Plagiarism Detection: Embedding Analysis on GPU

The Challenge: Paraphrased Plagiarism That String Matching Cannot See

A university group operating six institutions across England and Wales processes 280,000 student submissions per academic year through a commercial plagiarism detection service. The existing tool excels at finding verbatim copying and light paraphrasing but fails to detect sophisticated plagiarism: AI-paraphrased passages, contract-cheating essays written to order, and heavily reworded material that preserves ideas while changing every surface-level phrase. An academic integrity audit revealed that the existing tool flagged only 4.2% of submissions, while manual deep-review of a random sample suggested the true academic misconduct rate was closer to 11%. The gap — approximately 19,000 submissions per year — represents undetected violations that undermine academic standards and disadvantage honest students.

The group also needs to detect AI-generated content, an increasingly difficult task as language models improve. Existing AI detection tools have high false-positive rates on non-native English speakers’ work, creating equity concerns. The group needs a system that analyses writing at a deeper semantic level, processing submissions through its own UK-hosted infrastructure to avoid sending student work to external services.

AI Solution: Semantic Similarity and AI Content Detection

An embedding-based plagiarism detection system works by encoding every submission into dense semantic vectors and comparing them against: the group’s historical submission archive (cross-referencing 1.2 million past submissions), a reference corpus of known essay mill content, and established source material. Unlike string matching, semantic similarity catches paraphrased content because the meaning vectors remain close even when every word changes.

A second model — a fine-tuned classifier trained to distinguish human and AI writing patterns — provides AI-generated content detection with lower false-positive rates than commercial alternatives, particularly for non-native speakers. Both models run on a dedicated GPU server, processing submissions in batch overnight and flagging results for academic review.

GPU Requirements

Processing 280,000 submissions annually (roughly 23,000 per month during term) requires encoding each submission into vectors and comparing against the full archive. The encoding step is GPU-intensive; the comparison step leverages the vector index.

An RTX 5090 processes the monthly volume overnight with hours to spare. Private AI hosting ensures all student submissions remain within GDPR-compliant UK infrastructure.

Recommended Stack

• Sentence Transformers (all-MiniLM-L12 or E5-large) for encoding submissions into semantic vectors.
• FAISS or Qdrant for storing the archive of 1.2 million submission vectors and performing rapid similarity search.
• Fine-tuned DeBERTa for AI-generated content classification, trained on a balanced dataset of human and AI writing.
• FastAPI backend accepting submission uploads and returning similarity scores with flagged passages.
• VLE integration (Moodle/Canvas) via LTI for seamless workflow integration.

For processing handwritten exam submissions, add document AI for digitisation. Deploy an LLM via vLLM to generate detailed academic integrity reports from detection results.

Cost Analysis

The commercial plagiarism service costs the group approximately £350,000 per year across six institutions. The semantic similarity system supplements rather than replaces the existing tool initially, adding a deeper detection layer. The GPU server cost for running the additional analysis is a fraction of the existing contract. Over time, the group may reduce reliance on the external service, recapturing a significant portion of that annual spend.

Detecting the estimated 19,000 additional misconduct cases per year protects degree integrity, reduces grade inflation from undetected cheating, and supports fair assessment for the 90% of students who submit honest work.

Getting Started

Build your submission archive by encoding the last three years of student submissions (anonymised) into semantic vectors. Train the AI detection classifier on a dataset of 5,000 human-written and 5,000 AI-generated essays across disciplines. Run both systems in parallel with your existing tool for one semester, comparing detection rates and false-positive rates before making policy decisions based on the new system’s flags.

GigaGPU provides UK-based dedicated GPU servers for academic integrity workloads. Add an AI chatbot for student queries about academic integrity policies.