Paper Title
Glass-Box Academic Integrity: Explainable Structural Plagiarism and AI-Generated Text Detection using Graph Attention Networks

Abstract
With the introduction of large language models (LLMs), plagiarism has evolved from lexical copying to conceptual rewriting, in which the semantic conceptual structure is preserved but the surface-level wording changes. Lexical similarity, semantic embeddings, or perplexity scores are the foundations of traditional plagiarism and AI-written text detection techniques, which are frequently black-box binary outputs devoid of any meaningful explanation. This paper presents an explainable hybrid detection system that uses a Graph Attention Network (GAT) to identify structural plagiarism and semantically represents documents as conceptual graphs.The attention weights serve as localized evidence by drawing attention to questionable conceptual connections that are present in different documents. In order to detect machine-written text behavior, a stylometric component simultaneously analyzes lexical diversity and variance features. To effectively detect copy-and-paste, paraphrased, and AI-rewritten text in a single, cohesive framework, a score fusion layer combines transformer-based semantic similarity and graph-level structural similarity scores. With 87.4% accuracy and enhanced recall on highly paraphrased instances, experiments on a PAN-style dataset using LLM paraphrases demonstrate statistically significant improvement over TF-IDF, SBERT, and BERT baselines.. An edge removal validation test also verifies that high-attention edges are indicative of meaningful explanatory evidence.. The proposed approach bridges performance and interpretability, providing verifiable justification rather than black-box verdicts, and supports trustworthy academic integrity assessment in the era of generative AI. Keywords - AI Writing Identification, Concept Graph Modeling, Graph Attention Networks, Structural Similarity Analysis, Stylometric Features, Semantic Embeddings, Hybrid AI Systems, Trustworthy Artificial Intelligence.