생성형 인공지능을 활용한 방사선학 연구의 GPT 탐지 표절 의심률 분석 = Analysis of Plagiarism Suspicion Rates Detected by GPT in Radiological Science Research Using Generative Artificial Intelligence

국문 초록 (Abstract)

본 연구는 생성형 인공지능(Artificial Intelligence)을 활용한 방사선학 분야 논문을 대상으로 탐지된 GPT (Generative Pre-trained Transformer) 표절 의심률을 분석하였다. 한국연구재단 등재 학술지인 한국방사선학회논문지와 방사선기술과학지에 2022년 12월부터 2025년 6월까지 게재된 447편 논문을 대상으로 KCI 문헌 유사도 검사와 ㈜무하우의 카피킬러(GPT Killer 기능 포함)를 이용하여 유사도를 비교하였다. KCI 문헌 유사도는 6.8%±3.3%, 카피킬러 표절 의심률은 16.9%±11.1%로 유의한 차이를 보였으며(p<0.001), GPT 생성이 의심되는 논문은 전체의 27.1%였다. GPT 의심 논문 군의 KCI 유사도는 5.7%±2.7%, 카피킬러 표절 의심률은 27.1%±15.4%로 평균 20% 이상 높게 나타났으며, GPT 생성 의심 논문의 비율은 시계열적으로 증가하는 추세를 보였다. GPT 생성 문장은 주로 결론(19.2%), 토의(17.7%), 서론(14.8%) 순으로 많이 탐지되었고, GPT 탐지 문장의 누적 빈도가 증가할수록 표절 의심률이 상승하였다. 그 결과, GPT 활용 정도에 따라 표절 의심률 기준을 “양호(0% - 20% 미만)”, “유의(20% - 30% 미만)”, “의심(30% - 50% 미만)”, “위험(50% 이상)”의 4단계 운영기준을 제안한다. AI는 논문 작성에서 연구자의 전문성·경험·창의성·비판적 사고를 보조하는 도구로 한정적으로 활용되어야 하며, 최종 책임은 인간 저자에게 있음을 재확인해야 한다.

다국어 초록 (Multilingual Abstract)

This study analyzed the plagiarism suspicion rate associated with the use of Generative Pre-trained Transformer (GPT)-based generative artificial intelligence in the field of radiological science. The study compared the similarity of 447 papers published in the Journal of the Korean Society of Radiology and the Journal of Radiological Science and Technology, both journals registered with the National Research Foundation of Korea, from December 2022 to June 2025, using the Korea Citation Index (KCI) literature similarity and MUHAYU Co., Ltd.'s Copykiller (including the GPT Killer function). The KCI literature similarity rate was 6.8%±3.3%, and the Copykiller plagiarism suspicion rate was 16.9%±11.1%, showing a significant difference (p<0.001). 27.1% of the papers were suspected of being generated by GPT. The KCI similarity of the suspected GPT papers was 5.7%±2.7%, and the Copykiller plagiarism suspicion rate was 27.1%±15.4%, which was more than 20% higher on average. The proportion of suspected GPT-generated papers showed a time-series increase. The GPT-generated sentences were mainly detected in the order of conclusion (19.2%), discussion (17.7%), and introduction (14.8%), and the plagiarism suspicion rate increased as the cumulative frequency of GPT-detected sentences increased. As a result, we propose a four-level operating standard for the plagiarism suspicion rate according to the degree of GPT utilization: “Good (0% - less than 20%)”, “Significant (20% - less than 30%)”, “Suspicious (30% - less than 50%)”, and “Risk (50% or more).” It should be reaffirmed that artificial intelligence should be used only as a tool to assist researchers’ expertise, experience, creativity, and critical thinking in paper writing, and that the ultimate responsibility lies with human authors.