Hybrid Learning for AI-Generated Text Detection and Attribution

Abstract

The rapid advancement of large language models (LLMs) has significantly in- creased the quality and fluency of AI-generated text, making it increasingly difficult to

distinguish from human-written content. This challenge is further amplified in source attribution tasks, where texts generated by multiple LLMs often exhibit overlapping linguistic and stylistic characteristics. This study proposes a hybrid framework for AI-generated text detection and attribution by combining semantic and stylometric

representations to improve classification performance. The framework integrates con- textual semantic embeddings extracted using RoBERTa with handcrafted stylometric

features, while supervised contrastive learning is employed to enhance feature discrim- inability and class separability. The proposed approach addresses both binary classi- fication (human versus AI-generated text) and multi-class attribution for identifying

the generating model. The fused representation is processed through a multilayer per- ceptron (MLP) classifier for final prediction. Experimental evaluations conducted on

the MAGE Dataset demonstrate the effectiveness of the proposed framework, achiev- ing a 93.20% macro F1-score in binary classification and an 82.22% macro F1-score

in multi-class attribution. Comparative results show that the proposed model outper- forms traditional machine learning baselines, including Logistic Regression, Random

Forest, Support Vector Machine (SVM), demonstrating its robustness and effectiveness for AI-generated text detection and attribution tasks.