Contrastive siamese network for detecting AI-generated text across domains and models

Abstract

The rapid proliferation of large language models (LLMs) has raised growing concerns about distinguishing between human-written and AI-generated text. This work addresses the task of detecting AI-generated content by evaluating the latent similarity between a given input text and an alternative response generated for the same prompt, either known or inferred. Accordingly, CLAID (Contrastive Learning for AI Detection) is proposed as a Siamese Neural Network architecture utilizing BERT encoders and contrastive loss to capture semantic similarity between text pairs. Unlike prior approaches that rely on explicit classification or domain-specific features, our method focuses on modeling pairwise similarity, enabling a flexible and model-agnostic detection framework. To evaluate the generalization capabilities of the system, a comprehensive multi-domain and multi-model benchmark comprising three diverse datasets (i.e., HC3, DAIGT, and OUTFOX), encompassing a wide range of text genres, prompt structures, and generative models, has been constructed. Experimental results show that the proposed model achieves near-perfect classification accuracy across both single-domain and mixed-domain scenarios, demonstrating strong robustness to domain shifts, prompt variability, and authorship ambiguity. The model also exhibits strong data efficiency, attaining high performance with minimal supervision.

Ähnliche Arbeiten