China's AI Thesis Checks Fuel a Gray Market in Rewriting Services

Students Pay to Lower AI Scores as Graduation Pressure Builds

A 5,000-word thesis can be rewritten for as little as 32 yuan, about $4.45, in China's growing market for services that promise to lower the amount of text flagged as generated by artificial intelligence. More customized changes using algorithms can cost up to 20,000 yuan. The business has expanded during graduation season as universities tighten checks on academic work.

Vendors on major shopping platforms such as Taobao advertise "manual modifications" and "24-hour online consultation." Their promise is to reduce a paper's artificial intelligence-generated content rate, commonly called its AIGC rate, so that it can meet a university's requirements.

Some Chinese universities require that rate to stay below thresholds ranging from 20% to 40%. Others use the score as a reference and leave the final judgment to a supervisor. For students approaching graduation, a number produced by a detection system can mean hours of revisions and uncertainty about whether their work will pass.

The services have also exposed problems with the checks themselves. Students describe receiving sharply different results from different detectors, while some rewritten papers emerge with fragmented sentences, awkward language and broken arguments. An industry selling relief from automated scrutiny is often using more AI to deliver it.

As of June 13, 2025, the debate reaches beyond whether students should use tools such as ChatGPT. It concerns how universities judge independent work, how much confidence they place in software, and what happens when students pay unfamiliar vendors to make a detection score fall.

What the Vendors Sell

AIGC rate reduction services offer to rewrite or modify academic text so that detection systems judge it to look more like human writing. Vendors target systems used by Chinese universities, including the VIP Paper Check System and CNKI.

Some sellers claim to have proprietary databases and algorithms that identify words or sentence structures likely to trigger a detector. They then replace those features. The promised result is a lower suspected AIGC rate, achieved by changing the language and structure of the paper.

The label "manual modification" does not necessarily describe how the work is done. Much of the process still relies on AI. Vendors typically consult academic databases, compare content and use algorithms to adjust vocabulary and sentences. The service can therefore involve one automated system rewriting text to satisfy another.

Prices vary widely, from the 32 yuan offer for a basic thesis of 5,000 words to charges reaching 20,000 yuan for more customized modifications using algorithms. Sellers often provide little clarity about pricing or their educational qualifications before a student contacts them.

Many vendors ask customers to move conversations and payments from shopping platforms to private channels such as WeChat. That makes transactions harder to trace and can make it more difficult for students to collect evidence when seeking a refund or making a complaint.

University Rules Create Demand for Lower Scores

Chinese universities have introduced AI rules in response to concerns about academic fraud and the misuse of generative tools, including ChatGPT, Baidu's Ernie Bot and Alibaba's Tongyi Qianwen. Some require students to disclose their use of AI as well as meet limits on suspected machine-generated text.

The rules differ between institutions. A threshold that must be met before a thesis passes places a different kind of pressure on students from a score used to help a supervisor review the work. In either case, students have an incentive to reduce the percentage that a detector flags.

Luo Xueming, a senior expert at the Guangdong Modern Urban Industrial Technology Research Institute, describes the resulting demand as "pseudo-demand." His assessment connects the market for rewriting services to the introduction of the detection systems themselves.

The business extends beyond lowering AIGC rates. The broader gray industry also includes ghostwriting and the creation of PowerPoint presentations. Unclear academic boundaries around polishing with AI, plagiarism and adaptation have given these services room to operate.

For universities, the difficulty is deciding where useful assistance ends and misconduct begins. A student may use an AI tool to improve a sentence or to generate a substantial part of a thesis. Both involve AI, but they give supervisors different questions to examine about the student's own contribution.

A Student's Revisions Bring a Different Problem

Wen Suyu, a senior at a university in Guangdong, spent hours revising her dissertation to reduce its AIGC rate. Her changes brought little improvement in the score. Under pressure, she turned to a rate reduction service.

The service lowered the detected rate, but the revised writing sounded unnatural and robotic. Her experience captures a practical cost of the market: a paper can receive a more acceptable score while becoming harder to read.

Thousands of students across China share similar anxieties and advice on social media platforms such as Xiaohongshu. Their discussions describe a process that can consume substantial time without producing predictable results.

Common approaches include changing repeated sentence structures and transition words associated with AI writing, using several AI tools to vary the style, and manually editing generated drafts to add a more personal voice and irregularities.

Students also report differences of more than 50% in the AIGC rates assigned to the same text by different platforms. Such variation leaves them trying to satisfy a particular system while also keeping their writing coherent and academically appropriate.

Poor revisions create another burden. Fragmented sentences, unprofessional wording and logical inconsistencies can require further editing, even after a vendor has delivered the lower score it advertised.

How AI Detectors Reach Their Judgments

AI content detectors examine features such as sentence structure, patterns of wording and vocabulary diversity. They use those features to estimate whether text resembles writing produced by a machine. The resulting suspected AIGC rate is an assessment generated by the system.

CNKI and the VIP Paper Check System claim to identify text produced by major AI models. Their use has made the workings and reliability of automated detection an increasingly consequential issue for students and supervisors.

Zong Chengqing, a researcher at the Institute of Automation of the Chinese Academy of Sciences, explains that AI text is produced according to mathematical rules calculated by large models. Detection systems use similar mathematical analysis to look for the statistical patterns associated with machine-generated writing.

In simple terms, the detector searches for patterns in the words, rather than watching the writer produce them. Some experts put detection accuracy above 80%, but results in actual use have been less consistent.

A false positive occurs when a detector flags human writing as AI generated. That can send a student back to revise work they wrote themselves. Differences between systems add another complication, because each may assess the same passage differently.

An Official Warning Against Overreliance

Science and Technology Daily, an official publication of China's Ministry of Science and Technology, has warned against excessive reliance on AI detection. It highlighted cases in which classic essays and original research were incorrectly flagged as generated by AI.

In one widely cited example, a Chinese essay written 100 years ago received an AI-generated rating of more than 60% from a popular detection tool. The example illustrates how writing that predates today's generative systems can still contain patterns a detector associates with them.

A Science and Technology Daily editorial described the reliance on automated checks in these terms:

We aspire to eliminate the negative effects of AI with clear-cut solutions, but using AI to detect AI is essentially still a form of technological superstition.

Some universities in China and abroad have reconsidered or abandoned AI detection tools amid these concerns. Their response places greater weight on faculty judgment and a broader evaluation of academic work.

That approach gives supervisors a central role in examining the research, the argument and passages flagged by software. The task is to assess the student's work as a whole while deciding what revisions or explanations are needed.

September Rules Will Require AI Content Labels

China is also preparing a wider regulatory response to AI-generated content. Rules scheduled to take effect in September 2025 will require clear labeling of generated text, images, audio and video.

The Cyberspace Administration of China, together with other government agencies, has mandated both explicit and implicit markers. Explicit labels are visible to users. Implicit markers, including digital watermarks embedded in metadata, carry identifying information within a file's accompanying digital data.

The measures include penalties for failing to comply and form part of a broader campaign against misinformation, academic fraud and online deception. The main requirements cover several stages of producing and distributing content:

• Visible labels and implicit digital markers for AI-generated content.
• Compliance checks by online service providers before publication.
• Record keeping by service providers and users who post AI-generated content.
• Strict prohibitions on removing or altering AI content labels.

Enforcement faces technical difficulties. Real time applications and content moving between platforms complicate oversight. Watermarking and metadata have limitations, and labels can be removed or altered, making it harder to keep identifying information attached to content as it circulates.

The timing of the academic and regulatory developments is important:

• During the 2025 graduation season, demand for AIGC rate reduction services has grown alongside university checks.
• As of June 13, 2025, students and supervisors are navigating differing university rules and disputed detection results.
• In September 2025, China's new AI content labeling rules are scheduled to take effect.

Academic Integrity Requires More Than Editing

Li Juan, an associate professor at Central South University's Law School, estimates that 60% to 80% of students have used AI tools in their theses this year. The uses she describes range from refining sentences to composing large portions of the work.

Li acknowledges that AI can improve language quality. She also points to the danger of AI hallucinations, meaning convincing information generated by a model that is false. In academic writing, that can include invented statistics or legal cases.

A polished sentence can therefore still contain a factual error. Supervisors must review flagged content and guide students through revisions, while also checking the substance of arguments and evidence.

Chu Zhaohui, a researcher at the China National Academy of Educational Sciences, advocates putting more effort into academic ethics education:

A practical approach would be to strengthen academic ethics education among students.

His approach centers on independent thinking, research skills and integrity. Those skills help students understand their responsibility for the work they submit, including material they have asked an AI system to generate or revise.

The rapid development of AI has moved faster than comprehensive legal and ethical frameworks. Oversight also crosses several authorities, including education, market supervision, internet security, and science and technology. Clear boundaries between acceptable assistance and academic misconduct remain unresolved.

Students Face Financial and Academic Risks

Students purchasing rate reduction services face risks beyond an awkwardly rewritten thesis. Complaints include difficulty securing refunds, poor quality modifications and vendors becoming unreachable after payment.

Private conversations and payment arrangements that are difficult to trace can leave customers with little evidence to support a complaint or legal action. These practices weaken consumer protection and contribute to mistrust and dissatisfaction.

The same student may face several pressures at once: a deadline, an AIGC threshold, inconsistent detector results and a seller promising a quick solution. Paying for revisions can add a financial loss to the original academic problem.

Calls for greater oversight and clearer ethical standards reflect both sides of the issue. Students need understandable rules about AI assistance, while universities need ways to assess independent work that account for errors and variation in detection tools.

China's forthcoming labeling requirements will add obligations for identifying generated content. Within universities, supervisors still have to judge research quality, factual accuracy and the student's contribution. A lower detection score leaves those parts of the assessment to be completed.

Key Points

• China's graduation season has fueled a gray market promising lower AIGC scores for academic papers.
• Prices range from 32 yuan for a basic thesis of 5,000 words to as much as 20,000 yuan for customized modifications.
• Some universities set thresholds ranging from 20% to 40%; others use detection scores to inform supervisors' decisions.
• Services often use AI despite advertising manual editing, and students report awkward language and damaged arguments.
• Detection systems can produce false positives and sharply different results for the same text.
• China's AI content labeling rules are scheduled to take effect in September 2025.
• Named educational researchers call for stronger ethics education and a central role for supervisors in evaluating academic work.
• Private transactions, refund difficulties and unreachable vendors expose students to consumer risks.