This year's graduates are busy drawing a clear line between their theses and AI
Many experts said that AI detection is a helpless move by universities against the backdrop of an imbalance in the ratio of teachers to students and the absence of tutor guidance. Currently, various types of AI detection still have problems such as vague judgment criteria, inconsistent detection results among different systems, and widespread misjudgments. It is a sub - optimal choice, and its positioning as an early - warning tool should be clarified.
As the graduation season approaches, "AIGC rate (probability of artificial intelligence generation)" and "tutorials on reducing AI (artificial intelligence) content" have become hot topics on the Internet this year regarding graduation theses.
At Huazhong University of Science and Technology, graduate student Wang Tian was shocked to find that at least five or six classmates around him used AI tools like Doubao to complete their graduation theses, and the suspected AIGC rate in the CNKI system detection was very low, even close to zero. However, his own thesis, written word - by - word, had an AI rate as high as 36% in the first school - internal detection system, 16% higher than the school - stipulated range and far higher than his original estimate of 5%.
Cheng Chunfei from Zhengzhou University spent nearly 100 yuan on the Lixin Zhijiao thesis detection system to minimize the AI rate. Cheng Lu from Sichuan University used the platform's free detection times and paired them with AI de - weighting instructions to optimize the text, controlling the AI rate within 5%.
This is not an accidental individual phenomenon. As AI detection of graduation theses is incorporated into the university review process, the operation process of "writing with AI first, then testing with AI, and finally reducing AI content with AI" is becoming the new norm for many students when submitting their theses.
According to the "Insight Report on College Students' AI Usage Behaviors and Mindsets in 2025" jointly released by China Youth Daily, China Youth Media, and the social platform Soul App, 65.9% of the surveyed students turn to AI first when encountering problems. AI has surpassed traditional channels to become the preferred way to obtain information. The dependence of senior students is more obvious, with the daily multiple - use rate of senior students reaching 21.2%, far higher than the 8.0% of freshmen. 47.1% of students admitted that they "can't do without AI".
During the interviews, many experts said that AI detection is a helpless move by universities against the backdrop of an imbalance in the ratio of teachers to students and the absence of tutor guidance. Currently, various types of AI detection still have problems such as vague judgment criteria, inconsistent detection results among different systems, and widespread misjudgments. It is a sub - optimal choice, and its positioning as an early - warning tool should be clarified.
Su Yu, a professor at the School of Computer Science and Artificial Intelligence of Hefei Normal University and the developer of the "Yijianzhiwen" intelligent auxiliary system for degree theses, said that universities need to establish a complete and reasonable AI detection system. First, they should implement hierarchical management and define the scope in which students can use AI. Second, the usage methods and detection techniques should be traceable and explainable. Students can submit attachments on AI usage to explain which parts of the thesis were assisted by AI, and the detection system should also disclose the detection logic and standards for determining AI technology. Third, an appeal channel should be opened. If students believe that a certain paragraph of their text has been misjudged, they can apply for a re - check through their tutors and relevant university departments.
College Students' Responses
To avoid exceeding the AI detection standard, Cheng Chunfei adjusted her way of writing the graduation thesis. For her, a student majoring in English, a full - English manuscript is more likely to be detected with a high AI rate. Therefore, when writing the thesis, she only used AI to sort out some writing ideas and find literature sources. Then she read representative articles, master's theses, and books in the relevant field by herself. Finally, she re - organized the content in her own language based on these readings.
Cheng Chunfei said that this approach is more like using AI to assist in providing ideas rather than directly copying.
During the process of repeated detection and modification, she also found a way to reduce the AI rate for the second time: after getting the detection report, she manually modified fragmented and awkward sentences that did not meet the thesis language specifications. For large paragraphs of text suspected to be generated by AI, she used large models such as DeepSeek and Doubao to rewrite them. By customizing prompt words, she simplified long and complex sentences and optimized the writing style, using AI to inversely reduce the proportion of AI in the detection.
Cheng Chunfei said, "Currently, on social platforms such as Xiaohongshu and WeChat official accounts, there are various tutorial prompt words specifically used to reduce the AI rate everywhere."
Image source: Xiaohongshu
Many students also try to reduce the risk of exceeding the AI standard by selecting small - sample cases. Xian Jun, a senior student at Renmin University of China, said that when writing micro - quantitative and data - analysis theses in the finance major, many people choose niche samples such as rural banks for research. Since the repetition rate of such cases is low, it is naturally difficult to be judged as AI - generated.
Currently, the tolerance standards for AI - generated content in theses vary among universities. According to the public information from the Academic Affairs Office of Sichuan University, the proportion of AI - generated content in liberal arts graduation theses should not exceed 20%, and the upper limit for science, engineering, and medical theses is 15%.
Cheng Chunfei said that Zhengzhou University, where she is studying, has uniformly set the AI rate standard at 40%. Xian Jun also said that the school has not issued an official document yet, but the college has notified in the class group that the AI content in the thesis should be controlled within 10%.
Even after making various preparations in advance, students still face difficulties such as limited free detection opportunities in school - internal systems and relatively high costs for paid detection when revising their theses.
The school - internal system of Gao Lu's school uses the CNKI detection model and only provides two free detection opportunities. Her thesis requires the AI rate to be controlled within 15%. "The cost of a single CNKI detection for a 20,000 - word manuscript is 40 to 50 yuan."
Two of her roommates spent 80 yuan and 100 yuan respectively on VIP and CNKI for de - weighting and AI - reducing detection. Formal platforms such as CNKI, Wanfang, and VIP in the market all charge by the number of words, with the unit price ranging from 2 to 10 yuan per thousand words. Most universities also directly purchase services from third - party plagiarism - checking platforms such as CNKI, VIP, and Gezida.
To save money, many students will first take advantage of the free trial rights of platforms. For example, platforms such as PaperPass, PaperPure, and PaperYY have simultaneously launched AIGC detection and AI - reducing services, providing 2 to 5 free trials per day. The reports also mark the suspected AI paragraphs in different colors.
Gao Lu often uses the free detection times of the platform to conduct self - checks repeatedly until the text risk is reduced to within the qualified line before daring to submit it to the school - internal system. She said that after the actual detection, she found that the school - internal review scale is more lenient than expected. "As long as the whole text is not directly generated by AI, it can basically be below the standard without affecting the subsequent process."
The high pricing of formal platforms has also given rise to the gray service of low - cost AI - reducing agents on e - commerce platforms. Channels such as Pinduoduo, Xianyu, and Taobao are filled with various thesis detection and manual de - weighting packages. Taking the search results of the product keyword "Gezida" as an example, the overall service price ranges from 10 yuan to 50 yuan.
Xian Jun's roommate once had a bad experience: in the first AI detection, the rate was close to 30%. He spent more than 100 yuan to buy the highest - level manual AI - reducing service on the platform. After the modification, the sentences were not only awkward and incoherent, but the rewritten content was completely out of line with the professional logic. "It was simply for the sake of reducing the rate." He had to pay the customer service again for a second - round modification, and finally reduced the AI rate to about 9%.
In fact, such third - party services of unknown origin may also hide the risk of thesis privacy leakage.
Su Yu said that compliant detection platforms need to have corresponding data security and personal information protection mechanisms, including security verification for file uploads, role - based access control, literature download permissions, automatic invalidation of temporary links, and desensitization and backup of sensitive information. However, small - business platforms without formal service qualifications and vague privacy terms pose extremely high risks. The theses uploaded by students may be inappropriately retained, misused, or even used for other improper purposes.
Unstable Systems
During various AI detection processes, Xian Jun's biggest confusion is: on what basis does the system judge that a text is generated by AI?
Su Yu answered this question from a technical perspective: the so - called "AI flavor" refers to a series of characteristics in the language expression of text generated by large models. When generating text, large models usually predict subsequent content based on the context. The output content is affected by factors such as prompt words, training data, and generation parameters, and it is easy to form fixed and templated expressions, such as well - structured but mechanical sentence patterns and repeated use of conjunctions. At the same time, the essence of large models is probability - based output, and the viewpoints they generate are prone to factual errors or hallucinations. The density distribution of terms in the text is balanced but not in - depth, and the conclusions are generally vague.
According to Su Yu, currently, the mainstream AI detection tools mainly rely on two technical approaches to distinguish writing styles. The first is the statistical feature analysis method: when generating text, AI predicts and generates subsequent word elements based on probability distribution, and some text may exhibit relatively smooth and regular statistical features. Some detection tools also use perplexity, burstiness, text complexity, etc. as auxiliary features for judgment.
The second is the classifier model: developers use text written by humans and text generated by AI as training data to train a binary classification model, allowing the model to learn the subtle differences in semantics, syntax, and structure between the two types of text, and then use it to judge new text.
However, while the technical logic is one aspect, students still have to face a system full of uncertainties.
Wang Tian's thesis had an AI rate as high as 36% in the first school - internal detection, 16% higher than the school - stipulated range and far higher than his estimated 5%. The 5% mainly came from his use of AI tools to polish some original text.
Wang Tian told reporters, "Many of my friends wrote their theses with Doubao. I asked them if they had made any modifications themselves, and they all said no. But their detected rates were very low, even close to 0%. I wrote my thesis very carefully, but it was judged to have a high AI content. It's obviously a misjudgment."
In order to lower the detection value, Wang Tian had to spend two days modifying the manuscript. He even sacrificed some of the thesis structure, text smoothness, and logic to reduce the AI rate from 36% to 1.3%. He reduced the proportion to an extremely low level because the results of various free websites generally tend to be inflated.
Wang Tian said, "The suspected AI rate detected is often more than 20 percentage points higher than that of the school - internal official detection system."
Gao Lu also confirmed the large detection gap between different platforms: "I got a 15% rate on the free platform, but when I re - checked it on the school's official system, it was actually only 4% to 5%."
Wang Tian also conducted a set of comparative tests. He uploaded the original texts of four core journals in his major to PaperPure for detection. The AI proportion of many articles exceeded 20%, and two or three of them even exceeded 40%. This also made him question the rationality of the existing detection standards.
Wang Tian said, "When using the free platform, only half of the text marked as suspected AI paragraphs was actually polished by AI. The rest was all my original work, and the marked positions would change every time I checked. The most exaggerated thing is that even the acknowledgment part was judged as AI - generated by the system."
Photo provided by the interviewee
Not only the main body of the argument, but also fixed and standardized content is prone to misjudgment. All the fixed - template texts such as the description of the quantitative questionnaire research and greetings in Wang Tian's thesis were marked red, and the content could not be changed. Gao Lu's experimental steps were also judged as AI - generated, so she had to rewrite the original content with the help of large models to avoid being marked red.
Photo provided by the interviewee
In addition, the algorithms of different detection platforms are independent, and the judgment standards are not universal, which also makes it difficult for students to have a unified direction for modification.
Gao Lu said, "The marked - red parts in PaperPass, PaperYY, and the school system are completely different. Due to different judgment algorithms, the parts modified and optimized by off - campus tools may have no effect in the school - internal detection."
Helplessly, Wang Tian could only delete and replace some non - fixed professional expressions and logical conjunctions, disrupting the structure of some originally smooth texts. After the modification, although the AI rate met the standard, "it doesn't sound like normal language anymore."
Regarding whether universities should incorporate AIGC detection into the assessment reference standards for graduation theses, many interviewed students said that AI detection has a certain degree of rationality, but the existing technology and detection standards still have certain limitations.
Everyone unanimously suggested that schools should increase the number of free detections from the current 2 to 4 - 6 times. This can not only relieve students' anxiety about detection but also reduce the economic burden caused by additional detections.
In Su Yu's view, AI detection still has certain value. University teachers have limited manpower and energy and it is difficult for them to carefully review all graduation theses word by word. AI detection can quickly filter out the content suspected to be written by AI. On the one hand, it prompts students to actively verify the text source and AI usage, and supplement explanations or make modifications according to school regulations. On the other hand, it also provides a reference for teachers to review theses. "In essence, AI detection undertakes the