Documenting Disclosure: Limited Reporting of Generative AI Usage in Radiology Research Manuscripts.

Area of Science:

• Radiology
• Medical Informatics
• Scientific Publishing

Background:

• Large language models (LLMs) are increasingly used in medical research and manuscript development.
• Publisher policies requiring LLM use disclosure are becoming common, yet actual disclosure rates are unknown.
• Concerns regarding LLM-generated content (e.g., hallucinations, bias) necessitate transparent reporting.

Purpose of the Study:

• To examine the trends and rates of LLM use disclosure in radiology publications.
• To analyze factors associated with LLM disclosure, including manuscript type and institutional origin.

Main Methods:

• Bibliometric analysis of 1998 radiology publications across nine journals with LLM disclosure policies.
• Calculation of overall LLM disclosure rates.
• Logistic regression and mixed-effects models to assess temporal trends and relationships with peer review duration.

Main Results:

• Only 1.7% (34 of 1998) of manuscripts declared LLM use, primarily ChatGPT for readability.
• No significant increase in disclosure rates over time was observed.
• Secondary research manuscripts showed higher disclosure rates (3.9%) compared to primary investigations (1.3%).

Conclusions:

• Disclosure rates for LLM use in radiology publications are remarkably low, contrasting with reported adoption rates.
• Potential reasons for low disclosure include stigma, policy unawareness, or disagreement with requirements.
• There is a need for supportive research environments and clearer disclosure policies for LLMs in scientific writing.