AI-assisted Protective Action: Study of ChatGPT as an Information Source for a Population Facing Climate Hazards (2304.06124v2)

Abstract: ChatGPT has been emerging as a novel information source, and it is likely that the public might seek information from ChatGPT while taking protective actions when facing climate hazards such as floods and hurricanes. The objective of this study is to evaluate the accuracy and completeness of responses generated by ChatGPT when individuals seek information about aspects of taking protective actions. The survey analysis results indicated that: (1) the emergency managers considered the responses provided by ChatGPT as accurate and complete to a great extent; (2) it was statistically verified in evaluations that the generated information was accurate, but lacked completeness, implying that the extent of information provided is accurate; and (3) information generated for prompts related to hazard insurance received the highest evaluation, whereas the information generated related to evacuation received the lowest. This last result implies that, for complex, context-specific protective actions (such as evacuation), the information was rated as less complete compared with other protective actions. Also, the results showed that the perception of respondents regarding the utility of AI- assistive technologies (such as ChatGPT) for emergency preparedness and response improved after taking the survey and evaluating the information generated by ChatGPT. The findings from this study provide empirical evaluation regarding the utility of AI-assistive technologies for improving public decision-making and protective actions in disasters.