AIGC 查重让毕业生崩溃,学生自己写的段落被判定 AI (English)

Generated: 2026-06-23 12:54:38

---

Driven Crazy by AIGC Detection: Handwritten Student Essay Gets 55% AI Rate, Zhu Ziqing 62.88%, Preface to the Pavilion of the Prince of Teng 100%

At two in the morning, a junior from my college posted on Moments, followed by three losing-it emojis.

He said his final project was stuck on the AIGC check. Completely fell apart.

I was about to comfort him when he sent me a screenshot — Zhu Ziqing's Moonlight over the Lotus Pond, written by his own hand in 1927, was flagged by some AIGC detection system as 62.88% AI-generated.

I was genuinely stunned.

You think that's it? Later, someone ran Preface to the Pavilion of the Prince of Teng through the same system. Guess what? 100%. That's right — Wang Bo's timeless masterpiece, written at age 26, was declared by the system to be AI-written.

And Liu Cixin's The Wandering Earth? Over 50% AI rate.

Wait — when Liu Cixin wrote The Wandering Earth, ChatGPT wasn't even born yet, was it?

At this point, one question popped into my head: Is this detection platform actually measuring "whether it was written by AI" or "whether it's well-written"?

Tell me — is there any justice left?

---

Universities Take a One-Size-Fits-All Approach: Guilty or Not, Exceed the AI Rate and You're Done

Come graduation season 2025, guess how many universities issued AIGC detection notices?

Over 90%.

You read that right. More than nine out of ten universities nationwide have it in black and white: graduation theses must pass AIGC detection first. Exceed the AI rate? Disqualified from defense, delayed graduation.

Not "might be checked" — "will definitely be checked." It was that absolute.

At top schools like Tsinghua, Fudan, and Zhejiang University, the AI rate is capped at 15%. Ordinary 985 and 211 schools set it between 20% and 25%. Most undergraduate institutions allow 30% to 40%. Sichuan University is even harsher — no more than 20% for humanities, 15% for science and engineering.

And penalties aren't just "return for revisions" anymore — it's straight to "delayed defense."

What does a delayed defense mean? No diploma and degree certificate by July. Job offers might fall through. Graduate school enrollment can't go through on time. Tell me — is that harsh or what?

From an administrative standpoint, I can actually understand the schools. AI ghostwriting has become a real plague. There's a running joke that in today's college classrooms, the competition isn't about who studies harder — it's about who writes better prompts. People's Daily has also set the tone: AI ghostwriting is academic misconduct, and it will be investigated thoroughly.

But here's the thing — this detection tool itself — is it even reliable?

---

I Dug Into the Detection Principles — It's All a Statistical Game

At this point, let me play translator for a moment.

The mainstream AIGC detection tools out there aren't that complicated in principle. They rely on two core metrics: Perplexity and Burstiness.

What does that mean? Simply put, it's measuring how "predictable" your text is.

AI-generated content usually has a smoother word-choice pattern. The probability distribution for each word is more concentrated, so perplexity is low. Human writing, on the other hand, has random fluctuations, typos, sentence structure variations, and mental leaps — so perplexity is high. Burstiness looks at whether the variation in sentence length and structure feels natural.

Sounds reasonable, right?

But think about it — what does academic writing strive for? Standardized vocabulary, rigorous logic, neat sentence structures. A student who has undergone serious academic training will likely produce writing that is "smooth" and "predictable."

The better you write, the more you sound like AI.

Do you see it? This is a paradox that gives you chills when you really think about it.

And it gets even more absurd. These detection systems don't disclose their algorithms, the basis for their thresholds, or their training data. Even running the same article through the same system at different times can yield fluctuating results.

A student at a university in Jiangxi tested his thesis one day — 10.37% AI rate. The next day, same platform, same article — guess what? 27.54%! His professor sent it back and told him to reduce the AI percentage.

An associate professor at Renmin University submitted a paper that took his team three years of fieldwork to write — and it was flagged as "highly suspected AI-generated." That was original research grounded in three years of grassroots investigation.

The same system can't even reproduce its own results. Think about it — can something like that really be used as a verdict?

---

Graduates Driven Crazy: I Pay AI to De-AI So My AI Rate Goes Down

Now comes the most absurd part.

An undergraduate student, pretty good at writing, produced a decent thesis on his own. He ran it through the detector — 55% AI rate. He knew he hadn't used AI, but he had to get that number down.

What could he do?

First move: Change "Based on the above analysis, this study posits that the mechanism has a significant impact" to "From what we saw earlier, this mechanism has a pretty obvious effect." — Reduce formality, add traces of spoken language.

Second move: Pad polished paragraphs with a bunch of unnecessary transition words and repetitive expressions. Break up neat sentence structures.

Third move: Delete every "firstly, secondly, then, therefore" — turn the whole paper into a rambling narrative. Take any subdivision below the third-level heading and mash it into one paragraph, removing all logical signposts.

Fourth move: Rewrite all flagged sections in colloquial language. Turn "This method demonstrates superior accuracy and stability" into "Well, honestly, this method is pretty good too."

Fifth move: English abstract flagged? Throw the Chinese into Baidu Translate, translate it into Japanese first, then into English. Some people even use Belarusian as an intermediary, deliberately introducing grammatical errors.

But here's the real kicker — someone discovered that if you replace every comma in the paper with a period, changing nothing else, just the punctuation — the AI rate drops from 48% to 11.51%.

Amazing, right?

After all these maneuvers, the AI rate does go down. But what does the paper become? Poor logic, colloquial expression, like a diary entry written by a primary school student.

And that's not even the most ironic part.

People who actually use AI to generate large amounts of content can pass with ease. Because they do one thing: they get AI to reduce its own AI traces. Feed the high-AI text into another AI, have it replace synonyms, adjust word order, add colloquial expressions, and output a low-AI version. There are already "AI reduction" services with clear price tags — tens to hundreds of yuan per paper.

In this entire process, the only step that never touched AI was the part where the student actually wrote the paper.

AI is the problem, AI is the tool, and AI is also the solution. Think about it.

---

The Gray Market: A Multibillion-Dollar Anxiety Industry

Speaking of which, I did the math.

WeiPu charges