Abstract:
"The rapid advancement of AI-generated content, particularly AI-generated text (AIGT), has
raised significant concerns across various sectors including media, education, and finance.
While Large Language Models (LLMs) like ChatGPT have revolutionized content creation,
they have also introduced challenges in distinguishing between human-authored and AIgenerated text. This research addresses the critical need for an effective and transparent method
to detect AI-generated content, focusing on the lack of explainable solutions in current
detection systems.
To bridge this gap, it’s proposed an innovative approach that integrates state-of-the-art AIgenerated text detection methodologies with explainable AI techniques. This solution aims to
not only identify AI-generated text across various domains but also provide clear explanations
for the classification decisions. This approach involves developing a domain-agnostic model
capable of analyzing texts of varying lengths and contexts, moving beyond the limitations of
existing methods that often focus on specific areas such as short online reviews.
Preliminary results from the ALBERT-based model, tested on a dataset of 21,833 comments,
demonstrate promising performance. The model achieved an accuracy of 77.2%, precision of
79.8%, recall of 77.2%, and an F1 score of 76.6%. These metrics indicate the potential
effectiveness of the approach in accurately detecting AI-generated text while providing
explainable insights into the classification process. This research contributes significantly to
enhancing the integrity and reliability of digital information by offering a transparent and
robust solution for AI-generated text detection.
"
