Chinese students use AI tools to outsmart AI detectors

Chinese universities are increasingly implementing strict policies to limit the use of artificial intelligence in academic work, which is leading to a paradoxical trend: students are now using AI tools to outwit the very detectors designed to detect them, writes Rest of World.

The restrictions began with universities, including leading institutions such as Fuzhou University and Sichuan University, setting limits of 15% to 40% AI-generated content in their theses. Failure to meet these requirements can result in serious consequences, including delayed graduation or expulsion.

Xiaobin, a senior studying German literature, found herself in this situation just a week before her thesis was due. Although she had mostly written her 16-page paper herself and used ChatGPT and DeepSeek for minor revisions, a pre-check using the school’s recommended platform flagged half of her work as AI-generated. "The whole process felt absurd to me. … [I feel like] an innocent person being dragged to the gallows," Xiaobin told Rest of World.

Universities largely rely on tools developed by large academic technology companies such as China National Knowledge Infrastructure (CNKI), Wanfang Data, and Chongqing VIP, which traditionally offer plagiarism detection services. However, students and even some faculty members have expressed concerns about the accuracy and fairness of these AI detectors. Many students who claim to have used AI minimally or not at all report that their papers have been falsely labeled as having a large amount of AI-generated content.

Adding to the complexity, some of these same platforms, including Chongqing VIP and PaperPass, allegedly offer "AI reduction" services to help students bypass the very checks they are conducting, creating a profitable, if ethically questionable, cycle.

Students from at least eight universities have expressed confusion and frustration over the "sudden" implementation of the policy. Feeling immense pressure to get their degrees, many are turning to various services, both human and AI-driven, to change their work. These range from expensive "purely human transcription" services, some of which cost nearly $100, to cheaper AI chatbots that change vocabulary and syntax. The results have been mixed, with some services proving effective, while others introduce significant errors.

Dede, a college student in Fujian province, paid about 500 yuan ($70) to a tutor who promised to make corrections. Although the AI detector identified less generated content in her work after that, it became disjointed, with key terms misinterpreted. She gave an example of how "three knives," a traditional Fujian headdress, was sloppily changed to "three-bladed tools." Another student reported that the word "semiconductor" was changed to "0.5 conductor" by an AI tool that is supposed to help bypass AI detectors.

Faced with consistently high scores of AI-generated content, students are resorting to creative workarounds. Xiaobin, despite her extensive manual editing, found that her AI score stubbornly remained above 50%. Her final solution was surprisingly simple: replacing semicolons with commas. This made her sentences vague, but significantly lowered the AI detector’s score. "It feels like you’re being punished for writing too well," she commented. Xiaobin successfully graduated in mid-June with a thesis that her school recognized as 2% AI-generated.

The rapid adoption of these AI detection tools in China, despite their acknowledged limitations, contrasts with the ongoing debate in the United States and elsewhere about their reliability. Some Chinese professors are concerned about the unintended consequences of these policies. "The bigger issue is that these tools make students feel like using AI is something to be ashamed of," a communications professor from Shandong province said anonymously, comparing it to the past public avoidance of sex education. When something can’t be discussed honestly, it can’t be handled properly."

Some institutions, however, are advocating a more balanced approach. In May, Nanjing University issued a statement acknowledging the limitations of AI detectors and urging teachers not to rely solely on their results.