An explainable framework for assisting the detection of AI-generated textual content

Abstract

The recent development of generative AI (GenAI) algorithms has allowed machines to create new content in a realistic way, driving the spread of AI-generated content (AIGC) on the Internet. However, generative AI models and AIGC have exacerbated several societal challenges such as security threats (e.g., misinformation), trust issues, ethical concerns, and intellectual property regulation, calling for effective detection methods and a better understanding of AI-generated vs. human-written content. In this paper, we focus on AI-generated texts produced by large language models (LLMs) and extend prior detection methods by proposing a novel framework that combines semantic information and linguistic features. Based on potential semantic and linguistic differences in AI vs. human writing, we design our Semantic-Linguistic-Detector (SemLinDetector) framework by integrating a transformer-based semantic encoder and a linguistic encoder with parallel linguistic representations. By comparing a series of benchmark models on datasets collected from various LLMs and human writers in multiple domains, our experiments show that the proposed detection framework outperforms other benchmarks in a consistent and robust manner. Moreover, our model interpretability analysis showcases our framework's potential to help understand the reasoning behind prediction outcomes and identify patterns of differences in AI-generated and human-written content. Our research adds to the growing space of GenAI by proposing an effective and responsible detection system to address the risks and challenges of GenAI, offering implications for researchers and practitioners to better understand and regulate AIGC.