An exploratory study on the effectiveness of AI detection tools in identifying AI-generated articles

Accurate identification of AI-generated content is critical for preserving scientific credibility. This exploratory study was performed to assess the effectiveness of eight AI detection tools (free versions) in differentiating human-written from AI-generated articles within the oral and maxillofacial surgery field. The analysis included 24 human-written articles and 12 AI-generated articles produced using ChatGPT, DeepSeek, Gemini, and Copilot. The primary outcome was the detection effectiveness of each tool, expressed as a mean percentage score, for human-written and AI-generated text. Secondary outcomes were usability and processing limitations. The statistical analysis was performed in RStudio (P < 0.05). For published human-written text, QuillBot showed perfect detection (none detected as AI-written), and was fast and easy to use. For the AI texts, Copyleaks performed best (mean score 99.6/100), followed by Sapling (mean score 95.6/100). A weak, non-significant correlation was found between manuscript length and detection effectiveness for published human-written (ρ = -0.15, P = 0.44) and AI-generated texts (ρ = -0.08, P = 0.70). QuillBot appears to be an accessible and effective tool for distinguishing human- from AI-generated text. Its effectiveness could be enhanced when used alongside other detection tools like Sapling or Copyleaks, allowing articles produced with excessive reliance on AI to be detected.