Earticle

초록

영어

This study empirically validates the limitations of Vision AI(GPT-5.2) and the efficacy of Human-in-the-Loop collaboration using 280 real-world fire cases(N=280). While independent AI analysis showed potential discriminative capability (AC1=0.645), it exhibited significantly low reliability (κ=0.296) and distributional heterogeneity (Stuart-Maxwell test, p<.001), driven by “Multimodal Hallucination” and “Confirmation Bias” based on visual patterns. Conversely, incorporating expert contextual information effectively rectified misjudgments and reconstructed the AI’s causal reasoning logic. Consequently, this study demonstrates that AI cannot replace human investigators due to limitations in “Sensory Extension” and “Legal Agency,” suggesting that AI must be established as an “Intelligent Collaborative Tool” operating under the control of human investigators.

목차

Abstract
1. 서론
1.1 연구의 배경 및 필요성
1.2 연구의 목적
2. 이론적 배경
2.1 Vision AI의 특성
2.2 국내 화재조사 분야의 AI 연구
2.3 화재 조사의 과학적 방법론과 맥락(Context)의 중요성
2.4 Vision AI 오류 분석을 위한 이론적 정의
3. 연구 방법
3.1 연구 설계
3.2 연구 대상
3.3 연구 도구
3.4 연구 절차(Experimental Procedure)
3.5 자료 분석 방법
4. 연구 결과
4.1 기술 통계
4.2. 단독 감식(Zero-shot) 연구
4.3 멀티모달 감식(Human-in-the-Loop) 연구
5. 결론 및 제언
5.1 연구 요약
5.2 화재조사관의 불가체성 논의
5.3 제언: 화재조사 보고서 품질 고도화 도구로서의 활용
참고문헌

키워드

저자

• 최복무 [ Bok Moo Choi | 고창소방서 ]
• 홍순문 [ Sun Mun Hong | 덕진소방서 ]

참고문헌

발행기관

간행물

표지보기 관심저널 등록

• 간행물명

  한국화재감식학회 학회지 [Magazine of Fire Investigation Socity of Korea]

• 간기

  계간

• pISSN

  2092-531X

• 수록기간

  2009~2026

• 등재여부

  KCI 등재

• 십진분류

  KDC 539 DDC 628