(SINGAPORE 2026.6.11) China’s new university rules for curbing AI-generated content in theses are intended to safeguard academic integrity, but they have instead engendered a gray industry of paid “AI content reduction” service and a debate over how scholarship should be evaluated in the age of artificial intelligence, Chinese media reported.
Across China, a growing number of universities now require graduating students to pass AI-content detection checks before they are allowed to defend their theses. Institutions set thresholds ranging from about 15% to 40% for AI-generated content. Humanities papers tend to have slightly higher allowable limits, while science, engineering, and medical theses often face stricter caps.
The policies are devised out of concern that students may rely excessively on generative AI tools when producing academic papers. However, implementation has quickly run into practical problems. Chief among them is the lack of consistency among AI-detection platforms, which often produce sharply different results when evaluating the same document.
According to Chinese financial newspaper National Business Daily (每日经济新闻), the same thesis can receive AI scores that vary by more than 20 percentage points across different systems. One platform may flag a document as nearly fully AI-generated, while another gives it a reassuring moderate or even low score. This has raised doubts about whether such tools can reliably measure originality at all.
Students and educators also say they face a paradox: writing that is clear, structured, and logically organized is often more likely to be flagged as AI-generated, while less polished or more conversational writing tends to pass detection more easily.
As a result, many students have begun to deliberately alter their writing style to avoid being rated as high on AI. Instead of improving clarity, they sometimes weaken expressions, introduce informal phrasing, or even add grammatical mistakes in order to lower their “AI rate.” In extreme cases, students instruct AI tools to rewrite their own work in a fragmented style just to get past the zany detection systems.
One student, Hu Yun, an English major at a university in southwestern China, said she was shocked to receive an AI-detection score of 65% for her thesis, which she had written entirely herself.” “To reduce AI rates, some students ended up doing something completely counterintuitive,” she said, describing how academic writings were intentionally degraded to satisfy whacky algorithms.
Liu Feng, a recent graduate, said she spent several days repeatedly revising her thesis after running it through multiple detection systems. “I spent four or five days staring at unfavorable AI detection reports and rewriting sentences,” she told NBD. “In the end, my thesis didn’t improve academically. Most of my effort had gone into trying to satisfy the system.”
Her university required AI-generated content to be below 30%. However, even sections she wrote on her own were flagged as suspicious by detection software such as Weipu (维普). Core analytical passages were sometimes marked as likely AI-generated, while simpler, less formal sections passed without issue.
A similar experience was reported by Chen Lu, a postgraduate student at a medical university. She noted that even a highly accomplished peer—whose research had been published in a prestigious Science Citation Index (SCI) journal—received an AI score of around 30% for a thesis based on his original work.
“These results feel ironic,” Chen said. “The more logical and refined a writing is, the more likely it would be marked as problematic.”
Adding to the pressure is the financial cost of repeat testing. Many universities provide only one free-of-charge attempt with their platforms. Each additional test must be paid for by students themselves.
Fees vary by platform. China’s largest academic database provider charges about two yuan (S$0.38) per thousand Chinese characters, meaning a 100,000-character thesis can cost nearly 200 yuan per scan. Students reported spending significant amounts on repeat checks and plagiarism screening, with some describing the process as financially and emotionally exhausting.
Meanwhile, a commercial service has popped up to meet student demand for lower AI scores. Online marketplaces such as Taobao (淘宝）and Xiaohongshu （小红书） are now filled with advertisements on help for “AI rate reduction”.
One company charges 300 yuan for a 13,000-word thesis and promises guaranteed reductions in AI detection scores. Others offer automated systems that rewrite documents using AI tools designed to evade detection algorithms.
The sellers claim high success rates, with some asserting that 90% of their clients can get their AI scores reduced to around 10%. One reportedly completed over 4,000 orders in a short period.
NBD journalists testing these services found that a document initially flagged at over 95% AI-generated was reduced to around 10–15% after processing. However, the reduction came at a cost: the resulting text tended to be more awkward, less formal, and short on academic flavour.
One platform, called “Writing Dog” (写作狗）， claims to be able to restructure formulaic writing while preserving technical content. However, experts pointed out that such systems likely rely on large language model APIs combined with prompt engineering designed to evade detection tools rather than genuinely improve academic quality.
According to one expert cited by NBD, the actual cost of processing a thesis may be only a few cents, but students pay hundreds of yuan for each service. “The profit margins may exceed those of traditional ghostwriting,” the expert noted.
Unlike conventional essay mills that produce full papers, AI-rate reduction services scale more easily with automatic rewriting. The “efficiency” has helped widen their customer base: not only students who are reluctant to write would come to them, but also diligent students whose work might now be unjustly flagged by detection systems.
This has raised legal and ethical concerns. According to some legal experts, such services facilitate circumvention of academic integrity systems and may constitute illegal business practices. However, crackdown is prevented by the fact that many sellers use coded language or euphemisms.
At the heart of all these controversies is a deeper question: whether AI-detection tools are suitable for assessing academic works at all, NBD underlined.
Many researchers and education experts argue they are not. Chu Zhaohui, a researcher at the Chinese Academy of Educational Sciences ( 教育科学研究院）said current AI detection tools are neither accurate nor reliable enough to determine originality. “There is no trustworthy AI detection system at present,” he said。
Others want AI detection to be used only as a supplementary reference, not decisive metric. Chen Yixin, a managing director at a consultancy firm, said final assessments should rely on human academic review rather than automated scoring systems.
In response to growing criticism, some universities have begun adjusting their policies. Institutions such as Nanjing University have proclaimd that AI-detection results should be treated as references rather than definitive proof of misconduct.
At the national level, the China Association for Degree and Graduate Education （ 中国学位与研究生教育学会 ）issued new guidelines in May that shift focus away from raw detection metrics. Instead, they emphasize transparency in the use of AI tools, documentation of research processes, and oral questioning during thesis defenses.
The guidelines reflect recognition that AI tools are becoming embedded in academic workflows rather than being external threats. As one professor, Hu Yanping of Shanghai University of Finance and Economics, noted, AI is likely to become a fundamental part of research and learning.
“The challenge is not to eliminate AI, but reforming how we evaluate academic value,” he said. In his view, Chinese universities remain overly focused on publication counts and formal outputs, rather than real-world impact or innovation.
He asserted that research should be judged by whether it generates meaningful contributions to knowledge or society. In some cases, he suggested, even heavily AI-assisted work could be valuable if it solves important problems.
“Conversely, work produced entirely without AI may lack substance or originality, ” he quipped.
