The great detectives: humans versus AI detectors in catching large language model-generated medical writing

Background: The application of artificial intelligence (AI) in academic writing has raised concerns regarding accuracy, ethics, and scientific rigour. Some AI content detectors may not accurately identify AI-generated texts, especially those that have undergone paraphrasing. Therefore, there is a pressing need for efficacious approaches or guidelines to govern AI usage in specific disciplines. Objective: Our study aims to compare the accuracy of mainstream AI content detectors and human reviewers in detecting AI-generated rehabilitation-related articles with or without paraphrasing. Study design: This cross-sectional study purposively chose 50 rehabilitation-related articles from four peer-reviewed journals, and then fabricated another 50 articles using ChatGPT. Specifically, ChatGPT was used to generate the introduction, discussion, and conclusion sections based on the original titles, methods, and results. Wordtune was then used to rephrase the ChatGPT-generated articles. Six common AI content detectors (Originality.ai, Turnitin, ZeroGPT, GPTZero, Content at Scale, and GPT-2 Output Detector) were employed to identify AI content for the original, ChatGPT-generated and AI-rephrased articles. Four human reviewers (two student reviewers and two professorial reviewers) were recruited to differentiate between the original articles and AI-rephrased articles, which were expected to be more difficult to detect. They were instructed to give reasons for their judgements. Results: Originality.ai correctly detected 100% of ChatGPT-generated and AI-rephrased texts. ZeroGPT accurately detected 96% of ChatGPT-generated and 88% of AI-rephrased articles. The areas under the receiver operating characteristic curve (AUROC) of ZeroGPT were 0.98 for identifying human-written and AI articles. Turnitin showed a 0% misclassification rate for human-written articles, although it only identified 30% of AI-rephrased articles. Professorial reviewers accurately discriminated at least 96% of AI-rephrased articles, but they misclassified 12% of human-written articles as AI-generated. On average, students only identified 76% of AI-rephrased articles. Reviewers identified AI-rephrased articles based on ‘incoherent content’ (34.36%), followed by ‘grammatical errors’ (20.26%), and ‘insufficient evidence’ (16.15%). Conclusions and relevance: This study directly compared the accuracy of advanced AI detectors and human reviewers in detecting AI-generated medical writing after paraphrasing. Our findings demonstrate that specific detectors and experienced reviewers can accurately identify articles generated by Large Language Models, even after paraphrasing. The rationale employed by our reviewers in their assessments can inform future evaluation strategies for monitoring AI usage in medical education or publications. AI content detectors may be incorporated as an additional screening tool in the peer-review process of academic journals.