Plagiarism Rate vs. AI Rate: Why Did Your Paper Pass Similarity but Fail the AI Scan?
Passing a thesis plagiarism check does not guarantee passing an AI screening. Text-matching systems scan indexed databases for overlapping passages, while AI classifiers evaluate linguistic features in submitted prose. Understanding the technical divergence and operational limits of each tool helps explain why these scores diverge and why automated metrics cannot replace academic evaluation.
HumanPen Team
· 7 min read
Why Did Your Thesis Pass Plagiarism Checking but Fail the AI Rate?
A thesis plagiarism check (similarity report) and an AI detection check (AIGC report) operate on distinct technical foundations and evaluate completely different questions. Passing one screening provides neither a technical nor an institutional guarantee of passing the other.
In thesis submission workflows, authors sometimes encounter scenarios where a similarity score is low while the AI detection indicator remains elevated. This divergence stems directly from the fundamental division between database text matching and linguistic feature classification:
- Similarity checking systems compare submitted manuscripts against indexed repositories of academic journals, dissertations, conference proceedings, and public web documents to identify verbatim textual overlap.
- AI detection systems do not search a database of published literature; instead, they analyze the stylistic and linguistic features of submitted prose to estimate how closely the text aligns with patterns associated with machine-generated language models.
When an author uses a generative AI tool to draft or paraphrase sections of a manuscript, the resulting prose consists of new combinations of phrasing that have not been cataloged in published literature databases. As a consequence, the similarity engine finds no matching sources, yielding a low similarity percentage. However, those passages may still exhibit characteristic markers of machine-generated prose, leading an AI classification model to flag them. Conversely, standard literature citations or common technical phrasing may generate matches in a database-driven similarity report while reading as conventional human academic writing.
Similarity reports ask: "Does this text match passages in our indexed literature databases?" AI detection reports ask: "Do the linguistic features of this prose correspond to patterns identified by our classification model?" Because the two tools evaluate entirely different dimensions of a submission, observing differing outcomes across them is standard system behavior.
Database Text Matching vs. Statistical Language Classification
Academic similarity systems—such as CNKI Academic Integrity, VIP, Wanfang, or Turnitin Similarity—function as database comparison engines. Using comparison algorithms, these systems partition an uploaded document and scan it against indexed journal articles, theses, dissertations, and web repositories, highlighting overlapping passages and identifying source publications. The resulting similarity indicator reflects the proportion of matched text relative to the indexed corpus, a figure directly shaped by database coverage, comparison parameters, and user-selected filters such as quotation exclusions.
In contrast, AIGC screening systems function as feature classification tools. Rather than searching for specific pre-existing publications, these systems evaluate structural and stylistic features of submitted prose—such as vocabulary distribution and sentence formation—against a trained classification model. Implementation approaches, model architectures, feature sets, and scoring thresholds vary across different service providers.
This structural distinction determines how standard academic elements behave across the two types of reports:
- Quotations and standard expressions: Verbatim literature citations, statutory definitions, standardized methodology descriptions, and bibliographic entries are widely documented across academic databases. In a similarity system, these elements readily register as matches unless appropriate exclusion filters are enabled. In an AI screening system, models focus on evaluating connected prose, and individual services apply their own filtering rules and structural parsing methods to handle non-prose elements such as formulas, code, tables, or bibliographies.
- Filters and configuration dependencies: Similarity systems typically allow users and administrators to configure granular filters, such as excluding references or previous publications. AI classification systems do not match against external source documents; their output depends entirely on how the submitted text aligns with the service's internal classification model and threshold settings.
What Each Report Can and Cannot Tell You
Interpreting detection results requires recognizing the technical boundaries of automated software. Both similarity percentages and AI indicators represent computed outputs under specific system configurations, rather than definitive assessments of scholarship or intellectual integrity.
| What the Report Can Tell You | What the Report Cannot Tell You |
|---|---|
| Similarity Report: Identifies text passages that match indexed documents (such as journal articles, dissertations, or public web pages) within the system's database coverage; displays an overlap indicator based on the system's active scope and filter settings. | Similarity Report: Cannot determine whether matched text represents legitimate scholarly citation, standard terminology, or improper copying; cannot verify whether non-matching text was independently written by the author or generated through other means. |
| AI / AIGC Report: Highlights prose passages that match the model's learned patterns of machine-generated text based on its specific classification approach; provides an indicator reflecting the model's classification output under current service settings. | AI / AIGC Report: Cannot identify the specific AI model, prompt, account, or user involved in generating the text; a standalone automated score cannot itself establish academic misconduct or substitute for institutional academic evaluation. |
A similarity report directs readers to specific source literature; an AI report provides an algorithmic classification of prose characteristics. Both outputs serve as technical screening aids and cannot replace substantive evaluation by academic advisors and institutional review committees.
The Four Score Combinations and the Reality of "Dual-Low" Theses
Because text-matching and linguistic classification evaluate independent dimensions of a manuscript, four distinct combinations of outcomes can occur:
- Lower Similarity, Higher AI: The manuscript contains little textual overlap with indexed literature, but certain sections display features flagged by an AI classification model. This pattern frequently arises when generative tools are used to draft text directly or to mechanically paraphrase existing sources.
- Higher Similarity, Lower AI: The manuscript contains significant overlap with existing literature—such as standard definitions, regulatory excerpts, experimental protocols, or unexcluded references—while the narrative prose itself does not trigger AI classification flags. In this scenario, verifying citation formatting and exclusion settings is the appropriate step, rather than rewriting sentences.
- Higher Similarity, Higher AI: The manuscript displays both verbatim matches with published sources and passages flagged by AI screening models. This situation requires separate reviews: validating citations and sources for the similarity matches, and examining drafting history for the AI-flagged sections.
- Lower Similarity, Lower AI ("Dual-Low"): Neither screening indicator exceeded designated thresholds under the system's current parameters.
In academic discussions, achieving "dual-low" (双低) metrics refers to keeping both the similarity score and the AI screening indicator beneath an institution's or department's designated thresholds. However, remaining within automated screening limits merely indicates that initial screening parameters were satisfied; it does not in itself verify the intellectual rigor, factual accuracy, or original contribution of the research.
What to Do When Your Similarity and AI Results Conflict
When similarity and AI detection reports show unexpected results or appear to conflict, follow these practical steps:
- Examine the detailed official report: Look beyond the headline percentage score. Review highlighted sections, inspect which comparison databases were accessed, and verify whether quotation and bibliography exclusions were properly applied.
- Review citation and formatting standards: If the similarity indicator is elevated, examine the flagged passages individually. Verify whether they consist of standard technical terms, legal definitions, or authorized citations, and ensure formatting complies with institutional style guidelines.
- Avoid mechanical rewriting to manipulate scores: When addressing an AI screening notice, avoid arbitrary synonym substitution, distorted sentence structures, or third-party automated rewriting services. These approaches frequently introduce conceptual errors, damage academic clarity, and compromise readability.
- Maintain complete records of the drafting process: Preserve dated outlines, successive drafts, laboratory or field notes, reading summaries, and word processor version history (such as revision tracking in WPS, Word, or Tencent Docs) to demonstrate the chronological progression of the research.
- Consult with your advisor and follow institutional channels: If a screening indicator triggers department-level review or concerns, bring the detailed report and your drafting documentation directly to your thesis supervisor, following the official consultation and review procedures established by your institution.
Sources and Institutional References
- Standing Committee of the National People's Congress of the PRC / Ministry of Education: Academic Degrees Law of the People's Republic of China (《中华人民共和国学位法》, adopted April 2024, effective January 1, 2025; establishing statutory procedures for degree evaluation and institutional reviews). Ministry of Education regulatory portal: `https://www.moe.gov.cn/jyb_sjzl/sjzl_zcfg/zcfg_jyfl/202404/t20240426_1127804.html`
- CNKI (中国知网): Academic Integrity Literature Check and AIGC Detection Services. Official portal: `https://check.cnki.net/`
- VIP (维普资讯): Thesis Similarity Check and AIGC Text Detection Services. Official portal: `https://vpcs.cqvip.com/`
KEEP READING
Are Thesis AI-Lowering Services Reliable? How to Vet Third-Party Intermediaries
12 min read
Academic writingCNKI AI Detection Costs Over 100 RMB: Should You Pay for a Pre-Check?
8 min read
AI detectorsUniversity of Reading AI Detection Policy: Turned Off, Not Activated, or Unmentioned?
9 min read