Key points are not available for this paper at this time.
The emergence of cloud computing platforms has catalyzed the development of numerous generative artificial intelligence (AI) chatbots, offering users unprecedented access to information and reasoning capabilities at their fingertips, provided they have internet access. This manuscript explores AI's current perspectives on coastal management, focusing particularly on the implementation of natural infrastructure. By examining the insights provided by three widely used generative AI chatbots: ChatGPT (GPT-4), Google Gemini 1.5 Flash, and Microsoft Bing Copilot, the authors present a comparative analysis of AI-generated opinions on coastal management. The models were given identical prompts related to coastal management and natural infrastructure. The AI-generated responses were evaluated using a scored system assessing their accuracy, completeness, relevance, clarity, and depth. The analysis revealed significant variability in the performance of the three models. Bing Copilot, consistently delivered the most accurate and relevant responses. Gemini provided well-structured but often truncated answers, while ChatGPT's outputs were frequently generalized and lacked depth. A key finding was the presence of "response drift," where all models tended to reuse concepts from earlier in the conversation. Furthermore, the models' inability to access paywalled scientific literature and proprietary databases was identified as a critical limitation. The findings underscore the potential and limitations of these tools to inform and advance the field of natural infrastructure. The study highlights that the models lack the situational awareness and critical judgment necessary for complex environmental decision-making. The observed response drift and data access limitations can reinforce narrow perspectives and prevent a comprehensive evaluation of contested issues. A proposed framework suggests integrating AI within a cyclical and collaborative management process that prioritizes human interpretation and oversight to ensure contextually validated and robust outcomes.
Dillon et al. (Mon,) studied this question.