整形外科論文におけるAI生成テキスト：横断的分析

Large language models are increasingly being used in scientific writing, but their use in orthopaedic literature remains unclear.

We analyzed 196 articles published in March 2025 in 10 leading orthopaedic journals. GPTZero quantified artificial intelligence (AI)-generated text by article section. Composite AI scores were calculated and tested for associations with the h5-index, study design, level of evidence, authorship characteristics, and geographic region with use of nonparametric and both Pearson and Spearman correlation analyses.

AI-generated content was detected in 89.8% of articles. The mean AI score was 18.1% (median, 14.9%). Scores differed by section (p < 0.001) and were the highest in the Results. AI use correlated with the proportion of non-MD authors (ρ = 0.22) and with the total author count (ρ = 0.19), but not with the h5-index. No association with study design or level of evidence was found. Differences by geographic region were modest and not significant after correction.

AI-generated content appears to be widespread, particularly in Results sections. Its use varies by authorship characteristics and geography but not by study design or journal prestige. Clear disclosure standards are essential to guide responsible AI use in scientific writing.

This study is clinically relevant because transparency in scientific writing supports accurate interpretation of the evidence used in patient care. Identifying the prevalence of AI-generated text helps to protect the integrity of the orthopaedic literature.