Is Your Paper Flagged as AI? Here's How to Convert AI to Human Text and Reduce Detection.

This year, major institutions have rolled out significant updates, rendering old tricks like rephrasing sentences or retaining the original subject ineffective. To be honest, I never wanted to rely on such methods in the first place. The reason is that the "AIGC detection rate" itself is a false proposition. For instance, many essays written manually by netizens fail to pass AIGC detection. An article might pass the AIGC check of one institution but fail that of another—and none of these institutions have made their detection standards public. This is simply a game without rules. However, I still got carried away by netizens calling me "expert" repeatedly. After spending a month researching AIGC detection rates, I’m finally ready to share our findings.

1 Why Do Papers Have an "AI Vibe"?

Without proper guidance, the content generated by AI reads very mechanically—that’s what we call the "AI vibe." Why does this happen? I won’t speculate too much on this question. After searching through numerous detection institutions, I finally found some clues in Tencent Zhuque’s official information. Zhuque highlighted two key terms: Perplexity and Unexpectedness.

1.1 Low Perplexity

Perplexity refers to "the degree of surprise when a word appears in a specific position." To put it simply:

• High perplexity: The choice of words is unexpected and surprising.
• Low perplexity: The choice of words is highly predictable and meets expectations.
• AI writing has extremely low perplexity because AI always selects the safest and most common word combinations. For example, in the sentence "On my way home from work yesterday, I saw a...", AI will most likely choose common animals like "cat," "dog," or "bird"—it would never come up with a creative idea like "I saw a penguin in a suit making a phone call." This predictability makes AI-generated text easy for other AI detectors to identify—because the content is just too "normal."

1.2 Low Unexpectedness

Unexpectedness, in simple terms, refers to the "sense of jump" and "irregularity" in an article. Human writing is like a roller coaster: sometimes it includes long, elaborate passages, and sometimes short sentences pop up suddenly; one moment the style is formal and refined, and the next, a popular internet phrase appears; while analyzing a problem seriously, the writer might suddenly insert a personal comment. There are also shifts in vocabulary—from sophisticated terms to colloquial language.

A real person might write: "Literary creation requires profound cultural heritage and rich imagination. To be honest, how many people can really sit down and focus on that these days? Isn’t scrolling through your phone more enjoyable?" Such sudden style shifts are distinctly human.

In contrast, AI writing is like a smooth highway—always maintaining the same speed and style. It never "changes lanes suddenly" or "hits the brakes hard"; it won’t switch from an academic tone to a humorous, joke-like style, nor will it let its thoughts run wild when discussing serious topics. This excessive "perfect consistency" has become its biggest flaw.

This is also related to how AI is trained. Large AI models are trained on massive amounts of text data, and among this data, standardized and formal texts (such as news reports, textbooks, and academic papers) often account for a larger proportion, with higher quality and more structured content. In contrast, personalized and informal expressions (like internet slang or dialects) make up a smaller share of the training data. Naturally, AI learns more "standard expressions."

When generating text, AI works like playing a "word chain game"—it predicts and selects the next word with the highest probability of appearing. Since standardized corpus occupies a larger proportion during training, the probability of standard expressions is higher than that of personalized ones in most contexts.

2 How to Check the AI Generation Rate of an Article?

Wondering if an article was written by AI? Here’s a free and useful AI detection tool to help you easily identify AI-generated content!

2.1 ZeroGPT

ZeroGPT is a free, professional English AI detection website. It supports two methods of input: uploading files and pasting text. It also allows users to export PDF reports for detailed analysis. The website URL is:

https://www.zerogpt.com/

In addition to these free tools, there are many more professional paid options. When writing academic papers, many schools require detection by specific institutions—commonly used ones include Turnitin and iThenticate.

3 How to Reduce the AIGC Rate of a Paper?

3.1 Break AI’s "Uniform Characteristics"

AI-generated articles often have a fixed emotional tone—either sad throughout or objective and calm from start to finish. To eliminate the AI vibe, you need to create fluctuations in emotion within the article; this will make it sound more like it was written by a human.

3.2 Break AI’s "Standardized Vocabulary"

AI-generated text tends to use the most common and standard expressions. To get rid of the AI feel, opt for more personalized language instead of "textbook-style" phrasing.

3.3 Break AI’s Pursuit of "Perfection"

AI usually structures writing very neatly: each section has a similar length, and the logic is extremely rigid. To remove the AI vibe, you should intentionally disrupt this "perfect structure."