news

한어Русский языкEnglishFrançaisIndonesianSanskrit日本語DeutschPortuguêsΕλληνικάespañolItalianoSuomalainenLatina

When Shu Ran saw the red and yellow colors of graduation thesis in the AI ​​detection system, she felt like crying but had no tears. She asked herself in her heart: Do I look so much like AI?

In Shu Ran's 23,000-word thesis, dozens of paragraphs were detected as "medium risk" and "high risk". Shu Ran explained that this meant that although his graduation thesis was original, it could not meet the school's thesis review requirements because of the "high AI rate".

With the popularity of AI, many students will use AI to generate content when writing papers. According to the official websites of some universities, these schools have issued a notice on AIGC (generative artificial intelligence) detection for the 2024 graduates for the first time, clarifying that students should strictly abide by academic norms and academic ethics when conducting graduation projects (papers), avoid over-reliance on intelligently generated content, and ensure the independence and originality of graduation projects (papers).

The students interviewed all believed that AI detection can indeed prevent students from using AI to write papers, but the current AI detection function is not yet mature. "It is difficult to understand how it is judged to be AI writing, because even if it is written by humans or modified, it will be detected as AI writing again." Shu Ran said that using technology to punish technology will result in the opposite of the requirements of the paper.

For those students whose graduation papers were mistakenly judged to be written by AI, the machine gave them a number and they had no way to appeal. They could only modify the papers according to the machine's operating logic. An obvious paradox is that when artificial intelligence becomes more and more like humans, the sentences written by humans themselves also become more and more like artificial intelligence. At this time, the more AI is used to detect whether AI ghostwriting is used, the higher the possibility that their papers will be mistakenly judged to be written by AI ghostwriting.

From "medium risk" to "high risk"

At the end of April this year, Shu Ran, who was about to graduate from university, discussed with his supervisor and decided on the topic of his graduation thesis.

A month later, Shu Ran started writing the first draft of her thesis. During the period of writing the thesis, she had already entered a company for internship. After get off work every day, she racked her brains in front of the computer and wrote until midnight before taking a rest. During the two days of the weekend, she devoted all her time to writing the thesis.

Because she had discussed the structure of the paper with her supervisor in advance, Shu Ran's thoughts were clearer and smoother during the writing process. In addition, the supervisor put forward more detailed requirements for the format of her paper, including the table of contents, fonts, and lines of tables.

About half a month later, she finished the first draft of her thesis. According to the school's requirements, her graduation thesis had to be more than 20,000 words. After revising it over and over again, Shu Ran's thesis stayed at 23,000 words.

Another requirement of the school is that in addition to general plagiarism checking, students must use an AI tool called "Gezida" as a paper citation detection system, and the detection result must be "low risk".

Shu Ran found out through online search that this software can provide functions such as paper duplication check, format detection, format correction, paper management, etc. Paper duplication check mainly finds the repetition rate of the content in the paper by comparing it with the journal database.

Different duplicate checking systems have different algorithms and database sizes. For example, the rule for duplicate checking in CNKI is 13 characters, that is, 6.5 characters repeated in a row will be marked in red. The purpose of AI detection is to identify which content is generated by AI, and make judgments by analyzing language style, detecting duplicate content, grammar and logic.

One week after finishing the first draft of the paper, Shu Ran sent it to her supervisor. After the supervisor made suggestions for revisions, she "manually" revised everything from the title to the framework until she was satisfied. Two weeks before the defense, she submitted the paper to the AI ​​tool for testing.

Before Shu Ran submitted her paper, one of her roommates ran an AI test. The result showed that the paper had a 13% duplication rate, which indicated a "medium risk" in the AI ​​test. The roommate was not surprised by the result, and she told Shu Ran that part of her paper was written by AI.

At the time, Shu Ran thought it was "quite novel", but she thought that such a paper would have a high duplication rate, as her own papers were written word by word. When Shu Ran submitted the paper, the duplication rate in front of her was only 1%, but the AI ​​report also said "medium risk".

This means that she must lower the AI ​​rate from "medium risk" to "low risk".

Like Shu Ran, Sichuan university student Lin Tianle's graduation thesis was also judged to be written by AI. Two months before graduation this year, Lin Tianle suddenly received a notice from the school asking him to check the AI ​​rate. The day before the deadline for submitting the thesis, he learned that the school required students to use the VIP thesis detection system. The system shows that the AI ​​language model can be used to detect AIGC to quickly and accurately identify AI-generated content in academic texts. It can effectively identify whether the text is partially or completely generated by the AI ​​model. The detection result has nothing to do with the quality of the thesis, but only indicates the probability that the content fragments in the thesis are generated by AI.

Lin Tianle's school requires that the AI ​​rate should be lower than 30%. If the rate exceeds this, the paper will not pass the blind review and will not be able to graduate. After he uploaded the paper, the result was 37%. Lin Tianle said that the content marked as AI was written by him.

Before that, he had checked it on CNKI and found that the duplication rate was zero percent, so he was relatively confident about his paper. On the day when he was due to submit the first draft, his teacher suggested that he add some new content. He continued to revise it and finalized it in the afternoon. "I felt that I worked so hard, but I was still in a mess before submitting it, and the AI ​​rate was judged to be 37%. I was very angry."

Lin Tianle complained about the test results on social media. All pictures in this article were provided by the interviewees.

Because he had been recommended for graduate school, Lin Tianle had high standards for his thesis. Every morning at 9 o'clock, he went to the laboratory to clock in and write his thesis, and he worked until 8 or 9 o'clock in the evening, sometimes staying up until 11 or 12 o'clock. This lasted for a semester, and he wrote nearly 40,000 words.

He was originally full of confidence, and his doctoral senior also praised him for the rich content of his paper, the beautiful figures and charts, and the fact that the duplication rate never exceeded 3%. However, it was detected that it was written by AI, and Lin Tianle's mood hit rock bottom.

Has your AI rate come down?

When he first started revising his paper, Shu Ran followed the red and yellow prompts in the software and thought about the revisions himself. For sentences that were similar to other content, "it looked a bit like AI," he revised them word by word.

She majored in international economics and trade, and her thesis involved many conceptual explanations of professional terms. But what bothered her was that these fixed explanations were most likely to be marked in red by the AI ​​system, and she could not "create a new concept."

Her specific modification methods were to replace with synonyms, adjust the order of sentences, or use more "advanced" vocabulary to counter the "mechanical and low-level words and sentences given by AI." But she did not expect that the result of the retest would become "high risk."

Shu Ran was completely "destroyed" in her heart. She didn't know the algorithm and rules, she only knew that the software would mark different risk levels, red for high risk, yellow for medium risk, and green for low risk. In the first test results, there were 35 "medium risk" sections in her paper, but after the modification, it became more than 50 sections.

After Shu Ran revised the paper, it became "high risk".

After her paper was judged as "high risk", Shu Ran calmed down for a night. The next day, she began to search the Internet for ways to reduce the risk. At first, she tried several AI tools that could reduce the AI ​​rate, "but they were all useless."

Afterwards, she searched for some posts by college graduates sharing their experiences in reducing AI rates, saying that they should remove words such as "first", "secondly", and "and" from the paper, as well as change sentences and word order, etc. She used this method to revise the paper again, and the paper returned to "medium risk".

After a day of revision, she was back to square one. Shu Ran felt that this revision method was too slow, so she continued to search online. This time, she extracted the experience of others in reducing the AI ​​rate and summed up a "rule that she knew by heart", namely "complex to simple, say the opposite of what you mean, active to passive, simple to complex".

With only two days left before the final draft of the paper was due, Shu Ran and her roommates' papers were all judged to be "high risk" or "medium risk." "Some of them used AI to write, some copied from other people's papers, and some wrote their own." Shu Ran said that their test results were similar, with a low duplication rate, but a high AI rate.

We stayed up until 3 a.m. that night. The most common question among the people in the dormitory was, “Has your AI rate dropped?”

Although the school provided a free test opportunity, Shu Ran wanted to use it for the final draft, because once the test report came out, the system would directly submit it to the teacher. She spent more than 200 yuan to purchase four test opportunities.

She felt that the paper was like a physical examination. “All the examinations revealed diseases.” The report was very detailed. “Even a few wrong words would be marked in red.”

In the process of reducing the AI ​​rate, Shu Ran kept thinking about what his supervisor said: If the thesis fails, it will affect graduation. There will be spot checks after the defense, and if it fails, the consequences will be serious.

As the day was about to dawn, Shu Ran couldn't remember how many times she had revised the entire paper. After checking again, although there were still a few paragraphs on the report that showed "medium risk" and "low risk", the overall result was the three words "zero risk" that she was looking forward to.

She breathed a sigh of relief and couldn't wait to share her method of reducing the AI ​​rate with her classmates. She stayed up late to help her roommate's paper change to "low risk". "My roommate was so happy that he said you are the god of our dormitory."

But Shu Ran felt that she had also changed her roommate's paper to be "unsightly". In order to reduce the AI ​​rate, she, like other students, no longer cared about the fluency and smoothness of the paper content. However, in the end, she successfully passed the defense.

Like Shu Ran, Lin Tianle also searched the Internet for various methods to reduce the AI ​​rate. His classmates suggested that he use AI software to reduce the AI ​​rate, but he refused. He rewrote it three times and finally deleted it in anger.

But the AI ​​rate found in the paper was still 30.13%, and he felt that he might have to postpone his graduation. He remembered that night, almost all of his classmates stayed up all night to revise their papers because of the AI ​​rate.

Initially, the school notified students that their manuscripts had to be submitted before midnight of that day. However, due to the instability of the AI ​​detection system, many students failed to submit their manuscripts successfully, so the school had to postpone the deadline to 12 noon of the next day.

Lin Tianle's later experience in reducing the AI ​​rate was to use fewer conjunctions, more commas in sentences, and have multiple short sentences share a subject.

In order to be faster and more efficient, Lin Tianle deleted all the parts marked in red and rewrote them directly. After revising three times, there were still some parts marked in red in the paper, so he continued to rewrite it, and finally the AI ​​rate dropped by less than 7%.

The deadline was getting tighter and tighter, so he could only continue to delete the thousands of words marked in red. He knew that this would have a certain impact on the content of the paper. "It doesn't look as good as before, but there's nothing I can do. At least I can graduate successfully."

“Deliberately exposing human shortcomings”

When classmates asked Shu Ran about how to lower the AI ​​rate, she half-jokingly said that it was actually "deliberately exposing some small human shortcomings," such as writing in a less fluent manner.

The joy of successfully reducing the AI ​​rate did not last for a few minutes, and Shu Ran began to feel lost. The paper she had written with all her heart and soul, after being "patched up", finally became "fragmented" and did not achieve the goal she wanted.

This also made her think: when AI becomes more and more intelligent, and what it writes is "very advanced, very professional, and getting closer and closer to humans, and we use the AI ​​detection system to ask us to lower the AI ​​rate, isn't that making ourselves less human?" Shu Ran feels that this is a paradox.

Shu Ran’s paper was marked yellow by AI before it was revised.

After Shu Ran's paper was revised, it was no longer marked in yellow, but the sentences became incoherent.

Regarding the misjudged behaviors, the common question among the college students interviewed is, how does the program judge? "Many people use AI to make their spoken language more academic, but now they deliberately change academic language to more colloquial language to avoid detection," said Gu Siqi, a senior at a university in Zhejiang.

In December last year, Gu Siqi completed the opening report of her graduation thesis. In February this year, she started writing the first draft. It took her about two weeks to look up information, conceive and write the manuscript. She handed the first draft of the thesis to her teacher. After that, she revised it according to the teacher's suggestions, and an undergraduate graduation thesis was completed.

The school where Gu Siqi is studying also has new regulations this year. Unlike the paper version of the final version in previous years, students need to upload their papers through the CNKI system. After the final version was finalized, in May, the school notified the students that they would check the AI ​​rate. Gu Siqi thought, I have already written it, it won’t be changed again. "The teacher only said that if the paper is to be evaluated, the (AI rate) must be below 10%, but did not say how low the AI ​​rate of a qualified paper should be." Gu Siqi recalled.

After uploading the paper to the detection system, Gu Siqi found that a paragraph she wrote, "a very logical conclusion", was identified as written by AI. Although there is no specific ratio requirement, the teacher said to try to reduce the AI ​​rate of the paper. Before the reduction, the AI ​​rate of Gu Siqi's paper was 38%.

So she had to remove logical words, weaken the correlation between the previous and next sentences, or change the order of paragraphs, or compress several paragraphs into one. After the chaos, "it looked more like it was written by a machine." However, after a long time of modification, the AI ​​rate of the paper did drop.

However, Gu Siqi felt that the "highlights" in her paper eventually became "mediocre." "This made me feel very uncomfortable when I read the paper."

After she heard that AI could be used to reduce the AI ​​rate, she also tried using ChatGPT and other software. At that time, she wanted to remove the associated words, but the final result was "completely irrelevant" to what she wanted to write.

Gu Siqi feels that "there is a certain gap between AI logic and human logic. It will carry some human thoughts, but it is not as smooth as what we humans write, and it looks awkward." She said that AI can only grasp some keywords and identify whether the paper was written by AI. Therefore, using AI to change AI rate seems very contradictory.

After repeated revisions, Lin Tianle has lowered his "academic requirements." Forced changes to the logic of the article also make him feel "confused." "The conjunctions I have been using since I was a child can make the article more fluent, but now I can't use them. It seems that they have become exclusive to AI."

Although Lin Tianle's paper was finally rated as an excellent graduation paper, he still felt wronged and had nowhere to complain. "I didn't use AI to write it, so do I have to admit something I didn't do?" He was upset for several days.

Technology "sanctions" technology

Before and after a paper is produced and submitted, AI is used as an auxiliary tool, then as a system to detect AI ghostwriting, and then becomes a tool to reduce the AI ​​rate. "AI dominates everything," said Lin Tianle.

Lin Tianle's major is automation. Sometimes he would use AI to ask about the meaning of a certain code. "AI will give more patient and detailed answers than humans." He recognized that AI is a practical tool. However, he felt that the detection standard and accuracy of AI-based papers were not convincing.

Lin Tianle once told his thesis advisor about his confusion. The advisor felt that the detection system was not very smart at present, but the school had relevant regulations. Apart from encouraging him to speed up the revision, there was no other solution.

Lin Tianle's school conducted AI testing for graduation theses for the first time this year. Although the system is still in the public beta stage, it has already been used. "Didn't the Degree Law say that AI should be tested before, but the law eventually deleted the AI ​​clause. Because it can only be judged based on simple wording, and it is suspected."

He was referring to the "Degree Law of the People's Republic of China (Draft)" reviewed in August 2023. The draft clearly stipulates that if a degree thesis or practical results contains academic misconduct such as plagiarism, forgery, data falsification, and artificial intelligence ghostwriting, the degree certificate shall be revoked by the degree awarding institution after deliberation by the degree evaluation committee.

However, in the Degree Law, which was finally promulgated in April this year, AI ghostwriting was not included in the consideration of academic misconduct. "However, some schools have already started AI detection," said a teacher from a university in Chongqing. The introduction of the "AIGC (generative artificial intelligence) detection service system" by universities is mostly out of consideration for academic integrity, to ensure the originality of papers, and to resist the behavior of AI ghostwriting papers.

Notices from some schools show that AI tools can only be used for auxiliary tasks such as literature retrieval and data processing, and are strictly prohibited from being directly used in paper writing. However, different schools have different requirements for the proportion of AI-generated content.

While desperately trying to reduce the AI ​​rate, Gu Siqi had the same confusion: "Is the AI's judgment accurate and reliable?"

She heard from a classmate that the words she thanked a teacher in her acknowledgments were also marked as AI-written. She didn't know why her paper was judged to be written by AI, and she didn't know where to start to revise it. After the initial revision, she checked and found that it had only decreased by 1%.

There were a few paragraphs at the end that she really didn't know how to modify, nor could she delete. So she used the method of reducing the weight of the content, first translating it into English, then translating it back to Chinese, and the AI ​​rate eventually dropped to the 15% required by the school. In order to reduce the AI ​​rate, a classmate of hers translated Chinese into Estonian first, and then translated it back into Chinese.

There is a classmate in Gu Siqi's group whose paper introduction and acknowledgments were written by AI. After sending it directly to the teacher, the teacher "criticized" the student and said, "Such behavior is definitely academic misconduct. But if your own ideas are greater than the participation of artificial intelligence, AI is just a tool."

After the school required students to start checking AI rates, Gu Siqi discussed this with her tutor. Her tutor likes to use AI to assist in teaching and does not restrict students from using AI. The tutor believes that using AI can help students expand their thinking. "For example, if AI gives three ideas, if you can expand them to four or five ideas without being limited to AI's ideas, you will have the initiative to interact with AI."

In the future, Gu Siqi plans to study for a master's degree. She believes that all future papers may be tested by AI, which is necessary, "because many people now use AI to write directly." However, her own writing was also judged to be written by AI, which shows that the system is not perfect yet.

In June, Shu Ran passed the thesis defense and successfully received her diploma. The graduation thesis was lying on her computer. Occasionally, she would recall the "painful process" of reducing the AI ​​rate of the thesis, as if she had just completed a task "in a muddle".

She still knew nothing about the system that judged her as a ghostwriter.

(To protect the privacy of the interviewees, all the characters in this article are pseudonyms)