This project integrates AI large model capabilities to build a thesis inspection system, providing universities with thesis anti-AI detection capabilities: it supports two modes, text detection and file detection, accurately assesses the probability that a thesis was automatically generated by AI, and generates inspection reports with paragraph-level evidence localization for download and archiving, with supporting recharge billing, account permissions, and usage statistics, helping universities build a solid defense line for academic integrity against the backdrop of a surge in AI ghostwriting.
The thesis inspection system is a thesis anti-AI detection platform that integrates AI large model capabilities, providing academic integrity detection services for universities. The system uses AI large model semantic discrimination and multidimensional text feature analysis as a dual-channel engine to detect the probability that a thesis was automatically generated by AI. It supports two modes: text detection (paste-and-check snippets) and file detection (batch upload of Word/PDF). It automatically generates inspection reports containing overall probability, paragraph-level probability distribution, and highlighted localization of highly suspicious paragraphs for download and archiving, and is equipped with recharge billing, multi-role accounts, usage statistics, and open API integration capabilities. The system is positioned to provide auxiliary evidence for the determination of academic misconduct in universities—quantifying probability rather than directly convicting—forming a "plagiarism + AI ghostwriting" double insurance alongside duplication rate checking, helping universities safeguard academic integrity against the backdrop of a surge in AI ghostwriting.
With the popularization of generative AI large models such as ChatGPT, the phenomenon of "AI ghostwriting" in theses, thesis proposals, and coursework has spread rapidly in universities. AI-generated text is fluent, well-structured, and superficially highly original. Traditional duplication rate checking centered on "text copy ratio" is almost ineffective against it—a low duplication rate does not mean the student wrote it themselves.
At the same time, the "Degree Law of the People's Republic of China" explicitly lists "ghostwriting" as academic misconduct, and many universities have successively issued regulatory requirements for the use of AI tools in graduation theses. Academic integrity management is moving from "institutional provisions" to "implementation." But the reality is: teachers rely on naked-eye and experience-based identification of AI traces, which is inefficient and inconsistent in standards; questioning students without evidence easily triggers appeals and teacher-student conflicts; and each college seeks detection tools on its own, lacking a unified platform, with no way to guarantee thesis data security.
Against this background, this system integrates AI large model discrimination capabilities and multidimensional text feature analysis. Developed by Shenzhen Xiangming Technology Co., Ltd., it provides universities with an integrated thesis anti-AI inspection service of "text detection + file detection + inspection report download + recharge billing," so that academic integrity management in the AI era has evidence to rely on and tools to use.
Before the system was built, universities mainly faced six types of pain points in dealing with the AI ghostwriting problem:
| No. | Pain Point | Specific Manifestation |
|---|---|---|
| 1 | Surge in AI ghostwriting | Large models are readily available, and the proportion of AI-generated theses, proposals, and coursework is rising rapidly, making it hard to guard against |
| 2 | Traditional duplication checking "cannot detect AI" | Duplication rate detection only guards against "plagiarism." AI-generated text is superficially highly original with a low copy ratio, easily bypassing it |
| 3 | Manual identification is inefficient and subjective | Teachers judge AI traces one by one with the naked eye, which is inefficient and inconsistent in standards, and simply impossible to review at large sample sizes |
| 4 | Determinations lack evidentiary support | Questioning students based on feeling easily triggers appeals and teacher-student conflicts. Handling academic misconduct requires archivable evidence |
| 5 | Institutional implementation lacks tools | The Degree Law and campus AI usage regulations exist, but there is a lack of unified technical leverage to support implementation |
| 6 | Detection services are scattered and unmanaged | Each college finds tools on its own, with no unified platform and no usage management, creating high risk of thesis data leakage |
Centering on "professional detection, archivable evidence, platform-based services, and reassuring data," the system establishes four major construction goals:
Professional detection capability.With AI large model discrimination as the core engine, output the thesis "AI generation probability," with dual modes of text pasting and file upload covering various submission forms.
Archivable determination evidence.Generate inspection reports with paragraph-level localization and rich text and graphics, downloadable and archivable, as auxiliary evidence for academic integrity handling.
Platform-based service system.Integrate recharge billing, account permissions, and usage statistics to support a service model of unified procurement at the school level and unified quota distribution by colleges and departments.
Reassuring thesis data.Encrypted transmission and no retention of original text after detection protect student privacy and school academic data security.
The system is organized around three major stages: "intelligent detection services—report and evidence management—account and operations management," with a total of 9 core functional modules.
(1) Intelligent Detection Services
1. Text Detection
What it is:Paste a thesis snippet or full text to initiate detection, and the system returns the overall AI generation probability and paragraph-level assessment results in real time.
What problem it solves:Suitable for lightweight scenarios such as sample spot checks and quick pre-defense checks, with instant paste-and-check and immediate results.
Who it is for:Advisors, teaching secretaries, and students for self-checks.
2. File Detection
What it is:Supports uploading and detecting common thesis formats such as Word, PDF, and TXT, compatible with both single and batch modes, adapting to the volume of centralized submissions during graduation season.
What problem it solves:The entire thesis does not need to be copied and pasted, and batch submission makes unified department-level detection possible.
Who it is for:College teaching offices, degree offices, and academic affairs offices.
3. AI Generation Probability Assessment Engine
What it is:The system's core engine, integrating AI large-model semantic discrimination with multidimensional text feature analysis (sentence structure, vocabulary distribution, semantic fluency, etc.), outputting an overall AI generation probability and a paragraph-by-paragraph probability distribution.
What problem it solves:It breaks through the old approach of high false-positive rates in "keyword rule"-based methods, using large-model comprehension to identify the deep features of AI text, making detection more accurate and localization more granular.
Who it is for:A built-in system capability; all detection results are driven by this engine.
(II) Report and Evidence Management
4. Inspection Report Generation and Download
What it is:After detection is completed, a graphical report is automatically generated: overall probability, segmented probability distribution, highlighting of highly suspicious paragraphs, and feature explanations, supporting download for archiving and printing.
What problem it solves:It turns "I think it looks AI-written" into "the report shows which paragraphs are suspected of AI generation and with what probability," giving teachers and students evidence for communication and leaving a trace for handling.
Who it is for:Advisors, academic committees, and degree offices.
5. Detection Record Management
What it is:Historical detection records are retained uniformly, supporting retrieval by time, submitter, and file name, and allowing viewing and comparison of recheck results.
What problem it solves:After thesis revisions, rechecks have a baseline for comparison, and the entire detection process is traceable, meeting process archiving requirements.
Who it is for:Academic affairs offices and college administrators.
(III) Account and Operations Management
6. Recharge and Billing Management
What it is:Supports billing by number of detections or word count, with one-stop online recharge, order management, balance inquiry, and invoice application.
What problem it solves:Schools can make unified procurement and recharge, distribute quotas by department, keep detection services running sustainably, and make expenses clear and auditable.
Who it is for:School procurement departments and department administrators.
7. Account and Permission Management
What it is:Provides a multi-role account system for administrators, teachers, students, etc., distinguishing permissions for initiating detection, viewing reports, recharge management, and more.
What problem it solves:Students can self-check and self-correct, teachers can conduct targeted spot checks, and administrators can centrally manage quotas—one platform serves multiple roles.
Who it is for:Information Center, Academic Affairs Office.
8. Usage Statistics and Operations Dashboard
What it is:Visualizes data such as detection counts, word consumption, balance changes, and departmental usage distribution.
What problem it solves:Where the quota is spent and how much remains are clear at a glance, providing a basis for next year's procurement and service planning.
Who it is for:School management, department administrators.
9. Open Interfaces and System Integration
What it is:Provides standardized APIs that can connect to academic affairs systems and thesis management platforms, embedding AI detection into the thesis review workflow.
What problem it solves:Detection is no longer an isolated website, but a step in the dual-check process of "plagiarism check + anti-AI" for dissertations.
Who it is for:School IT departments, thesis platform vendors.
The system adopts a four-layer architecture, with equal emphasis on detection capability and data security:
Application layer:Web detection portal, report center, and admin backend, serving the detection scenarios of teachers and students and the operational scenarios of administrators, respectively.
Service layer:Detection task scheduling, report generation engine, account billing and permission services, supporting high-concurrency centralized detection during graduation season.
Engine layer:Dual-channel analysis integrating AI large model discrimination and multi-dimensional text feature analysis, outputting overall and segment-level AI generation probabilities.
Security layer:End-to-end encrypted transmission, original text is not retained after detection is completed, and thesis data and account data are stored in isolation.
Large model semantic discrimination, not keyword rules.The engine uses large model semantic understanding of text to identify AI-generated features, resulting in lower false positive rates and stronger resistance to rewriting and polishing than traditional "keyword + rule" solutions.
Probability output + paragraph-level localization.It does not give a crude "yes/no" conclusion, but quantifies probability and localizes it to specific paragraphs, allowing advisors to focus on checking high-suspicion paragraphs, and the report can be used directly for teacher-student communication.
"Auxiliary evidence" rather than "final adjudication".The system is positioned to provide objective reference basis for academic integrity handling, with the final determination right reserved for advisors and academic committees—this is the professional prerequisite for detection results to withstand appeal scrutiny.
Dual-mode detection fits the real process.Text pasting serves fragment spot checks, and batch file upload serves full-paper submission during graduation season; both forms are covered.
Billing and platform-based design.Recharge billing + multi-role accounts + usage statistics transform the detection service from a "one-time tool" into a normalized capability that schools can operate long-term.
After the system went live, it brought significant improvements across four dimensions: efficiency, evidence, institutionalization, and operations:
Single-paper detection returns in minutes, greatly reducing teachers' manual preliminary screening workload and significantly easing the pressure of batch submissions during graduation season.
Detection results change from "subjective feeling" to "paragraph-level probability evidence," giving teachers and students a basis for communication and significantly reducing disputes caused by judgments.
It provides an execution lever for the implementation of schools' AI usage regulations, shifting academic integrity management from post-hoc dispute handling to pre-check self-inspection and in-process spot checks.
Self-inspection and self-correction before submission have become the norm, with proactive revision replacing passive accountability, forming positive guidance.
Recharge billing and usage statistics make detection quotas and expenses clear and auditable, enabling sustainable service operation.
Safeguarding academic integrity.In an era where AI ghostwriting is normalized, it fills the key gap of "detecting AI" for universities, forming a double safeguard alongside plagiarism rate checking.
Reducing the burden on teachers.The division of labor between machine preliminary screening + manual review frees teachers from visually screening each paper one by one.
Fair and evidence-based judgments.Quantified probabilities and paragraph localization make handling able to withstand appeal scrutiny, reducing the risk of misjudgment and teacher-student conflicts.
A lever for institutional implementation.It turns degree laws and campus AI usage regulations from provisions into executable daily processes.
Worry-free data security.Encrypted transmission and no retention of original text provide reliable protection for student privacy and school academic data.
Sustainable operations.The billing model and usage management support schools in purchasing on demand and distributing by department, with controllable costs.
Generative AI will not be absent from any campus, and the defense line of academic integrity also needs to be upgraded simultaneously. With "AI large model discrimination + paragraph-level evidence + platform-based operations" as its core capabilities, this system helps universities turn anti-AI paper checking from "not knowing where to start" into "having evidence to rely on." Shenzhen Xiangming Technology Co., Ltd. will continue to iterate its detection engine and institutional service capabilities, working with universities to safeguard academic originality and integrity.
——Shenzhen Xiangming Technology Co., Ltd. · Case Library
Q1: What is the paper checking system?
A: This is an anti-AI paper detection platform integrating AI large model capabilities. It detects the probability that a paper was automatically generated by AI, providing universities with an integrated academic integrity checking service covering text detection, file detection, inspection report download, and recharge billing.
Q2: What core functional modules does the system include?
A: There are 9 core modules in total, organized into three major stages: intelligent detection services (text detection, file detection, AI generation probability assessment engine), report and evidence management (inspection report generation and download, detection record management), and account and operations management (recharge and billing, accounts and permissions, usage statistics and operations dashboard, open interfaces and system integration).
Q3: How does the system determine whether a paper was written by AI?
A: The core engine adopts a dual-channel approach of "AI large model semantic discrimination + multi-dimensional text feature analysis": the large model understands the generation characteristics of the text at the semantic level, while feature analysis captures statistical clues such as sentence structure, vocabulary distribution, and fluency. After cross-assessment of the two channels, it outputs the overall AI generation probability and paragraph-by-paragraph probability distribution, with a lower false positive rate than keyword rule-based solutions and stronger resistance to rewriting and polishing.
Q4: What is the difference between text detection and file detection?
A: Text detection is aimed at lightweight scenarios—paste a paper excerpt or full text, and it is checked immediately upon pasting and obtained immediately upon checking, suitable for tutor spot checks, quick checks before defense, and student self-inspection; file detection supports Word, PDF, and TXT uploads and is compatible with batch mode, suitable for department-level centralized full-paper submission during graduation season. Both modes share the same assessment engine.
Q5: Can the inspection report serve as a basis for handling academic misconduct?
A: The report includes the overall probability, paragraph-level probability distribution, highlighting of high-suspicion paragraphs, and feature explanations, and can be downloaded and archived. It is an important **auxiliary evidence** for academic integrity handling. However, the system is positioned as "auxiliary evidence" rather than a "final ruling"—whether academic misconduct is constituted ultimately remains with the tutor and academic committee, who make a comprehensive judgment based on the writing process and defense performance. This is also the professional premise that enables detection results to withstand appeal scrutiny.
Q6: How is paper data security guaranteed?
A: Transmission is encrypted throughout, the original text is not retained after detection is completed, and paper data and account data are stored in isolation; schools can submit papers centrally through a unified platform, avoiding the risk of paper data leakage caused by teachers and students using external tools on their own.
Q7: How does recharge billing work?
A: It supports billing by number of detections or by word count. After schools make unified procurement and recharge, quotas can be distributed by department and by role; online recharge, order management, balance inquiry, and invoice application are completed in one stop, and the usage statistics dashboard makes quota consumption and expenses clear and auditable.
Room 1409, Building C, Shenzhen Century Industrial Center, No. 743 Zhoushi Road, Hezhou Community, Hangcheng Sub-district, Bao'an District, Shenzhen
