Title: 'Your AI Text is not Mine': Redefining and Evaluating AI-generated Text Detection under Realistic Assumptions

Abstract: While there is widespread consensus that AI-generated content presents significant societal risks, the field of AI-generated text detection lacks a unified definition of what constitutes a harmful application. Current methodologies and datasets frequently establish their own, often implicit, criteria that bear only a loose connection to practical, real-world requirements. To bridge this conceptual gap, this study systematically delineates different definitions of AI-generated text and analyzes their specific attributes. To facilitate this investigation, we introduce AITDNA, a novel benchmark comprising human-machine co-authored texts annotated with comprehensive genesis data, including full histories of editing and AI interaction. Our evaluation of existing machine-generated text detectors reveals that while these tools may excel under specific definitions, they generally fail to function as effective broad-spectrum detectors. The associated code and data are made publicly available.

Source: arXiv Generated at: 2026-06-04 00:00:00 UTC