Major Chinese IT companies, including ByteDance, Tencent, and Alibaba, temporarily restricted access to their generative AI services during the national entrance exams, Gaokao, which took place from June 7 to June 10, 2025. These measures are aimed at preventing the use of artificial intelligence for cheating during exams that determine further admission to universities for over 13 million high school graduates.
This is reported by Finway
How the AI Service Restrictions Worked
Users reported that ByteDance’s Doubao, Tencent’s DeepSeek and Yuanbao, Alibaba’s Qwen, and Moonshot’s Kimi disabled the ability to answer questions and analyze images. Some applications informed users of the temporary unavailability of features specifically due to the exams. Even attempts to bypass the restrictions by rephrasing queries were unsuccessful — the AI stubbornly refused to respond. This caused frustration among students who were used to utilizing AI for study preparation, and sarcastic comments regarding the reasons for the service blockades appeared on social media.
The authorities intensified control, including biometric measures and communication jammers.
Unprecedented Security Measures During Exams
In addition to digital restrictions, the Chinese authorities significantly increased oversight of exam administration in physical classrooms. In many regions, AI-based systems have been implemented to detect suspicious behavior among students. Surveillance cameras record participants’ actions, and the data collected is analyzed after the tests are completed. Biometric identification, thorough screening of digital devices, and the use of devices to block radio signals have been introduced. Access to the venues where exams take place is strictly regulated, and in some cities, traffic has even been restricted and mass events postponed to ensure students arrive on time for their exams.
It is worth noting that the security of Chinese AI models also remains a subject of international discussion. Anthropic’s CEO, Dario Amodei, stated that the DeepSeek R1 model, developed by the eponymous Chinese startup, showed the worst results in data protection tests.