With the rapid advancement of large language models (LLMs), extremely complex texts that are identical to human-written ones have been generated. However, this also raises risks such as misinformation and academic dishonesty. As the responsible usage of LLMs becomes increasingly critical, identifying content generated by LLMs has become an essential task. Despite various proposed detection methods, a thorough understanding of their successes and limitations is still lacking. This paper presents a review of the literature on detecting texts generated by LLMs. LLM-generated text detection approaches are neural-based, feature-based, watermarking, and humanassisted; this paper highlights these approaches. This paper highlights significant challenges of existing detection approaches, such as robustness against text perturbation, generalization across domains and models, lack of interpretability, reliance on model access constraints (black-box vs. white-box), scalability limitations, computational costs, sensitivity to text length, and data scarcity for training. Furthermore, the quick development of LLMs continues to pose evolving challenges to detection techniques.
Paper
Full text
Detection of Large Language Model-Generated Text: A Comprehensive Review
Semantic Scholar · 2025
Abstract
With the rapid advancement of large language models (LLMs), extremely complex texts that are identical to human-written ones have been generated. However, this also raises risks such as misinformation and academic dishonesty. As the responsible usage of LLMs becomes increasingly critical, identifying content generated by LLMs has become an essential task. Despite various proposed detection methods, a thorough understanding of their successes and limitations is still lacking. This paper presents a review of the literature on detecting texts generated by LLMs. LLM-generated text detection approaches are neural-based, feature-based, watermarking, and humanassisted; this paper highlights these approaches. This paper highlights significant challenges of existing detection approaches, such as robustness against text perturbation, generalization across domains and models, lack of interpretability, reliance on model access constraints (black-box vs. white-box), scalability limitations, computational costs, sensitivity to text length, and data scarcity for training. Furthermore, the quick development of LLMs continues to pose evolving challenges to detection techniques.