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

TL;DR

This work defines mixtext as the mixed-content result of AI and human revisions and introduces MixSet, the first dataset dedicated to studying such cases. Using MixSet, the authors conduct extensive evaluations of prevalent MGT detectors in binary and three-class settings, revealing that existing detectors struggle to reliably identify mixtext, with performance highly dependent on the revision operation and the generating model. Key findings show no clear preference for labeling mixtext as MGT or HWT, limited cross-LLM transfer, and that retraining on MixSet improves mixtext detection, particularly for model-based detectors like Radar and GPT-sentinel. The results argue for the development of finer-grained, mixtext-aware detectors and highlight MixSet as a critical resource for advancing detection robustness in real-world AI-human collaboration scenarios.

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.

Figures (37)

• Figure 1: Three kinds of text: Machine Generative Text (MGT), Human Written Text (HWT), and mixtext. The text come from users is classified by detectors . The text in red is the HWT polished by LLMs.
• Figure 2: Accuracy of different dectors on MixSet. (Above) Model-based methods; (Below) Metric-based methods. P.T. and P.S. signify token and sentence-level polish, respectively; C. for complete, R. for rewrite; Adapt T. and Adapt S. for token and sentence-level adapt. See \ref{['Sec:Section 3']} for details on revising operations.
• Figure 3: The process of MixSet generation. We perform distinct operations in HWT and MGT. In HWT, three operations—polish, rewrite, and complete—are employed. In MGT, we utilize LLama2 and GPT-4 to aid in humanization and conduct the adaptation operation manually.
• Figure 4: Length distribution of the HWT, MGT, and MixSet.
• Figure 5: Cosine similarity of the MixSet