LLM-as-a-Coauthor: Can Mixed Human-Written and Machine-Generated Text Be Detected?

TL;DR

This paper investigates the challenge of detecting mixed human and machine-generated text, introduces a new dataset called MixSet, and evaluates existing detectors, revealing their limitations in identifying subtle mixtext scenarios.

Contribution

It defines mixtext, creates the first dataset for it, and assesses current detectors, highlighting the need for more fine-grained detection methods.

Findings

Existing detectors struggle with mixtext detection.

Detectors are less effective on subtle modifications.

Current methods lack robustness and generalization.

Abstract

With the rapid development and widespread application of Large Language Models (LLMs), the use of Machine-Generated Text (MGT) has become increasingly common, bringing with it potential risks, especially in terms of quality and integrity in fields like news, education, and science. Current research mainly focuses on purely MGT detection without adequately addressing mixed scenarios, including AI-revised Human-Written Text (HWT) or human-revised MGT. To tackle this challenge, we define mixtext, a form of mixed text involving both AI and human-generated content. Then, we introduce MixSet, the first dataset dedicated to studying these mixtext scenarios. Leveraging MixSet, we executed comprehensive experiments to assess the efficacy of prevalent MGT detectors in handling mixtext situations, evaluating their performance in terms of effectiveness, robustness, and generalization. Our findings reveal that existing detectors struggle to identify mixtext, particularly in dealing with subtle modifications and style adaptability. This research underscores the urgent need for more fine-grain detectors tailored for mixtext, offering valuable insights for future research. Code and Models are available at https://github.com/Dongping-Chen/MixSet.