Am J Obstet Gynecol. 2024 May 04. pii: S0002-9378(24)00571-4. [Epub ahead of print]
BACKGROUND: ChatGPT, a publicly available artificial intelligence (AI) large language model, has allowed for sophisticated AI technology on demand. Indeed, use of ChatGPT has already begun to make its way into medical research. However, the medical community has yet to understand the capabilities and ethical considerations of AI within this context, and unknowns exist regarding ChatGPT's writing abilities, accuracy, and implications for authorship.OBJECTIVES: We hypothesize that human reviewers and AI detection software differ in their ability to correctly identify original published abstracts and AI-written abstracts in the subjects of Gynecology and Urogynecology. We additionally suspect that concrete differences in writing errors, readability, and perceived writing quality exist between original and AI-generated text.
STUDY DESIGN: Twenty-five articles published in high impact medical journals and a collection of Gynecology and Urogynecology journals were selected. ChatGPT was prompted to write 25 corresponding AI-generated abstracts, providing the abstract title, journal-dictated abstract requirements, and select original results. The original and AI-generated abstracts were reviewed by blinded Gynecology and Urogynecology faculty and fellows to identify the writing as original or AI-generated. All abstracts were analyzed by publicly available AI detection software GPTZero, Originality, and Copyleaks and were assessed for writing errors and quality by AI writing assistant Grammarly.
RESULTS: One hundred fifty-seven reviews of 25 original and 25 AI-generated abstracts were conducted by 26 faculty and 4 fellows. Fifty-seven percent of original abstracts and 42.3% of AI-generated abstracts were correctly identified for an average of 49.7% across all abstracts. All three AI detectors rated the original abstracts as less likely be AI-written than the ChatGPT-generated abstracts (GPTZero 5.8 vs 73.3%, p<0.001; Originality 10.9 vs 98.1%, p<0.001; Copyleaks 18.6 vs 58.2%, p<0.001). The performance of the three AI detection software differed when analyzing all abstracts (p=0.03), original abstracts (p<0.001), and AI-generated abstracts (p<0.001). Grammarly text analysis identified more writing issues and correctness errors in original than AI abstracts, including lower Grammarly score reflective of poorer writing quality (82.3 vs 88.1, p=0.006), more total writing issues (19.2 vs 12.8, p<0.001), critical issues (5.4 vs 1.3, p<0.001), confusing words (0.8 vs 0.1, p=0.006), misspelled words (1.7 vs 0.6, p=0.02), incorrect determiner use (1.2 vs 0.2, p=0.002), and comma misuse (0.3 vs 0.0, p=0.005).
CONCLUSIONS: Human reviewers are unable to detect the subtle differences between human and ChatGPT-generated scientific writing due to AI's ability to generate tremendously realistic text. AI detection software improve identification of AI-generated writing but still lack complete accuracy and require programmatic improvements in order to achieve optimal detection. As reviewers and editors may be unable to reliably detect AI-generated pieces, clear guidelines for reporting AI use by authors and implementing AI detection software in the review process will need to be established as AI chatbots gain more widespread use.
Keywords: AI chatbots; AI detection; AI ethics; AI writing; artificial intelligence; authorship; large language models; plagiarism; research ethics; research policy