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

Neural network-based detectors generally outperform metric-based methods in both human authorship attribution and LLM-generated text detection.

“Neural network-based detectors generally outperform metric-based methods in both human authorship attribution and LLM-generated text detection problems [He et al., 2023;Zhang et al., 2024].”

Human-authored texts often feature unique stylistic nuances, such as variable sentence structures and emotive language.

“Human-authored texts often feature unique stylistic nuances, such as variable sentence structures and emotive language [Zhang et al., 2024].”

Other researchers handle LLM-revised human texts and human-revised LLM texts as a single category, alongside purely human-written and purely LLM-generated texts, approaching human-LLM co-authored text authorship attribution as a three-class classification problem.

“Other researchers handle LLM-revised human texts and human-revised LLM texts as a single category, alongside purely human-written and purely LLM-generated texts, approaching human-LLM co-authored text authorship attribution as a three-class classification problem [Zhang et al., 2024;Richburg et al., 2024].”

MIXSET offers a more comprehensive dataset that includes text refined by LLMs through polishing, completion, and rewriting operations.

“MIXSET [Zhang et al., 2024] offers a more comprehensive dataset that includes text refined by LLMs through polishing, completion, and rewriting operations.”

Human-authored texts tend to be more coherent and exhibit greater lexical diversity compared to LLM-generated texts.

“Human-authored texts tend to be more coherent and exhibit greater lexical diversity compared to LLM-generated texts [Guo et al., 2023;Dugan et al., 2023;Zhang et al., 2024].”

Self-BLEU are useful metrics, with a lower score indicating higher textual diversity.

“Additionally, Self-BLEU are useful metrics, with a lower score indicating higher textual diversity [Zhang et al., 2024].”

Current work simplifies human-written text, LLM-generated content, and texts coauthored by humans and LLMs into three broad categories, without differentiating between individual human authors and specific LLMs.

“Current work simplifies human-written text, LLM-generated content, and texts coauthored by humans and LLMs into three broad categories, without differentiating between individual human authors and specific LLMs [Zhang et al., 2024;Richburg et al., 2024].”