Lightweight AI-Generated vs Human-Authored Text Classifier based on Transformer Hidden State Analysis Using MiniLMv2

The increasing realism and sophistication of text generated by large language models (LLMs) such as ChatGPT, GPT-3, and GPT-4 have introduced substantial challenges in distinguishing between AI-generated and human-authored content. This issue has significant implications across various sectors including education, journalism, cybersecurity, and digital communication ethics. As LLMs evolve rapidly in scale and capability, the ability to detect synthetic text is becoming vital to maintain information integrity, prevent misinformation, and support ethical digital discourse. Goal. This research aims to develop a lightweight, accurate, and computationally efficient model that can reliably differentiate between AI-generated and human-written texts. The solution targets real-world deployment scenarios where low-latency inference and minimal resource consumption are essential. Design. To meet these objectives, we fine-tuned MiniLMv2—an efficient transformer distilled from BERT—augmented with a binary classification head. A diverse, domain-balanced dataset of human and AI texts was used to train and evaluate the model. Techniques such as GELU activation, Adam-based optimization, and dropout regularization were employed to enhance generalization and stability. The training pipeline was implemented using the modular and scalable FastAI framework. Our model offers an interpretable and practical solution, contributing to the ongoing discourse on trust, transparency, and accountability in AI-generated content detection. This work is intended for researchers in NLP, developers creating AI transparency tools, policymakers focused on digital integrity, educators, content moderation teams, and cybersecurity professionals. It also has strong applicability in online learning platforms, digital publishing systems, and real-time moderation pipelines.