Understanding the Effects of Human-written Paraphrases in LLM-generated Text Detection