文献速递技术与安全

记录披露:放射学研究稿件中生成式人工智能使用情况的报告有限

Documenting Disclosure: Limited Reporting of Generative AI Usage in Radiology Research Manuscripts.
PubMed 文献 · Acad Radiol · 2025-07-17

该文献计量分析纳入九种要求披露LLM使用的放射学期刊的1998篇稿件,仅34篇(1.7%)声明使用LLM,其中ChatGPT占94.1%、主要用于可读性/语法(97.1%);披露率随时间无显著增加(OR 1.06,95%CI 0.98–1.16,p=0.15),与同行评审时长无关;二次研究披露率3.9%高于一次研究的1.3%(p<0.001,Cramer's V 0.08)。

要点

  • 1998篇放射学稿件中仅34篇(1.7%)披露LLM使用,二次研究3.9%高于一次研究1.3%(p<0.001)。
  • 披露者中ChatGPT占94.1%(32/34),用于可读性/语法者占97.1%(33/34)。
  • 披露率随时间无显著增加(OR 1.06,95%CI 0.98–1.16,p=0.15)。

核心结论

  • 尽管调查显示研究者广泛使用LLM,放射学稿件中LLM披露率极低(1.7%),二者存在差异,可能源于真实未使用、担心污名、认为不披露有优势、不认同对微小编辑的披露要求或对政策不了解等原因;作者认为需要更包容的研究环境并制定细致的披露政策。

研究设计与样本

  • 对九种要求披露LLM使用的放射学期刊进行文献计量分析,纳入一次研究和二次研究、排除短篇出版物,共分析1998篇稿件;采用逻辑回归评估披露率的时间趋势,线性混合效应模型评估披露状态与同行评审时长的关系,卡方分析检验稿件类型与披露率的关联。

诊断标准与关键数值

  • 总体LLM披露率为1.7%(34/1998)。
  • 披露者中ChatGPT占94.1%(32/34),用于可读性/语法目的者占97.1%(33/34)。
  • 披露LLM使用的稿件多数来自非英语国家机构(22/34,64.7%)。
  • 二次研究稿件披露率3.9% vs 一次研究稿件1.3%(p<0.001),效应量小(Cramer's V 0.08 [0.04, 1.00])。

局限性

  • 披露率极低可能反映真实未使用、担心污名、认为不披露有优势、不认同对微小编辑的披露要求或对政策不了解等原因,原文未进一步区分。

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)。

英文原文(摘要)· 点击展开

RATIONALE AND OBJECTIVES: Large language models (LLMs) show promise in radiology through various clinical applications as well as in assisting with research manuscript development. Recent studies show 52.6% of medical researchers use LLMs in manuscript development, with non-medical researchers showing similar rates. Given concerns about hallucinations, bias, etc., many publishers now require disclosure of LLM use. While most medical imaging journals have LLM policies as of 2025, the actual disclosure rates for LLM usage remain unknown. Our study examines trends in LLM disclosure by analyzing 1998 radiology publications for LLM disclosures.

MATERIALS AND METHODS: A bibliometric analysis of nine radiology journals with LLM use disclosure requirements was performed. The study included primary investigations and secondary research while excluding short-form publications. The LLM disclosure rate was calculated overall. Logistic regression assessed temporal trends in disclosure rates, while a linear mixed effects model evaluated the relationship between disclosure status and peer review duration. Chi-square analysis examined associations between manuscript type and disclosure rates.

RESULTS: Of 1998 manuscripts, 34 (1.7%) declared LLM use. Most disclosures involved ChatGPT (32, 94.1%), primarily for readability/grammar purposes (33, 97.1%). The majority of manuscripts disclosing LLM use originated from institutions in non-English speaking countries (22, 64.7%). No significant increase in disclosure rates over time was observed (OR: 1.06 [95% CI: 0.98, 1.16], p=0.15), and no relationship with peer review duration (coefficient: -4.85, SE=11.25, p=0.67) was found. Secondary research manuscripts disclosed LLM use more frequently (3.9% vs. 1.3%, p<0.001) with a small effect size (Cramer's V: 0.08 [0.04, 1.00]).

CONCLUSION: Our findings demonstrate remarkably low disclosure rates in radiology manuscripts despite surveys indicating significant LLM adoption among researchers. This discrepancy may result from true non-use, fear of stigma, perceived advantages of undisclosed use, disagreement with disclosure requirements for minor editing, or policy unawareness, among other reasons. These findings suggest a need for more accepting research environments that recognize legitimate LLM benefits while developing nuanced disclosure policies addressing risks.

来源:Acad Radiol,查看原文。版权归原作者所有。

引用

  1. Jonah Barrett D, Heng R, Perchik JD, et al. Documenting Disclosure: Limited Reporting of Generative AI Usage in Radiology Research Manuscripts. Acad Radiol. 2025. PMID: 40675866. doi:10.1016/j.acra.2025.06.057

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。