A Deep Learning Framework for Detecting AI-Generated and Paraphrased Plagiarism in Research Contents

Abstract

The proliferation of large language models (LLMs) such as GPT-3.5, GPT-4, and their open-source successors has fundamentally altered the landscape of academic writing, blurring the boundary between human-authored and machine-generated scholarly text. Simultaneously, paraphrase-based plagiarism — long a challenge for string-matching plagiarism checkers — has become easier to produce and harder to detect as paraphrasing tools themselves increasingly rely on neural sequence-to-sequence models. Conventional plagiarism-detection systems, which rely primarily on lexical overlap and fingerprinting, are structurally unable to catch either AI-generated substitution or deep semantic paraphrase, since both preserve meaning while altering surface form [1,2]. This paper proposes an integrated deep learning framework that combines a shared transformer encoder with two specialised branches — an AI-generated text detection branch exploiting perplexity and burstiness statistics, and a paraphrase/semantic-similarity branch built on a Siamese cross-encoder architecture — fused through a weighted ensemble layer and accompanied by an attention-based explainability module. We review 40 works spanning classical plagiarism detection, statistical and neural AI-text detectors, and paraphrase identification, synthesise their reported performance figures into a comparative framework, and analyse the problem from technological, pedagogical, ethical, and publishing-industry perspectives. We further incorporate practice-based insights on plagiarism-free content creation and AI's dual role — as both a threat to and a defender of academic integrity — drawn from recent applied scholarship [3–5]. The proposed framework is positioned not as a replacement for editorial judgement but as a decision-support instrument that produces an interpretable originality report rather than a single opaque similarity score. We conclude with a discussion of limitations, including adversarial evasion, dataset bias, and the moving target problem posed by continuously improving generative models, and outline a research agenda for robust, explainable, and fair AI-assisted integrity screening.