Coastal management is at a crossroads, facing intensifying challenges driven by, urbanization, and ecological degradation (EPA n.d.; The Guardian, 2025). Rising sea levels, extreme weather events, and coastal erosion increasingly threaten both human communities and natural ecosystems (EPA n.d.). Traditional gray infrastructure strategies, such as seawalls and levees, while essential, often struggle to keep pace with these multifaceted and evolving threats. Moreover, these hard engineering approaches can have unintended consequences, such as habitat degradation and disruption of natural sediment transport processes (Yale Environment Review n.d.). In response, there is a growing interest in integrating nature-based solutions (NbS) into coastal resilience strategies. Natural infrastructures like wetlands, dunes, and mangroves offer adaptive benefits, including storm surge mitigation, shoreline stabilization, biodiversity enhancement, and carbon sequestration (NOAA n.d.). Coastal wetlands in the United States alone are estimated to provide $23.2 billion in annual storm protection services (NOAA n.d.). However, the successful implementation of NbS requires interdisciplinary collaboration and sufficient site-specific data (Center for Coastal Solutions n.d.).

Artificial intelligence (AI) is emerging as a powerful tool in environmental management, offering new capabilities to process vast datasets, simulate ecological scenarios, and optimize conservation interventions (Dalton, 2024). AI has already demonstrated its value in applications ranging from biodiversity monitoring to climate modeling (Financial Times, 2025). Yet, the role of AI in shaping guidance and decisions in coastal management, particularly in relation to nature-based infrastructure remains an evolving and under-explored domain.

Many AI chatbots are now available, each with its own strengths and weaknesses. Examples include ChatGPT, DeepSeek, Claude, Meta AI, Google Gemini, Microsoft Copilot, Poe, Perplexity, Le Chat Mistral, HuggingChat, Grok, Pi, and Merlin. PC Magazine recently published a comparative review of these tools, asking each chatbot the same series of questions and evaluating their accuracy, consistency, complexity, and depth, along with core features such as table creation, citation support, and summarization (Forlini and Circelli, 2025). The authors emphasized that this type of assessment has clear limits: evaluating an AI model’s performance across the full range of possible topics would require an “army of experts,” and even then, performance may shift from one day to the next due to updates. Still, their testing revealed distinct differences, and they noted that users will likely form their own conclusions after spending significant time with each tool. Their rankings listed ChatGPT as Best Overall, Google Gemini as Best Value, Microsoft Copilot as Best for Windows Users, and Claude as Best for Privacy.

This manuscript investigates contemporary perspectives on coastal management as generated by three widely used generative AI chatbots. By analyzing their responses to a series of structured prompts, the study provides a comparative discussion of AI-generated insights and evaluates the use of this technology within the framework of coastal management projects. A systematic scoring metric is employed, assessing each response across five key dimensions: accuracy, completeness, relevance, clarity, and depth. The authors hope that the discussion of their findings will lead to reflection of coastal practitioners on how AI should and successfully can be utilized in their workflows.

The results of the comparative analysis show that while AI chatbots are capable of producing informative and well-structured content that can aid in broadening public understanding and supporting educational or preliminary planning efforts, their limitations remain evident. In particular, the absence of contextual awareness, nuanced judgment, and domain-specific expertise, as well as the constraint posed by the training data to rely on publicly available sources, excluding much of the peer-reviewed literature behind proprietary databases. These restrictions can lead to oversimplified, partially accurate, or outdated conclusions, particularly in dynamic, interdisciplinary fields (McGovern et al., 2021). Consequently, although generative AI shows promise as a supplementary tool for exploring sustainable, nature-based approaches to coastal management, critical decisions should continue to rely on the expertise and discernment of subject matter professionals.

Additionally, by critically analyzing how generative AI interprets and formulates guidance on coastal management, this research underscores the necessity of understanding the internal logic and data-driven heuristics that underpin AI-generated outputs. As AI becomes more integrated into environmental planning workflows, especially in areas such as climate adaptation and coastal resilience, it is imperative to interrogate how these systems process complex, interdisciplinary knowledge.

The central question posed is whether AI can meaningfully advance coastal management and nature-based infrastructure, or whether it merely reflects the biases embedded in its training data and the assumptions embedded in user prompts. AI’s current capabilities suggest it can synthesize information, identify general patterns, and provide accessible entry points into complex topics. However, its outputs are constrained by the scope, quality, and timeliness of the data it has been exposed to, as well as by its lack of contextual awareness and domain-specific judgment. While AI can simulate expertise in language, it does not possess an embodied understanding of ecological systems, or the situational insight required to navigate place-based environmental and social complexities. As these systems continue to evolve, their role in supporting scenario planning, stakeholder communication, and early-stage design processes may expand. Yet, realizing this potential requires a clear understanding of AI’s operational boundaries or what it can and cannot reliably do. Only by acknowledging these limitations can AI be responsibly integrated as a tool that augments, rather than replaces, the analytical rigor and critical thinking of domain experts in coastal science and engineering.

This study aims to assess the capabilities of generative artificial intelligence (AI) platforms in addressing topics relevant to coastal management, particularly with respect to natural infrastructure. A mixed qualitative-quantitative approach was used, employing a structured set of questions, standardized scoring criteria, and comparative evaluation across three leading free to use AI platforms. By evaluating AI’s capabilities against structured, interdisciplinary questions, this study underscores the potential of these tools to complement the expertise of engineers, ecologists, and coastal managers in advancing nature-based solutions for sustainable coastal management. The following subsections outline the steps taken in question development, AI selection, scoring, and reliability, along with a reflection on potential limitations inherent in AI-generated content.

A key observation emerged from the differences in responses generated for each reviewer: despite using the same prompt structures, model outputs were not identical. Variations reflected subtle shifts in phrasing, prompt history, and user-AI interaction dynamics, highlighting the stochastic nature of large language models and their sensitivity to input framing.

Across the session, responses evolved in tone and content, with later answers frequently reflecting terminology, conceptual frameworks, or biases introduced earlier. This was particularly evident in repeated emphasis on nature-based solutions, an area all models were more comfortable discussing, sometimes at the expense of balanced treatment of grey infrastructure or hybrid approaches. In several instances, prior responses appeared to prime subsequent answers, reinforcing specific narratives or limiting exploration of alternative strategies. For example, once natural infrastructure was discussed positively, later responses often defaulted to emphasizing these solutions without re-evaluating their trade-offs in context.

While this continuity occasionally enhanced coherence and readability, especially in models like Gemini that prioritized structured brevity, it also introduced redundancy and narrowed the scope of insights. Bing Copilot, although the strongest overall, exhibited this bias when referencing similar case studies or recommendations across multiple answers, sometimes offering less diversity in perspectives as the session progressed. ChatGPT, by contrast, tended to generalize its framing early and maintain that vagueness, while Gemini truncated nuanced arguments, possibly due to its lightweight optimization for speed.

Overall, this behavior underscores a structural limitation of generative AI in multi-turn analytical tasks: context retention improves flow but may constrain depth and diversity of content, particularly in domains like coastal management where interdisciplinary and contested perspectives are vital. Recognizing and mitigating this response drift is essential if these tools are to support complex decision-making processes without reinforcing narrow heuristics.

In a recent article in Nature (Naddaf, 2025), author’s found similar cases of ChatGPT and Gemini responding sycophantically, with a tendency for “people pleasing” and praising the questioner, to the detriment of depth, completeness and accuracy of the response. The findings of this study, support those claims and also conclude that AI platforms, such as Bing CoPilot and Gemini, can support the exploration of nature-based solutions by synthesizing complex coastal management data, terms and presenting clear, actionable insights. However, their role does not replace professional expertise, which should be used to guide critical decisions in the application of natural infrastructure.

The approach of using generative AI for this comparison highlights several important limitations of generative AI in this context. One critical issue is that these models can only generate responses based on the data they were trained on or have access to through publicly available information. Many high-quality scientific studies, technical reports, and regulatory documents remain behind paywalls or institutional access barriers, which prevents AI models from incorporating them into their responses. As noted by Himmelstein et al. (2018), approximately 28% of scholarly literature remains closed access, limiting broader public and algorithmic reach. This means that even the most accurate AI tools may not be able to reflect or reference the most recent or authoritative research unless it has been made open access or widely cited in public forums.

Furthermore, proprietary databases, such as those managed by Elsevier, Springer, and other commercial publishers, are typically excluded from the training corpora of AI models due to licensing restrictions. As a result, generative AI tools may produce answers based on secondary sources or older, less comprehensive data, which can be problematic for research fields like environmental science where timely, high-resolution data and current methodologies are essential (van Eck et al., 2013).

This comparative evaluation of ChatGPT (GPT-4), Google Gemini 1.5 Flash, and Microsoft Bing Copilot highlights both the promise and constraints of generative AI in the context of coastal management. While all three models demonstrated a baseline understanding of core topics, such as natural infrastructure, climate adaptation, and ecosystem services, their outputs varied significantly in accuracy, completeness, and depth. Notably, Bing Copilot outperformed its counterparts, often delivering responses with greater clarity and relevance, likely benefiting from real-time internet access (Microsoft, 2023a). Gemini followed closely, showing strength in concise structuring, while ChatGPT, despite its fluency, lagged in completeness and frequently produced generalized or vague content.

A shared strength across all platforms was the recognition of natural infrastructure as a vital component in coastal resilience. Concepts such as wetland preservation, dune restoration, and mangrove buffers were consistently cited for their roles in mitigating storm surge and erosion, echoing findings from the IPCC (2023) and research on ecosystem-based adaptation (Temmerman et al., 2013; Sutton-Grier et al., 2015). However, these insights often lacked contextual nuance. AI responses tended to reiterate well-established benefits of nature-based solutions (NBS) without critically engaging with trade-offs, feasibility under different environmental stressors, or the long-term performance of these systems under sea-level rise and socio-political uncertainty (Narayan et al., 2017; Spalding et al., 2014).

In terms of depth and contextualization, none of the models consistently demonstrated a robust understanding of the interdisciplinary nature of coastal planning. The complexity of balancing engineered infrastructure with ecological, social, and regulatory constraints was often oversimplified or omitted. Bing Copilot stood out by referencing recent research and providing greater topical relevance, though even its highest-scoring responses lacked deeper comparative evaluation between grey and green infrastructure strategies (Arkema et al., 2013). Gemini, optimized for efficiency, frequently truncated its answers, while ChatGPT’s outputs, although coherent, remained mostly generic and detached from current scientific discourse.

A critical limitation identified was the presence of contextual bias and response drift introduced by sequential prompting. Presenting all eight questions within a single session led the models to reuse terminology and frameworks from earlier interactions, sometimes reinforcing useful continuity but often narrowing the diversity of insights. For instance, once nature-based strategies were introduced positively, subsequent responses tended to default to these solutions even when a more balanced perspective would have been appropriate. This phenomenon aligns with observations in AI discourse, where context retention improves conversational flow but can constrain exploration of complex, multi-perspective topics (Bender et al., 2021; Birhane et al., 2022).

These findings underscore a central tension in using generative AI for applied environmental decision-making: while the models can articulate plausible and well-structured ideas, they lack situational awareness, ecological embodiment, and the ability to interpret socio-political dynamics, elements that are foundational in successful coastal management (Barbier et al., 2008; Foley et al., 2005). This absence creates a risk that confidently worded outputs may be misinterpreted as authoritative, leading to uncritical acceptance and marginalization of expert judgment.

To meaningfully integrate AI into coastal management workflows, it must be framed explicitly as a decision-support tool, not a substitute for expert analysis. AI can augment practitioner capacity during early stages, such as scoping literature, generating conceptual framings, or communicating ideas to stakeholders, but must be complemented by domain expertise to ensure scientific rigor and contextual relevance (Lal et al., 2022; Rolnick et al., 2022).

In Figure 5, a proposed theoretical framework outlines the cyclical and collaborative process for coastal management that would strategically integrate Generative AI chatbots with human expertise. The process begins within a broad Coastal Management Context, which considers ecological dynamics, socio-political conditions, and local knowledge. An Interdisciplinary Expert Team, comprising scientists, engineers, and policymakers, utilizes Generative AI (Chat Bots) to assist with tasks like literature synthesis and supplemental scenario ideation. However, the framework emphasizes that both AI outputs and human insights are filtered through a critical stage of Human Interpretation and Oversight. This ensures that all information is contextually validated and subjected to ethical and professional judgment before leading to Coastal Management Decisions and Outcomes. Finally, a Feedback and Learning loop is acknowledge where the results of these decisions are made public and used to update both the system’s assumptions and the AI algorithms, creating an adaptive and continuously improving management cycle. This model demonstrates the belief that AI has a place in coastal management as a supportive tool that can increase work efficiency and augment critical thinking and decision making by human experts.

In conclusion, AI chatbots hold promise as accessible, rapid-response tools for conceptual exploration, public education, and early-stage synthesis in coastal resilience planning. However, their effective use depends on critical interpretation, expert oversight, and a clear understanding of their operational boundaries. As these technologies evolve, future work should prioritize model training on interdisciplinary, site-specific datasets and enhance their capacity to reflect the complexity, trade-offs, and uncertainties inherent in coastal systems.