AI Detectors Fail to Distinguish AI Editing from Plagiarism

A recent study investigates the efficacy of commercial AI detection tools used to uphold academic integrity, revealing significant flaws in their operation. The research indicates that these detectors struggle to differentiate between content that has been lightly edited or refined using AI tools and content that has been entirely generated by large language models (LLMs). This fundamental limitation poses a serious challenge for educational institutions attempting to implement fair policies regarding AI assistance. In a controlled experiment, the study analyzed published English abstracts, comparing older human-written texts with newer ones and those subjected to light AI editing. The findings showed that AI-edited abstracts, representing guideline-compliant AI assistance, were flagged as AI-generated by detectors like Pangram and GPTZero at rates between 64% and 80%. Conversely, unmodified original texts from 2023-2025 were flagged at 9-15%, with non-STEM fields showing higher rates. The critical implication is that students who use AI responsibly for editing or refining their work face a higher risk of being falsely accused of misconduct than those who might use "humanizer" tools to obscure full LLM drafts. The study concludes that AI detector scores should not be used as the sole evidence for academic misconduct, highlighting a policy failure in their current application.