrance exam, the various professional qualification exams all decide countless people’s fates, and all face the same potential motive to cheat. Why do the platforms loudly suspend functions only during the gaokao, while ignoring the thousands of other exams across the year?

Because the gaokao is where society’s spotlight falls. It is the most sensitive in public opinion, the one whose blowback is fiercest if something goes wrong. That the platforms perform only for the gaokao shows that what they are optimizing is not the scenario with the highest risk of cheating, but the scenario with the highest risk of blame to themselves.

---

They never had the ability to be the gatekeeper

The deeper reason is that, technically and structurally, platforms simply lack the conditions to be a gatekeeper of exam fairness.

In time, they cannot know the full map of exams. Across a year there are thousands of exams nationwide, differing in schedule and subject. A platform cannot, and has no obligation to, track the precise timing of each one. In content, they cannot discern a question’s intent. Feed a hard physics problem into the chat box, and the platform has no way to tell whether the screen on the other side belongs to a student revising at home or a cheating ring inside an exam room.

Put these two blind spots together and you get a knot that cannot be untied. For a platform to truly carry out “preventing cheating” across all exams, it would have to do both impossible things at once.

Follow the logic to its end: if a platform had to answer for every exam, and it can distinguish neither the timing nor the intent, then its one foolproof option would be to disable the answering function permanently. To destroy the tool itself in order to keep it from being misused, this reductio’s endpoint exposes the fatal feature of security theater. It does not scale. A real line of defense, like signal jamming, can cover every exam room indiscriminately. Security theater can only be performed under the brightest spotlight, the gaokao, for everyone to see.

---

A useless move that blurred a clear allocation of accountability

The original chain of accountability was perfectly clear.

Cheating in the gaokao has long been a criminal offense. Deterrence comes from criminal law. Securing the exam room comes from the exam organizers, through screening and physical jamming. The consequences of cheating, and of organizing it, are borne by those involved.

In this chain, every actor is defined and capable of bearing its accountability. And the platform’s only sensible role here is not the gatekeeper before the fact, but the evidence source after it. Deter with criminal law, isolate with physical means, trace and obtain evidence with the platform’s data. That is each part in its proper place.

When a platform disables its answering function, it looks like doing a bit extra. In fact it uses a seemingly accountable move to package a self-interested act of risk avoidance. The cost of this move lies not only in its uselessness, but more in the fact that it muddied water that had been clear. It gives the public the illusion that preventing cheating has become the tech companies’ accountability, quietly diluting an allocation of accountability that was already clear. If nothing goes wrong, everyone is happy and the firms collect a round of applause for social responsibility. If something does, this blackout still stops nothing. What it guards against was never cheating, only the platform’s own risk of being blamed. And that risk should never have been carried by an AI platform that has neither the obligation nor the ability to stand guard.