| Abstract |
| This exploratory study compared subjective evaluations of selected human-written disaster warning messages with versions restructured using generative Artificial Intelligence (AI). In total, 18 messages covering 6 hypothetical scenarios―earthquakes, tsunamis, wildfires, landslides, floods, and major industrial accidents―were evaluated. Each included a human-written original and versions restructured using ChatGPT-5 and Gemini Advanced. Data from 41/42 participants were analyzed. Participants rated perceived comprehension, accuracy, reading speed, reading ease, and actionability using a five-point Likert scale. Repeated measures ANOVA was followed by Bonferroni-adjusted comparisons and Pearson correlation analyses. AI-restructured messages generally received higher subjective ratings than selected human-written originals, although differences varied by disaster type and measure. Significant "main" effects of disaster and message type, as well as marked interactions, were observed. Variations between AI models were limited, except that Gemini received higher ratings for perceived accuracy and actionability regarding the major industrial accident scenario. All five measures were positively correlated. However, the absence of a comparable human-revision condition prevented the separation of the effects of AI use from those of message revision. The small sample, hypothetical scenarios, and self-reported measures also limit the generalization to objective performance or actual protective behavior. These findings provide exploratory evidence that generative AI may support disaster warning message restructuring and inform future controlled studies utilizing human-revised messages and objective measures. |
|
|
| Key Words |
| Disaster Warning Messages, Generative Artificial Intelligence, Risk Communication, Communication Effectiveness, Actionability, 재난문자, 생성형 인공지능, 위험 커뮤니케이션, 정보 전달 효과, 대처가능성 |
|
|
 |
|
