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

understanding-the-effects-of-human-written-paraphrases-in-llm-generated-text-detection

/pdfs/understanding-the-effects-of-human-written-paraphrases-in-llm-generated-text-detection.pdf

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

The paper "**Understanding the Effects of Human-written Paraphrases in LLM-generated Text Detection**" investigates how human paraphrasing impacts the reliability of AI text detection systems. The research examines the effectiveness of current detection methods when faced with *human-modified AI-generated content*. Through extensive experiments, the authors demonstrate that even minor human edits can significantly reduce detection accuracy, with paraphrasing reducing detection rates by up to 50%. The study identifies **key patterns** in how different types of human modifications affect detector performance, from simple word substitutions to complex structural changes. Results show that current detectors struggle most with *hybrid content* combining AI generation and human editing. The research provides valuable insights for developing more robust detection systems and highlights the challenges in distinguishing between purely AI-generated text and human-edited AI content. These findings have important implications for *academic integrity*, *content authenticity*, and *digital forensics*.