机器生成文本检测综述

A Comprehensive Survey of Machine-Generated Text Detection

摘要: 随着人工智能生成内容(AIGC)与人类文本之间的界限日益模糊，机器生成文本检测成为自然语言处理的重要研究方向。文章综述机器生成文本检测的技术演变，包括主动嵌入隐秘信号的水印技术、基于特征统计的传统方法、利用预训练语言模型(如RoBERTa、DeBERTa)进行判别的监督学习方法，以及结合模型预测不确定性、特征分布差异的概率检测方法。近年来，局部化检测与可解释性分析成为新的研究热点，使检测系统能够识别具体生成片段并解释判别依据。然而，跨模型泛化、多语言场景与对抗鲁棒性仍是亟待解决的难题。未来的研究将致力于构建具有更强鲁棒性和可解释性的检测框架，结合因果推理与多模态信息提升检测性能，助力推动LLM生成文本检测技术的实用化与规范化发展。

Abstract: As the boundary between Artificial Intelligence-Generated Content (AIGC) and human-written text becomes increasingly blurred, machine-generated text detection has emerged as a critical research direction in natural language processing. This paper reviews the technological evolution of machine-generated text detection, including watermarking techniques that actively embed hidden signals, traditional methods based on feature statistics, supervised learning approaches leveraging pre-trained language models (e.g., RoBERTa, DeBERTa) for discrimination, and probability-based detection methods that incorporate model prediction uncertainty and feature distribution differences. In recent years, localized detection and interpretability analysis have become new research hotspots, enabling detection systems to identify specific generated segments and explain the basis for discrimination. However, cross-model generalization, multilingual scenarios, and adversarial robustness remain pressing challenges to be addressed. Future research will focus on constructing detection frameworks with enhanced robustness and interpretability, integrating causal reasoning and multimodal information to improve detection performance, thereby advancing the practicalization and standardization of LLM-generated text detection technologies.