Scientists at the University of Kansas have created an AI-detection tool capable of identifying scientific text generated by AI with 99% accuracy, specifically targeting outputs from systems like ChatGPT. Developed by chemist Heather Desaire and her team, the tool distinguishes AI-generated text from human-written scientific literature found in peer-reviewed journals, improving accuracy compared to broader detection methods. The study, published in Cell Reports Physical Science, highlights the urgent need for precise detection to uphold scientific integrity, as AI generators can fabricate facts that infiltrate academic publications. Desaire’s approach focuses on the distinctive traits of scientific writing, utilizing a small, carefully curated dataset comprising only 64 human-written and 128 AI-generated documents. This method relies significantly on human insight rather than traditional large-scale machine-learning strategies, allowing the team to discern subtle differences effectively. Desaire aims to make this tool accessible to other researchers, encouraging those without formal computer science backgrounds to engage in AI and detection research. Her work emphasizes that with the right guidance, individuals from various fields can contribute valuable insights to the evolving domain of AI technology.
