China's AI apps shut down image recognition during gaokao

During China’s grueling gaokao exam week, construction sites shut down to keep the streets quiet. Now, cloud-based AI models are getting muted too.

On June 9th, Bloomberg reported that top Chinese AI chatbots temporarily disabled key features to stop students from cheating. Alibaba’s Qwen and ByteDance’s Doubao cut image recognition from answering test questions, while Tencent’s Yuanbao and Moonshot’s Kimi completely shut down photo-recognition during testing hours. The viral DeepSeek chatbot also blocked services during specific times to protect exam integrity, according to The Guardian.

The exam kill-switch

If you asked Kimi or Yuanbao to analyze an image during the blackout, the apps were pretty blunt about why. The chatbots told users their tools were paused to "ensure the fairness of the college entrance examinations."

None of the tech companies made big public announcements. Word of the sudden shutdown actually spread directly from students chatting on Weibo.

More than 13.3 million students sat for the gaokao between June 7th and 10th. Phones and laptops are already strictly banned inside Chinese exam rooms, but disabling AI vision features acted as an extra cloud-level safety net just in case someone slipped a device past security.

Going analog to beat the bots

Cheating via chatbot has become a headache for schools everywhere, not just in China.

While Chinese platforms are flipping software switches, American classrooms are going fully analog. The Wall Street Journal reported in May that sales of old-school paper "blue books" have surged across US universities over the past two years as professors abandon digital submissions to keep students honest.

Whether it’s pulling the plug on server features or forcing everyone back to pen and paper, schools and platforms are scrambling to patch the same loophole.

Why it matters

For regular users, losing photo-recognition for a weekend is a minor inconvenience. Nobody is canceling their subscription over a three-day exam blackout.

For the AI industry, we think this is a smart, coordinated defensive play. In high-stakes environments, waiting for an AI cheating scandal to hit the news cycle usually brings harsh regulatory backlash. By quietly pulling the plug on sensitive features beforehand, Chinese tech giants showed they can self-police and align with public priorities without being forced to.

That said, hitting an emergency off-switch every June is basically putting a band-aid on a broken bone. Banning image processing stops casual snaps, but text-based queries can still slip through. More importantly, as small on-device models get smarter and run locally without an internet connection, remote cloud blackouts won't work forever. Eventually, how we test students will have to evolve, because you can only unplug the cloud for so long.