Which LLMs are Difficult to Detect? A Detailed Analysis of Potential Factors Contributing to Difficulties in LLM Text Detection

The paper investigates the detectability of AI-generated (AIG) texts across various LLM families (Llama, OpenAI, Claude, etc.) and writing domains. By training 92 DistilRoBERTa-based classifiers using the LibAUC library, the authors identify that OpenAI models, particularly when performing "rewriting" tasks, are significantly harder to detect due to their human-like linguistic complexity.