文献速递显微修复重建烧伤

无声的作者:一项检测整形与重建外科期刊中AI辅助写作的初步研究

The Silent Author: A Pilot Study Detecting AI-Assisted Writing in Plastic and Reconstructive Surgery Journals.
PubMed 文献 · Ann Plast Surg · 2026-03-23

该初步研究从2024年7月至2025年7月发表的整形与重建外科及烧伤期刊中抽取10%(n=67)文章,并与2014至2015年匹配对照比较,使用RoBERTa分类器和困惑度评估检测AI样文本特征。结果显示当代文章26.9%含至少一个被标记部分,显著高于对照的4.5%(P<0.001),提示需明确披露政策与审稿人培训。

要点

  • 当代整形外科文章中26.9%含AI样文本特征,显著高于2014-2015年的4.5%。
  • 被标记内容主要集中在方法学(9.1%)和摘要(7.5%)部分。
  • 检测不能确证作者身份,但提示需标准化披露政策和审稿培训。

核心结论

  • 当代整形与重建外科出版物中AI样文本特征比前AI时代队列更常见。
  • 检测不能确认作者身份,但强调需要更清晰一致的披露政策、标准化分级报告政策和审稿人培训。

研究设计与样本

  • 初步分析抽取2024年7月1日至2025年7月1日发表文章的10%(n=67),来自领先的整形与重建外科及烧伤期刊。
  • 使用相同标准组建2014至2015年匹配对照队列。
  • 使用RoBERTa分类器结合基于困惑度的评估分析文章,标记潜在AI样文本特征。

主要发现

  • 文章层面,67篇中18篇(26.9%,95%CI:17.7-38.5)含至少一个被标记部分。
  • 共20个子部分被标记(5.3%,95%CI:3.5-8.1)。
  • 被标记内容集中在方法学(9.1%)和摘要(7.5%),其他部分较低。
  • 对照队列中3篇论文(4.5%,95%CI:1.5-12.5)和3个子部分(0.9%,95%CI:0.3-2.7)被标记,分别约增加6倍和5倍。
  • 文章层面差异有统计学意义(χ²=11.1,P<0.001)。

临床怎么用

  • 临床医生阅读整形外科文献时应意识到AI辅助写作的可能性,关注方法学和摘要部分的AI样特征。
  • 期刊和审稿人需采用标准化披露政策和审稿人培训,以促进AI在外科出版中的负责任整合。

局限性

  • 检测工具不完美,检测不能确认作者身份。
  • 期刊缺乏披露政策共识,快速采用与有效监管之间存在差距。

以上解读依据论文摘要(全文未开放获取或许可不允许翻译转载)。

原文

BACKGROUND: Academic publishing underpins surgical decision-making, but the rapid adoption of generative artificial intelligence (AI) raises concerns about research credibility and patient safety. To the best of our knowledge, no prior pilot study has examined its presence in plastic and reconstructive surgery. Detection tools remain imperfect, and journals lack consensus on disclosure policies, leaving a gap between rapid adoption and effective oversight. METHODS: This pilot analysis sampled 10% (n=67) of articles published between July 1, 2024, and July 1, 2025, across leading plastic and reconstructive surgery and burn journals. We assembled a matched control cohort from 2014 to 2015 using identical criteria. Articles were analyzed using a combined RoBERTa classifier and perplexity-based evaluation to flag potential AI-like textual characteristics. RESULTS: At the article level, 18 of 67 articles (26.9%, 95% CI: 17.7-38.5) contained ≥1 flagged section, with 20 subsections (5.3%, 95% CI: 3.5-8.1) flagged. Flagged content clustered in methods (9.1%) and abstracts (7.5%), with lower prevalence in other sections. In our control cohort, 3 papers (4.5%, 95% CI: 1.5-12.5) and 3 subsections were flagged (0.9%, 95% CI: 0.3-2.7), representing ~6-fold and 5-fold increases, respectively. This difference was statistically significant at the article level (χ²=11.1, P <0.001). CONCLUSIONS: AI-like textual characteristics were more frequently detected in contemporary plastic and reconstructive surgery publications than in the pre-AI cohort. Although detection does not confirm authorship, these findings underscore the need for clearer and more consistent disclosure, standardized and graded reporting policies, and reviewer training to enable responsible integration of AI into surgical publishing.

来源:Ann Plast Surg,查看原文。版权归原作者所有。

引用

  1. Happonen HDI, Bollard SM, Dolan RT, et al. The Silent Author: A Pilot Study Detecting AI-Assisted Writing in Plastic and Reconstructive Surgery Journals. Ann Plast Surg. 2026. PMID: 41860427. doi:10.1097/SAP.0000000000004726

供执业医师工作参考,不替代临床判断;数值与分级以指南原文为准。