Universities spend money to detect AI, while students spend money to reduce AI similarity

This graduation season, a CCTV reporter interviewed a graduate named Ge Jiayi. She saw no reason to feel guilty: when writing her thesis, AI only helped her look up some theoretical materials, while the abstract, project plan and core creative content were all original. However, the test results showed that suspected AI-generated content accounted for 56%; for the section of the project plan she drafted with her team, the figure reached 97%.

The machine ruled that her work was not written by a human. What she had to do was revise her writing until the machine acknowledged it as human-generated.

She was not the unluckiest one. This graduation season, some graduates tried to reduce the 62% AI content rate, but after revision and re-testing, the rate rose to 94% instead of dropping. On e-commerce platforms, "reducing AI content rate" has long been a clearly priced business: a graduate named Zhang Yue paid 30 yuan at the rate of "10 yuan per 100 words" for 386 words judged as "suspected AI-generated".

It has long been established how unreliable this kind of detection is. The real question that no one has asked is: since everyone knows it is inaccurate, why do schools keep purchasing it every year?

To figure this out, one has to look through a type of document that few people read: the procurement announcements of universities. The purchase reasons written by schools on these announcements are not the same as what students imagine.

Even OpenAI Shut Down Its Own Detector

In June this year, a reporter from Yicai conducted an experiment: asking DeepSeek to generate a 1,000-word article, and submitting it to two major detection platforms respectively. For this 100% AI-generated text, HowNet judged the AI content rate as 0%, and Weipu judged it as 55.71%. One completely missed the detection, and the other only recognized half of it. For an article with a fully clear origin, neither system gave an answer close to the fact.

This is not a unique problem with domestic systems. On July 20, 2023, OpenAI quietly took down its own AI text classifier, leaving a note on its official website: due to excessively low accuracy. According to the data released by OpenAI itself, this tool can only recognize 26% of AI texts, and will misjudge 9% of human-written texts as AI-generated. The company that developed ChatGPT cannot handle this matter itself.

Turnitin, the most widely used tool by universities around the world, went even further. Its official documentation explicitly states that the detection results "should not be used as the sole basis for taking adverse measures against students"; for the range below 19%, no specific number is displayed at all, only an asterisk, because this section is most likely to cause misjudgments. A Stanford team published a set of data in the *Patterns* journal in 2023: 91 TOEFL essays written by non-native English speakers were submitted to seven mainstream detectors, with an average of 61% being judged as AI-generated, and 19% being unanimously identified by all seven detectors as not written by humans. The reason does not lie in writing proficiency: the more constrained and standardized the word choice and sentence patterns are, the more "predictable" the text appears to the machine.

The problem lies at the root, and no iteration of algorithms can fix it. Cai Hailong, Vice Dean of the College of Education of Capital Normal University, explained the difference between the two in a CCTV interview: plagiarism check compares the thesis with the corpus sentence by sentence to make a deterministic judgment; AI detection uses AI to guess human texts, which "is essentially a probability-based classification".

Different platforms have different models, and the features they refer to are not limited to "perplexity", but they all share one thing in common: they can only judge how statistically similar a piece of text is to AI-generated content, and cannot restore its real generation process. As long as what is submitted is still a final text, rather than writing records, no matter how upgraded the algorithm is, it can only calculate the "similarity" more precisely, and will never be able to confirm the "authenticity".

A probabilistic game has been turned by schools into a rigid index with two decimal places. Sichuan University requires that the AI content rate for liberal arts should not exceed 20%, and for science, engineering and medical disciplines, it should not exceed 15%; Hongshan College of Nanjing University of Finance and Economics sets the reference line at 40%, and students exceeding the line will be urged by their supervisors to make rectifications. The standards are set by each school itself. The same thesis may have a completely different fate if submitted to a different school. Probability is flexible, but the red line is rigid, and the students caught in the middle are the ones who get punished.

AIGC Detection Inherits the Procurement Rationale from the Plagiarism Check Era

Don't schools know all this? At the very least, misjudgment is no longer an unknown risk: CCTV and Guangming Net have presented the cases to the public, and even the detection platforms themselves have included the technical limitations in their official documents. The risks are obvious, so why do schools still purchase these tools? The announcements hold their own answers.

When universities procure thesis detection systems, they generally follow the "single source" procedure, which means no bidding and a designated supplier, and the reason for single-source procurement must be made public. An announcement from the Academic Affairs Office of Xinxiang University listed three reasons for HowNet's student thesis detection system: accuracy, uniqueness, and continuity. Reading further, "accuracy" refers to the supplier having the largest and most comprehensive academic resources in China; "uniqueness" means it is "the only detection system that is not separately open to individual students"; "continuity" means the university has used it since 2013, and "the system retains a large number of our undergraduate theses".

The April 2023 announcement of Changzhou University is similar, with a price of 39,600 yuan per year, and a three-year contract signed at once.

The first point is interesting. In the era of plagiarism check, it is reasonable to write "accuracy" in this way: plagiarism check relies on the database, and the larger the database, the more comprehensive the detection, so the accuracy can indeed be guaranteed by the scale of resources. However, AIGC detection has no comparable database; it is a probability classification, and no one can guarantee the misjudgment rate, yet the procurement rationale is directly carried over. When AIGC detection was introduced into schools, it was rarely re-demonstrated as a new product. A more common practice is bundling: in the June 2025 procurement announcement of Zhejiang University of Mechanical and Electrical Engineering, 500 AIGC detection quotas are bundled with HowNet's scientific research achievement detection system in the same 55,000-yuan project. What schools take over is not just a new algorithm, but also the existing account system, historical thesis database, and the migration cost of replacing suppliers. A system that has been used for ten years, with one more service option checked, can "manage AI", which is more convenient than any new solution.

As for "not separately open to individual students", which sounds like a deliberate obstacle to students, was clearly written as an advantage in the announcements back then: since students cannot access the system used by schools, there is only one outlet for detection results, and authority is guaranteed through monopoly. Now students can buy various self-test services on the market: HowNet AIGC detection costs 2 yuan per thousand characters, Weipu charges 20 yuan per article, and PaperPass bundles plagiarism check and AIGC detection at 1.5 yuan per thousand characters. However, the version and corpus used for self-test may not be consistent with those of the school's procurement, so after spending money, the numbers still do not match.

Ge Jiayi cannot see any of these procurement reasons. She has no idea what standards the system uses to judge her, nor can she figure out how the 56% figure is obtained. She only knows that she truthfully declared the use of AI according to the rules, but what she got in return was the obligation to prove the innocence of every piece of her original writing.

A Detection Report Is Cheaper Than a Manual Judgment

There is something even more rigid weighing on schools. The Degree Law, which came into effect on January 1, 2025, clearly stipulates that if there is academic misconduct such as ghostwriting, plagiarism, or forgery in a degree thesis, the degree can be revoked; during the first draft review of the law, "using artificial intelligence to ghostwrite a degree thesis" was once explicitly listed. Coupled with the normalization of random spot checks on undergraduate graduation theses, schools must prove to their superiors: I have exercised management.

What can be used to prove it? A detection report with a percentage is the cheapest option of all. It takes a lot of working hours for supervisors to read every thesis carefully, and strict defense will lead to the responsibility of a rising graduation delay rate. With a machine report costing tens of thousands of yuan per year, the responsibility shifts from "the school failed to exercise proper control" to "the student failed the machine check". A supervisor's comment that "this thesis reads like AI-generated" cannot be entered into the form, but "AI content rate 38.6%" can be registered, sorted, archived, and used for accountability. As for how far the number is from the truth, it cannot be seen from the form, and no column needs to fill in this information.

This kind of thinking was thoroughly described by someone thirty years ago. The historian of science Theodore Porter explored in *Trust in Numbers* why various institutions are obsessed with quantification, and the answer is a sentence that has been widely quoted later: quantification is a "decision-making method that does not seem to be making decisions". He also found that the impulse to replace personal judgment with numbers appears precisely in the most vulnerable institutions: the more untrusted an institution is, and the more exposed it is to external accountability pressures, the more it needs a seemingly objective number to stand in front. Applied to today's universities, this judgment hardly needs any translation: the Degree Law and thesis spot checks have put schools in a position of being held accountable, and the judgment of supervisors cannot withstand this pressure, but numbers can.

What works best for schools is not even a machine that draws conclusions for them, but a machine that gives numbers and notes "for reference only". When management is needed, the percentage becomes the basis for rectification; when disputes arise, the final judgment can be pushed back to the supervisor. Being able to provide seemingly objective numbers without taking full responsibility for them is where the real value of such tools lies.

What schools buy is never detection, but exemption from responsibility.

The Payer Does Not Bear the Consequences of Misjudgment

No other instrument in the world works this way. If the scale in the cafeteria is off by two liang, students will protest on the spot, and the scale will be replaced; if the thermometer has an error of two degrees, the hospital will return it the next day. Why is AI detection, which is known to all for its mistakes, still renewed every year?

The difference is that the reading of the scale is verified on the spot by the person who pays, while for the AI content rate reading, the paying school does not verify it, and the students stuck by the numbers have no procurement rights. For an instrument, as long as the payer and the victim are not the same group of people, "accuracy" is no longer its quality indicator. Even for the school that buys it, the deterrent effect itself is a function: a surveillance camera that cannot clearly capture faces, as long as its red light is still flashing, no one will act recklessly in the corridor.

A misjudgment, on the school's back end, is a report to be reviewed and a little extra working time; on the student's side, it means multiple top-ups, an all-nighter, and several hours of deliberately modifying coherent sentences to make them less smooth — some delete "in conclusion", some break up parallel sentences, and some even have to revise their work back as required by their supervisors. When the costs of both sides are added up, the price signal perceived by the procuring party is so weak that it can be ignored.

If One Number Is Not Credible, Add Four More

After schools find that the system is inaccurate, they often do not return it, but buy several more sets. In the April 2026 notice of the College of Grassland Science and Technology of Sichuan Agricultural University, an undergraduate graduation thesis has to go through five machine checks: anti-plagiarism detection, AIGC detection, Weipu format detection, plus two sets of AI evaluation from Weipu Smart Evaluation and Huachen Smart Evaluation, before being reviewed by the supervisor. The five checks follow one logic: if a single number is unreliable, let a bunch of numbers corroborate each other.

Students' countermeasures, in turn, feed the system. The AI content reduction service on e-commerce platforms charges 3 yuan per thousand characters, and one merchant has sold 4,166 items; a reporter from *National Business Daily* tested it in June this year, and the AI content rate of an article was reduced from 95.77% to 11.3% after processing, at the cost of the language becoming awkward and no longer resembling normal thesis expression. Most of the time, the so-called AI content reduction is also completed by AI: AI writes the thesis, AI eliminates AI traces, and then it is submitted to AI for detection. When these tricks are passed to schools, the conclusion is often only one: the existing detection is not strong enough, and it needs to be upgraded.

The economist Charles Goodhart described this cycle long ago, and the anthropologist Marilyn Strathern condensed it into one sentence: when a measurement indicator becomes a target, it is no longer a good measurement indicator. The AI content rate perfectly embodies this statement: once linked to the qualification for thesis defense, what students modify is no longer the thesis itself, but that number.

The failure of the indicator is not an accident, but the destined outcome of this way of use.

Let's lay out the whole chain: schools submit the record of "completed review" to the regulatory authority, detection providers get demand from both school procurement and student self-test, and AI content reduction merchants take orders one by one in the gaps. After going around a full circle, no party pays based on "whether the judgment is correct".

If this machine were really accurate, part of the business would collapse. There is no evidence that any platform deliberately makes mistakes, but the structure is clear: misjudgment leads to re-detection, which generates revenue; misjudgment leads to panic, which attracts customers for AI content reduction merchants. Inaccuracy is the productivity of this business.

The Most Expensive Detection Is Done by Humans

To be fair to schools, AI ghostwriting is a real problem. If left unmanaged, graduation theses will sooner or later become a relay race of large language models.

But "needing to manage" does not mean "managing in this way". Some universities have already demonstrated another management method. In its May 2025 notice, Nanjing University explicitly acknowledged that AIGC detection results are "probabilistic analysis based on algorithm models, with technical limitations, and are only used as auxiliary references for academic standardization, not as the basis for judging thesis originality". At the same time, it requires students to file applications for the whole process of AI use and retain process records. Professor E Haihong of Beijing University of Posts and Telecommunications went a step further in his suggestion: different disciplines should not adopt a one-size-fits-all approach, and the final conclusion should be drawn by discipline experts and supervisors.

These solutions lead to the same goal: take the judgment power back from the machine and return it to humans. The only difficulty is one word: expense. The cost of a supervisor reading a thesis carefully is much higher than that of a machine scanning it; vouching for a student that "this is his own work" even requires the supervisor to put his own academic credit as collateral.

So most schools choose machines. The reason is not necessarily that machines are more accurate, but that it does not matter even if they are not accurate.

Next March, a new round of procurement announcements will be posted on the campus network, and the reasons will probably still be those few: the most comprehensive resources, the only channel, ten years of use. In the dormitory in the early morning, there will be another Ge Jiayi, revising her own sentences over and over until they no longer sound like herself, then paying 20 yuan to ask another machine to prove that she is a human.

At that time, schools will get the report, students will get their innocence, and every party will get what they want. There is only one thing: after all the money is spent, no one still knows whether that thesis was really written by her.

This article is from the WeChat Official Account "Singularity Outside", author: wiwi, and published by 36Kr with authorization.