Why Human-Written Chinese Papers Get Flagged as AI: The Truth About Bullet Points, Parallelism, and "In Summary"

AI text classifiers do not evaluate the intellectual substance of your research; they are trained machine-learning models that evaluate statistical patterns and structural regularity across sequences of tokens. Structured enumeration (such as "first, second, third"), balanced parallel clauses, and formulaic academic transitions (such as "in summary" / 综上所述 or "from this it can be seen" / 由此可见) represent highly standardized, uniform phrasing in formal Chinese. Because trained classifiers have learned to associate this extreme syntactic symmetry and formulaic predictability with generative AI training distributions, carefully organized human academic prose can trigger elevated suspicion scores.

This dilemma—where careful academic rigor can lead to an automated warning—has attracted widespread public attention. In an investigative report by BeiQing ShenYiDu (Beijing Youth Daily) titled "Graduation Theses Trapped in the 'AI Rate': Methodically Structured Bullet Points and Frequent Parallel Sentences Judged as Not 'Human'", a student recounted being flagged for methodical enumeration and balanced parallel phrasing, asking: "Just because I like to use methodical bullet points and parallel sentences, does that mean I cannot be 'human'?" This tension has led many students to wonder whether domestic detection systems maintain an internal blacklist of forbidden transitional phrases.

The underlying technical reality is straightforward: no detection vendor or published machine-learning study has ever demonstrated that any single Chinese phrase or grammatical construct constitutes an exclusive signature of artificial intelligence. Classifiers respond to distributional patterns—syntactic predictability, token-level transition smoothness, and mechanical pacing across a passage—not to specific vocabulary words.

Understanding why methodical academic prose gets flagged requires examining how modern AI detection technology actually functions. A widespread misconception on student forums suggests that detectors calculate mathematical "perplexity" (how surprised a model is by the next word) and "burstiness" (variation in sentence structure and length) in real time to grade submissions. As detailed in our analysis of how perplexity and burstiness relate to AI detection, those metrics primarily describe early, first-generation demonstration tools. Commercial detection vendors have largely moved beyond them: GPTZero has stated that as of autumn 2023 it no longer uses perplexity and burstiness, while Turnitin has clarified that its model "is not explicitly programmed to evaluate specific signals such as 'burstiness,' 'perplexity'."

Today's production detectors—including major academic screening platforms and neural text classifiers—operate as fine-tuned machine-learning classifiers. These deep-learning models are trained on massive datasets of paired human-written and machine-generated texts to identify complex, high-dimensional linguistic features.

However, standard academic writing instruction in Chinese universities actively cultivates structural habits that overlap closely with the statistical patterns classifiers identify:

• Numbered Enumeration: Outlining measures as "First, establish mechanisms; second, improve procedures; third, strengthen supervision" creates rigid structural symmetry;
• Parallelism and Syntactic Balance: Aligning grammatical structures across paired clauses creates uniform sentence lengths and repetitive syntactic frames;
• Transitional Formulae: Expressions such as "in summary" (综上所述), "based on the above analysis" (基于上述分析), and "thus it can be seen" (由此可知) provide standardized logical bridges whose textual environments are highly predictable.

When a student adheres strictly to formal academic norms by organizing arguments into highly structured, symmetrical paragraphs, the linguistic profile of their prose becomes exceptionally uniform. To a trained neural classifier, this structural regularity closely mirrors the distribution of synthetic text seen during model training.

This defines the crucial evidence boundary: automated detectors are sensitive to rigid structural templates, mechanical symmetry, and extreme organizational uniformity; however, no published technical source treats the mere presence of "in summary" or bulleted lists as conclusive proof of AI authorship. An elevated percentage represents a probabilistic classification based on learned feature patterns, not a verified factual determination of misconduct.

Online search results and social media forums are crowded with commercial marketing posts claiming that "deleting these 20 phrases will instantly reduce your AI rate below 5%." These content farms assemble arbitrary lists of supposed "taboo phrases," instructing students to eliminate standard academic markers such as "in summary," "it is evident," "in-depth exploration," "from the perspective of," and "on one hand... on the other hand."

Commercial platforms promote these phrase-deletion recipes because simplistic word replacement is easy to monetize through automated rewriting tools and fee-based rephrasing services. Analytically and methodologically, this advice is deeply flawed:

1. Classifiers analyze context windows, not static keyword blacklists: Mechanically replacing "in summary" (综上所述) with a colloquial synonym like "broadly speaking" (总的来看) leaves the underlying sentence architecture, contextual embeddings, and structural probability distribution virtually unchanged.
2. Arbitrary deletion compromises academic cohesion and rigor: Scholarly arguments depend on explicit transitional signposts to convey logical relationships. Forcing students to dismantle organized bullet points into disconnected conversational fragments or replace precise technical terms with informal language directly degrades thesis quality, triggering legitimate reprimands during faculty review.
3. Institutional rules dictate tool use, not universal bans or broker claims: Student inquiries often ask whether using rewriting or humanizer software is strictly prohibited. Acceptable tool usage is not governed by a universal rule; what is permissible depends entirely on the specific guidelines of the student's university, academic department, and individual assignment. Some faculties allow generative or assistive tools for language polishing provided the assistance is explicitly declared, whereas others prohibit any automated rewriting. Readers should consult their institution's academic conduct code and speak directly with their supervisor. Crucially, students should beware of commercial intermediaries who claim "guaranteed pass rates" or promise "100% human scores." Reporting by Xinhua News Agency has highlighted the prevalence of paid "AI-reduction" services, while China National Radio (Voice of China) investigated abuses in the market, noting fees ranging from dozens to over a thousand yuan, sellers who simply use AI with prompts to rewrite texts ("用AI降AI率"), and instances where students in disputes were threatened with having their theses uploaded online or reported to their universities. Because detection models operate on statistical probabilities, no commercial tool can legitimately guarantee specific detection outcomes.

Methodical structure and stylized transitions are fundamental conventions of academic discipline, reflecting departmental standards rather than generative shortcuts:

• Engineering and Computer Science: Technical methodologies, algorithmic pipelines, testing parameters, and system architectures are typically communicated via numbered lists, discrete stages, and reproducible sequences to preserve technical precision.
• Law, Management, and Public Administration: Statutory analysis, institutional assessments, and administrative recommendations rely heavily on parallel clauses and standardized summations to ensure unambiguous legal and procedural interpretation.
• Medicine and Life Sciences: Empirical reporting adheres to standardized reporting formats (such as clinical trial descriptions and control group metrics), where standardized phrasing minimizes interpretive ambiguity.

As Professor E Haihong stated in Xinhua's reporting: "Different disciplines should not apply a one-size-fits-all approach to AIGC rates; final determinations should be made by subject-matter experts or faculty advisors." Statistical classification tools cannot distinguish between an author's rigorous adherence to professional scholarly conventions and an algorithm's statistical token completion.

Ultimately, the authoritative determination of academic originality rests with human scholars: faculty advisors, peer reviewers, and departmental thesis committees. University integrity policies focus on plagiarism, fabrication, and unauthorized generative tool use; clear organization, balanced syntax, and rigorous reasoning do not themselves constitute academic misconduct. Academic integrity should not be compromised to placate a statistical tool.

When an independently drafted Chinese thesis receives an elevated AI detection score due to methodical structure, parallel phrasing, or standard academic transitions, degree candidates should avoid panic and resist commercial rewriting brokers promising guaranteed scores. What forms of revision or assistance are permitted depends on your institution's specific policies, but submitting unverified commercial rewrites risks compromising your work. Instead, follow this four-step institutional verification protocol:

1. Preserve and Secure a Comprehensive Draft Trail: The most definitive defense against an automated flag is an unalterable chronological record of the writing process; Export and archive timestamped document versions: Word files with tracked changes and automated local backup iterations, cloud version histories from WPS or Tencent Docs, or commit logs from Overleaf and Git repositories; Assemble secondary developmental evidence: annotated research literature, handwritten laboratory notebooks, raw survey datasets, primary interview recordings, and formal progress reports submitted during thesis milestones.
2. Examine and Cross-Check the Detailed Screening Report: Obtain and thoroughly review the complete, unabridged PDF or HTML report generated by the institution's designated detection platform rather than relying solely on the summary percentage; Identify precisely which sections, paragraphs, or blocks were highlighted, noting whether the software flagged them as generative AI patterns or as traditional literature overlap; Verify whether structural exclusions were correctly configured, ensuring that the bibliography, acknowledgments, appendices, mathematical formulas, and block citations were not mistakenly aggregated into the scored word count.
3. Articulate Disciplinary Conventions and Methodological Logic to Your Advisor: Schedule an in-person meeting with your faculty supervisor, presenting both the flagged report and the underlying research drafts; Clearly explain the disciplinary necessity of the flagged structures: "Section 3 uses a three-part comparative list in accordance with standard engineering benchmarking conventions, and the concluding transitional sentence reflects the direct deduction from the preceding empirical dataset"; Rely on your advisor's academic expertise to validate that the flagged passages represent appropriate disciplinary expression rather than automated generation.
4. Inquire About Institutional Review Thresholds and Appeal Procedures: Consult your department's official thesis management guidelines through the Academic Affairs Office (教务处) or Graduate School (研究生院) to determine whether the AIGC score serves merely as an advisory screening alert or as a strict threshold; Request your university's formal appeal procedures and required forms. Under China's Academic Degrees Law, statutory provisions govern degree reviews: Article 39 provides for presenting statements and arguments (陈述和申辩) before a decision not to award or to revoke a degree, Article 40 provides for academic review (学术复核) of academic evaluation conclusions, and Article 41 provides for review (复核) of decisions not to accept, not to award, or to revoke a degree. Because an initial AIGC screening score does not by itself constitute a statutory disciplinary decision, students facing disputes should follow their institution's established review procedures, submitting their draft trails and writing history for human faculty evaluation.

1. BeiQing ShenYiDu (Beijing Youth Daily) (北青深一度 / 北京青年报): "Graduation Theses Trapped in the 'AI Rate': Methodically Structured Bullet Points and Frequent Parallel Sentences Judged as Not 'Human'" (《困在“AI率”中的毕业论文：有条理地分点叙述，常用排比句，被判定为不是“人”》), published August 8, 2025. URL: `https://news.qq.com/rain/a/20250808A03F3K00`
2. Xinhua News Agency (新华社): "Thesis AIGC Rate: Are Detection Standards Scientific?" (《论文AIGC率，检测标准科学吗？》), published May 23, 2026 via Xinhua WeChat (first published by Science and Technology Daily / CCTV News on May 22, 2026), examining university screening standards and paid AI-reduction services. URL: `https://m.thepaper.cn/newsDetail_forward_33210288`
3. China National Radio (央广网): "Investigation into Abuses in Thesis AI-Reduction Services" (《代降论文AI率服务乱象调查》), published July 14, 2025 (first published by Legal Daily), reporting on paid services using AI prompts to rewrite text, fees from dozens to over a thousand yuan, and dispute threats. URL: `https://news.cnr.cn/native/gd/20250714/t20250714_527260647.shtml`
4. Guangming Daily Online (光明网): "My Self-Written Thesis Was Diagnosed as AI" (《“自己写的毕业论文，被确诊AI”》, July 22, 2024, repost of The Paper) and "Thesis 'AI-Reduction' Has Become a Business: AI Detection Must Not Become Formalism" (《论文“降AI率”已成生意，AI检测不能沦为形式主义》, June 11, 2026, commentary referencing National Business Daily's June 10, 2026 report 《高校严查论文AI率，帮学生“降写”成了火爆生意》). URL: `https://m.gmw.cn/2024-07/22/content_1303798841.htm` / `https://guancha.gmw.cn/2026-06/11/content_38823723.htm`
5. Beijing Daily Client (北京日报客户端): "Thesis AI Detection Is Driving This Year's College Graduates Crazy" (《论文AI检测，正在“逼疯”这届大学生》), published July 19, 2026. URL: `https://xinwen.bjd.com.cn/content/s6a5ccb7de4b03fa51a81b89b.html`
6. Standing Committee of the 14th National People's Congress: Academic Degrees Law of the People's Republic of China (《中华人民共和国学位法》), adopted April 26, 2024 at the 9th session of the 14th NPC Standing Committee, effective January 1, 2025. URL: `https://www.12371.cn/2024/04/28/ARTI1714265542656277.shtml`