AI-generated Text Detection: A Multifaceted Approach to Binary and Multiclass Classification

TL;DR

The paper addresses the challenge of detecting AI-generated text and attributing it to the originating LLM within the Defactify AAAI 2025 shared task. It extends prior work by proposing two architectures—an Optimized model and a Simple model—capable of both binary and multiclass classification. The Optimized architecture fuses RoBERTa-base detector outputs, E5 embeddings, and stylometric features to produce document-level representations, while the Simple architecture uses E5 embeddings plus stylometry fed into a gradient-boosting classifier for multiclass discrimination. On the test set, the Optimized architecture achieves F1=0.994 for Task A, and the Simple architecture achieves F1=0.627 for Task B, indicating strong performance for binary detection and a path to improve multiclass attribution. Overall, the work highlights the utility of stylometry and multi-representation fusion in AI-generated text detection and model attribution, with clear avenues for enhancing multiclass performance.

Abstract

Large Language Models (LLMs) have demonstrated remarkable capabilities in generating text that closely resembles human writing across a wide range of styles and genres. However, such capabilities are prone to potential misuse, such as fake news generation, spam email creation, and misuse in academic assignments. As a result, accurate detection of AI-generated text and identification of the model that generated it are crucial for maintaining the responsible use of LLMs. In this work, we addressed two sub-tasks put forward by the Defactify workshop under AI-Generated Text Detection shared task at the Association for the Advancement of Artificial Intelligence (AAAI 2025): Task A involved distinguishing between human-authored or AI-generated text, while Task B focused on attributing text to its originating language model. For each task, we proposed two neural architectures: an optimized model and a simpler variant. For Task A, the optimized neural architecture achieved fifth place with $F1$ score of 0.994, and for Task B, the simpler neural architecture also ranked fifth place with $F1$ score of 0.627.

Figures (2)

• Figure 1: Proposed optimized neural architecture
• Figure 2: Simple architecture