Prioritise risks and improve adaptation strategies in the Veneto coast through the application of a custom AI tool

Prioritise risks and improve adaptation strategies in the Veneto coast through the application of a custom AI tool Maria Katherina Dal Barco a,b , Veronica Casartelli b , Marcello San` o a,b,c , Sebastiano Vascon a,d , Silvia Torresan a,b,* , Andrea Critto a,b a Department of Environmental Sciences, Informatics and Statistics, Ca’ Foscari University of, Venice, Venice, Italy b CMCC Foundation - Euro-Mediterranean Center on Climate Change, Italy c Griffith University, Gold Coast, Australia d European Centre for Living Technology (ECLT), Ca’ Foscari University of Venice, Venice, Italy ABSTRACT Climate change has emerged as one of the most severe global challenges of our time, with rising temperatures and unprecedented shifts in climate patterns. Coastal areas are particularly vulnerable, facing compounded impacts from sea-level rise and increasingly frequent extreme weather events, demanding urgent need for proactive and comprehensive adaptation measures to protect coastal regions, recently defined as sentinels of climate change. A paradigm shift towards a multi-hazard risk perspective is increasingly recognised as essential in risk assessment and management. Moreover, Artificial Intelligence (AI) have emerged as promising tools to aid decision-making processes in coastal risk management and climate change adaptation. This study introduces COAST-AId, a custom Large Language Model designed to facilitate the analysis and synthesis of diverse information relevant for climate risk assessment and management along the Veneto coast. The tool facilitates the application of the risk assessment framework proposed in the European Climate Risk Assessment analysing the specific climate risk challenges of this region. The framework combines three key dimensions – i.e., risk identification, risk analysis, policy analysis – to prioritise risks and define urgent actions. The application of the COAST-AId tool was performed in close cooperation with local stakeholders involved in the MYRIAD-EU project where a systemic multi-hazard risk framework is considered to support the development of disaster risk management and climate adaptation pathways. The tool’s performance was evaluated by stakeholders, highlighting critical risks in the Veneto coastal as well as opportunities for enhancing coastal resilience and improving risk reduction and adaptation strategies at the regional to local scale. 1. Introduction Climate change impacts have become increasingly severe along coastal areas worldwide, where climate-related threats, such as rising sea-levels and changing patterns, frequency and intensity of extreme events (e.g., coastal storms, floods; [1,2] [3]), affect natural ecosystems and the communities ([1,2]). Addressing these challenges requires a shift from traditional single hazard assessment to a multi-hazard and multi-risk framework, analysing the complexity and spatio-temporal overlapping nature of multi-risk events [4–7]. This paradigm shift allows for a more comprehensive understanding of hazards, simplifying the development of targeted solutions to address multiple risks that coastal areas face. These interactions can intensify impacts, resulting in more severe outcomes, which requires holistic approaches that consider multi-hazard, multi-scale, and multi-sector perspectives [8–11]. Coastal local authorities and policymakers are increasingly faced with the challenging task of balancing coastal development with * Corresponding author. CMCC Foundation - Euro-Mediterranean Center on Climate Change, Italy. E-mail addresses: [email protected] (M.K. Dal Barco), [email protected] (V. Casartelli), [email protected] (M. San` o), [email protected] (S. Vascon), [email protected], [email protected] (S. Torresan), [email protected] (A. Critto). Contents lists available at ScienceDirect International Journal of Disaster Risk Reduction journal homepage: www.elsevier.com/locate/ijdrr https://doi.org/10.1016/j.ijdrr.2025.105818 Received 11 April 2025; Received in revised form 10 September 2025; Accepted 12 September 2025 International Journal of Disaster Risk Reduction 130 (2025) 105818 Available online 13 September 2025 2212-4209/© 2025 The Authors. Published by Elsevier Ltd. This is an open access article under the CC BY-NC-ND license ( http://creativecommons.org/licenses/by-nc-nd/4.0/ ). risk management. An integrated approach to assess multiple risks along coastal areas is needed to identify, on one hand, the main causes contributing to this rise of coastal risks and to identify, on the other hand, the most appropriate adaptation and risk management measures. According to the framework developed within the H2020 MYRIAD-EU project [12], this study uses a multi-risk assessment approach and follows the international and European risk assessment guidelines for climate change adaptation and disaster risk reduction [13–18] and the international ISO standard (ISO/IEC 31010:2019) to perform a scoping climate risk assessment at the regional to local scale. Risk assessment is used to facilitate a shared understanding of the risks affecting the Veneto region, as well as their relative significance (priority) in the context of both current (reference) and future scenarios. The identified, assessed and prioritized risks are then used as the basis for risk management planning and, therefore, the identification of risk reduction measures. More specifically, the research activities in this study involved the identification and prioritization of the main climate-related hazards and risks affecting the coastal areas of the Veneto region, and followed the main phases of the standard [19], i.e., problem formulation, risk identification, risk analysis, risk evaluation and risk treatment. Furthermore, advances in Machine Learning (ML) and Artificial Intelligence (AI) in general have opened new possibilities for multirisk assessment, exploiting vast data sources for environmental observation and monitoring (e.g., remote sensing data from Copernicus Sentinels, USGS Earth Explorer; [20–23]). In particular, custom Large Language Models (LLMs), such as those based on OpenAI’s Generative Pre-training Transformers (GPTs; [24]), have shown potential in analysing and synthesizing diverse data sources to create enhanced frameworks for multi-risk assessment and resilience building. These models operate as context-enhancing tools, leveraging extensive pre-training on vast datasets to provide nuanced insights across domains [25]. Unlike other methods such as retrieval-augmented generation (RAG) or fine-tuning, custom GPTs allow to use a combination of prompt engineering and instruction tuning to align outputs with specific user needs, incorporating domain-specific knowledge uploaded on a size-limited vector database, without requiring extensive retraining [26]. This setup allows these tools to adapt flexibly to varied applications, including multi-risk assessment, where they synthesize information from sources such as scientific studies, historical data, and real-time inputs [25,27]. This paper introduces COAST-AId, COastal multi-risk Assessment SupporTed by AI-baseD tools, a custom GPT-based tool designed to support multi-risk assessment and management in the Veneto coastal area, one of the climate change hotspots in the Mediterranean basin [28,29]. COAST-AId follows the approach proposed in the first European Climate Risk Assessment (EUCRA) report, which provides an evaluation of climate risks across Europe [16]. The report identifies 36 major climate risks that threaten European critical Fig. 1. The case study area: the coastal areas of the Veneto region (Italy). M.K. Dal Barco et al. International Journal of Disaster Risk Reduction 130 (2025) 105818 2 sectors, including food, ecosystems, infrastructure, economy and finance, and health. Many of these risks have already reached critical levels and could escalate to catastrophic consequences without urgent and decisive action. The EUCRA approach includes four consecutive steps, i.e. (i) risk identification, (ii) risk analysis, (iii) policy analysis, and (iv) urgency evaluation. In the risk identification, the major climate risks have been selected for the European regions through an initial screening of the relevant scientific literature, which followed the criteria adopted by the IPCC AR6 (e.g., high magnitude and likelihood of severe consequences, high irreversibility, potential cascading effects beyond system borders, worsening effect of climate change over time). The risk analysis step assesses the severity of each major climate risk (classified as limited, substantial, critical, or catastrophic) across three timeframes (current, mid-century, or late century). Furthermore, the confidence level is determined based on the type, quantity, quality, and consistency of evidence, as well as the degree of agreement among the sources. The policy analysis step evaluates temporal horizon, ownership and readiness of EU policies to examine the level of risk recognition and management. Finally, the urgency to act is determined by combining the findings from the previous steps into a matrix, hereinafter referred as the “EUCRA matrix”. The outcome considers five categories: watching in brief, sustain current action, further investigation, more action needed e urgent action needed. For more details on the classification of each level, refer to Supplementary Material SM1. COAST-AId is designed to support the application of the EUCRA approach to the regional (subnational) scale, rapidly delivering accessible and actionable information on key risks impacting the Veneto coastal area. This enables local policymakers and stakeholders to enhance their risk knowledge and foster adaptation strategies, paving the way for more effective and sustainable regional resilience planning. Therefore, following the EUCRA approach, COAST-AId supports three primary objectives: (i) identifying major risks for the Veneto coast – risk identification, (ii) analysing the severity of these risks – risk analysis, and (iii) evaluating current policies and adaptation strategies – policy analysis. Following the performance evaluation provided by local stakeholders and AI-experts, these outputs are incorporated into the EUCRA matrix applied to calculate the urgency to act for the identified major coastal risk within the Veneto region. This integration offers valuable insights to refine local policies and supports the development of a resilient coastal management plan aligned with the 2024 Regional Strategy for Climate Change Adaptation in the Veneto region (Regione del Veneto, 2024). This paper explores the application of COAST-AId, carried out in close collaboration with local stakeholders. The tool represents a preliminary attempt to apply the EUCRA approach at a regional-to-local scale, with a specific focus on the Veneto coast (Section 3). Key outputs, including the stakeholders’ evaluation of COAST-AId’s performance, are discussed in Section 4. 2. Case study area The coastal area of the Veneto region (Fig. 1), stretching approximately 169 km along the North Adriatic Sea, presents a diverse and complex landscape shaped by both natural and anthropogenic influences [30]. The coastaline belongs to the provinces of Venice and Rovigo, and is characterised by low-lying sandy shores across the entire littoral, which is predominantly shaped by the presence of six river outlets (i.e., Adige, Piave, Brenta, Tagliamento, Livenza, Sile) and the delta system of the Po river [31,32]. Due to wave and tidal energy, these shores feature prominent spits and barrier islands belonging to the city of Venice (i.e., Lido and Pellestrina), enclosing morphologically distinct lagoons connected to the sea through tidal inlets. The evolution of these beaches relies on a delicate balance between the transgressive marine regime, driven by rising sea levels, and the diminishing fluvial sediment supply. These natural sites are deeply blended with the urban area, which have seen substantial expansion since the post-war economic boom and the subsequent rise of mass tourism throughout the region, which have significantly altered the natural sediment balance, reducing the resilience to coastal hazards [33]. This unique geomorphology renders the region particularly susceptible to extreme events as a result of a multitude of hazards [9, 33–37]. In particular, rising sea-levels, intensified storm surges, and increased frequency of extreme weather events have amplified coastal related risks, impacting water quality, biodiversity, and the livelihoods of local communities [21,36,38]. Recent notable extreme events include the exceptional high tide of November 2019 and the compound hazard events that led to the Vaia storm (also known as Storm Adrian) between October and November 2018 [36]. The November 2019 high tide was driven by a combination of atmospheric and oceanographic factors, including strong Sirocco winds across the Adriatic basin, a peak in the astronomical tide, an unusually high average sea-level in the Adriatic and the passage of a small cyclone over the Northern Adriatic and the Venice lagoon [36]. The Vaia storm, a powerful Mediterranean storm with hurricane-force wind gusts and heavy rainfall, predominantly impacted the mountainous areas of the Veneto region, causing extensive forest destruction [39]. It also triggered an unusually prolonged high tide in Venice, leading to storm surge events along the coastline and sediment overflow at the Veneto river mouths [40]. In response to these unprecedented multi-hazard and multi-risk scenarios, various strategies have been implemented to safeguard coastal zones and enhance urban resilience. The most implemented measures, aimed at managing coastal-related risks, include both ‘grey infrastructure’, such as groins, breakwaters, and tidal gates, as well as ‘nature-based solutions’ like beach nourishment, dune restoration and conservation of natural areas (Regione del Veneto, 2024). Recently, the Environment and Ecological Transition Unit of the Veneto Region drafted the prelimary document of the Regional Strategy for Climate Change Adaptation (SRACC; Regione del Veneto, 2024). The SRACC outlines an adaptive framework tailored to the region’s diverse geographic and environmental features, M.K. Dal Barco et al. International Journal of Disaster Risk Reduction 130 (2025) 105818 3 encompassing alpine, riverine, lacustrine and coastal areas. The strategy has been conceived as a strategic document built upon the findings of five technical-scientific reports, 1 which provide a comprehensive foundation on climatic conditions, existing adaptation plans, and multi-hazard analyses related to climate change. Notably, indices have been developed to classify relative risk levels across multiple climate-related hazards in the Veneto Region (Regione del Veneto, 2024). Among these risks, coastal flooding is the only coastal risk explicitly addressed in the current strategy, with its impacts projected to intensify significantly due to climate change, leading to an increased number of affected receptors (Regione del Veneto, 2024). However, the land-sea interface in the Veneto region is exposed to a broader spectrum of critical coastal risks that remains unaddressed [21,41]. This study aims to fill this gap by conducting an analysis of coastal risks in the Veneto region. Specifically, it aims to identify and prioritise the major coastal risks and to offer robust evaluation guidelines that could inform regional adaptation and risk management planning processes, making them effectively respond to the complex and evolving challenges posed by climate change. 3. Methodology This study aims to support the prioritization of coastal risks and enhance adaptation and risk management strategies for coastal areas in the Veneto region, addressing climate change challenges through the application of a customized tool based on a pre-trained large language model (LLM), GPT4o [42]. The proposed methodology followed four steps, as depicted in Fig. 2. Fist, COAST-AId’s instructions and knowledge were designed and prepared (Section 3.1). Then, the tool was applied to the coastal areas of the Veneto region following the European Climate Risk Assessment (EUCRA) approach as a reference framework and methodological guide. This application was aimed to identify the region’s most significant coastal risks, assess their severity, and evaluate current policy readiness to address these challenges (Section 3.2). To ensure the reliability of the results, AI experts and local stakeholders were actively engaged in evaluating COAST-AId’s performance and outputs (Section 3.3). This validation was a crucial step towards achieving the overarching goal of determining the urgency of actions required to reduce the identified coastal risks (Section 3.4). The following sections provide a detailed description of each phase. 3.1. COAST-AId development The assessment of multiple risks within the complex system of the Veneto region’s coastal area required the integration and analysis of a wide range of data sources in PDF format. These include reports, peer-reviewed papers, and deliverables from EUand nationally funded projects focusing on the investigated area. To address this challenge and extract the most relevant information for coastal risk assessment, a custom large language model (LLM), called COAST-AId (COastal multi-risk Assessment SupporTed by AI-baseD tools, https://chatgpt.com/g/g-12xYfUHJh-coast-aid), was developed. COAST-AId was preferred over off-the-shelf models as it ensures higher contextual relevance, domain-specific accuracy, and stronger alignment with the project’s objectives. Preliminary tests with general-purpose models (e.g., ChatGPT-3.5) revealed limitations in maintaining factual consistency and adequately capturing the complexity of coastal multi-risk dynamics. These findings underscored the need for a structured, prompt-engineered approach capable of reducing factual drift and enhancing domain alignment. To achieve this, COAST-AId was developed as a custom LLM trained exclusively on climate-related sources, including scientific literature, technical reports, and policy documents. By operating within a closed, vetted dataset, the tool minimizes hallucinations and factual inaccuracies, while preserving transparency over data provenance. Moreover, the model allows the integration of user-defined instructions to prioritise specific objectives (e.g., identifying key risks for a case study), thereby ensuring outputs remain tailored to real-world decision-making needs. These features are essential for supporting multi-risk climate assessments, which demand flexibility, contextual depth, and methodological rigor. COAST-AId was built entirely within OpenAI’s custom GPT (Generative Pre-trained Transformer) environment, using the GPT-4o architecture. No external infrastructure, fine-tuning, or local implementation of retrieval-augmented generation (RAG) was applied. Instead, the system leverages OpenAI’s GPT-4o ′ s native capabilities as a state-of-the-art multimodal transformer, capable of processing long-context text and structured documents with high reasoning efficiency [43,44]. All customization of COAST-AId was achieved through prompt engineering, rather than altering model weights or applying supervised fine-tuning. Custom behavior was implemented using system-level instructions and curated few-shot examples. This approach enabled domain-aligned reasoning and task-specific formatting, drawing on well-established prompting strategies such as chain-ofthought, decomposition, and few-shot prompting [26]. Knowledge integration was conducted through OpenAI’s native file upload functionality within the Custom GPT interface. Input documents, including scientific publications, policy reports, and stakeholder contributions, were uploaded directly to the platform. Upon upload, OpenAI’s backend automatically performed semantic chunking, embedding generation, and indexing into an internal vector store. At inference time, user prompts are transformed into embeddings and semantically matched against the uploaded knowledge base using OpenAI’s built-in vector search. The most relevant content chunks are then dynamically injected into the model’s output pipeline [45]. 1 Technical-Scientific Reports produced to support the Veneto Region Strategy for Climate Change Adaptation (SRACC): https://www.regione. veneto.it/web/ambiente-e-territorio/rapporti-tecnico-scientifici. M.K. Dal Barco et al. International Journal of Disaster Risk Reduction 130 (2025) 105818 4 All input documents were uploaded during the development and configuration of the tool, which means that COAST-AId’s knowledge is strictly limited to those specific sources. It does not access external web-based inputs. A complete list of input documents is provided in Supplementary Material SM2. Currently, the prototype tool is set to private mode, due to the inclusion of various non-open-source materials, such as subscriptionbased scientific papers and regional reports shared under restricted agreements. To promote transparency and ensure replicability, web search functionality was disabled, allowing the authors full control over all input data used during model operation. In the initial stages of the development, COAST-AId was tested iteratively using a subset of the 60 documents from the final dataset. This approach supported the refinement of both the tool’s linguistic behaviour and its understanding of domain-specific terminology. Starting with a smaller, curated selection of documents, allowed identification of comprehension gaps and informed improvements in Fig. 2. Workflow adopted to prioritise risks and improve current adaptation strategies in the Veneto coast. Table 1 Final set of prompts used for applying the EUCRA approach to the coastal areas of the Veneto region. EUCRA level Prompts Technique Risk identification Previous assessments have identified sea-level rise as a significant climatic hazard and urban development as a non-climatic hazard responsible of the occurrence of coastal erosion risk in the Veneto coastal area. Considering this example, list the most important climatic (atmospheric, marine, and terrestrial) and non-climatic hazards responsible for each coastal risks in the Veneto coastal area. Few-Shot Prompting [46] First, identify the most relevant coastal risks occurring along the Veneto coastal area. Next, for each risk, list the combinations of climatic and non-climatic hazards previously identified as the most important hazards responsible for these risks. Explain how these hazard combinations contribute to each specific risk in the coastal areas of the Veneto region. Decomposition [49] Risk analysis First, explain the rationale behind the management measures implemented for each coastal risk in the Veneto coastal area. Then, assess how these measures have either reduced or increased the impacts of the other risks, considering only the risks identified so far for the Veneto coastal area. Provide examples and explain why these effects occurred. Step-Back Prompting [50] As a data analyst, review each attached file and count the number of documents focused on each coastal risk identified so far for the Veneto coastal area. Provide a step-by-step summary of your counting process for each file. Role Prompting [48] For each coastal risk identified so far for the Veneto coastal area, provide: •the climatic and non-climatic hazards responsible for the risk. •the level of severity (Catastrophic, Critical, Substantial, Limited – based on the European Climate Risk Assessment classification) for the current situation (until 2040), mid-century (2040–2060), and end of the century (until 2100). •for each time frame, explain in detail the reasons for the assigned severity level, reasoning step-by-step. •the measures implemented to address the risk. •the sectors affected by this risk. Zero-Shot Chain-of-Thought Prompting [51] Policy analysis Using the provided definition of ’policy horizon’, determine the policy horizon along the Veneto coastal area related to each coastal risk identified so far for the Veneto coastal area as a consequence of climatic and nonclimatic hazards. Explain your reasoning step-by-step, considering the severity of the risk over time. Chain-of-Thought Prompting [47] As a policy analyst, identify the policies issued by at the sub-national/local scale of the Veneto region that cover each coastal risk identified so far for the Veneto coastal area. Provide brief descriptions of each policy and explain how they address each Veneto coastal risk. Role Prompting [48] Using the provided categories of ’policy readiness’, evaluate the norms and policies related to each coastal risk identified so far for the coastal area of the Veneto region. Explain your reasoning step-by-step to determine whether the policy readiness is Low, Medium, Advanced, or Very Advanced. Chain-of-Thought Prompting [47] Urgency to act First, summarize the information related to ‘risk severity’, ‘confidence’, ’policy horizon’, and ’policy readiness’ for each coastal risk identified so far for the Veneto coastal area. Next, using the European Climate Risk Assessment classification, identify the appropriate class of ’urgency to act’ (i.e., ’urgent act needed’, ’more action needed’, ’further investigation’, ’sustain current action’, ’watching brief’). Explain your reasoning stepby-step. Decomposition [49] M.K. Dal Barco et al. International Journal of Disaster Risk Reduction 130 (2025) 105818 5 output quality. This phase concluded with the integration of advanced prompt design techniques, balancing technical expertise and creative intuition to enhance user interaction with the model. The prompting strategies adopted were guided by the framework proposed by Schulhoff et al. [26], and include. •Few-Shot Prompting, providing the model with a few examples to learn from, enabling the generation of contextually relevant and informed responses [46]. •Chain-of-Thought Prompting, encouraging the model to reason step-by-step through a problem, promoting logical reasoning and ensuring structured and accurate responses [47]. •Role Prompting, instructing the model to adopt the role of an expert (e.g., jurist, data analyst), ensuring methodological task execution within the context of that role [48]. •Decomposition, breaking down a complex task into smaller, manageable steps, allowing the model to address each part accurately [49]. •Step-Back Prompting, where the model reflects on broader implications, promoting deeper analysis and thoughtful conclusions [50]. •Zero-Shot Chain-of-Thought Prompting, directing the model to provide step-by-step reasoning, guiding it to produce detailed and well-reasoned responses without prior examples [51]. These prompting techniques were pivotal in fine-tuning the interaction between COAST-AId and the user, culminating in a refined set of prompts specifically designed to apply the EUCRA approach to the Veneto region’s coast, as further detailed in Section 4.1 and Supplementary Material SM4. 3.2. Application of COAST-AId to the veneto region COAST-AId was created aiming to facilitate the application of the EUCRA approach to the local scale, making the complex data landscape of the Veneto coastal area more accessible to policymakers and stakeholders. The European approach includes four consecutive steps: risk identification, risk analysis, policy analysis, and urgency evaluation, which will be introduced in the following sections. 3.2.1. Risk identification To support the identification of major climate risks (first level of the EUCRA approach, see Section 1) affecting the coastal area of the Veneto region, the input documents (detailed in Supplementary Material SM2) were analysed and synthesised using the COASTAId tool. The prompts specifically designed for this step aimed to generate a comprehensive understanding of the Veneto coastal system, encompassing both climatic and non-climatic drivers, as well as their interactions leading to the identified coastal risk. The final output of this step is a conceptual framework that provides a comprehensive overview of the Veneto coastal system, which will be presented in Section 4.2, and laying the foundation for the subsequent risk analysis, presented in Section 3.2.2. 3.2.2. Risk analysis As outlined in Section 1, the second level of the EUCRA approach involves risk analysis, focusing on the assessment of both the severity of each identified major coastal risk and the corresponding confidence level. The classification for risk severity follows the EUCRA framework (i.e., limited, substantial, critical, catastrophic) across three distinct timeframes (i.e., current, mid-century, late century), as detailed in Supplementary Material SM1. To assess the EUCRA-defined confidence level, the volume of information available for each major risk derived directly from the input documents of the COAST-AId tool was used as a proxy. This adjustment ensured that the assessment aligns with the data availability and context-specific characteristics of the Veneto region while maintaining methodological consistency with the overall risk analysis framework. In particular, the greater the quantity and relevance of the information, the higher the confidence level is assigned to that coastal risk, ensuring a tailored evaluation that accurately reflects the data’s robustness. In particular, the confidence levels are categorised as follows: low if a maximum of five documents are identified, medium for six to ten documents, high for more than ten documents concerning the designated risk for the Veneto coast. 3.2.3. Policy analysis Following the EUCRA approach outlined Section 1, the third level encompasses the analysis of current policies, including evaluations of risk ownership, policy horizon and policy readiness. As at the European level, where the 2021 EU adaptation strategy provides an overarching framework for adaptation policies, but implementation is mainly through sectoral policies [16], the Strategy for Climate Change Adaptation of the Veneto Region (i.e., Strategia Regionale di Adattamento ai Cambiamenti Climatici per il Veneto – SRACC; [52]) sets strategic objectives to be implemented and/or being implemented within sectoral planning processes. The SRACC itself provides a comprehensive list and preliminary analysis of existing plans that include a number of adaptation and risk reduction measures. For the purpose of this study, which focuses on the regional level and aims at supporting policy makers in Veneto, only policy and plans associated with regional ownership were processed by COAST-AId. For the remining policy indicators (i.e., temporal horizon and policy readiness) the classification aligns with the EUCRA framework, reported in Supplementary Material SM1. M.K. Dal Barco et al. International Journal of Disaster Risk Reduction 130 (2025) 105818 6 3.3. Engaging experts and local stakeholder to evaluate COAST-AId performance The deployment of LLMs in real life settings is still in its infancy and it has been limited by challenges related to their reliability, the generation of hallucinations, and replicability in different contexts. To address these challenges, several frameworks have been developed to systematically assess LLM outputs, most of which rely on human evaluation [53]. In this study, experts and local stakeholders were engaged during the MYRIAD-EU project workshop in which the COAST-AId tool and its outputs were presented. The tool was developed as a web-based, chat-style interface within the OpenAI Custom GPT environment. However, stakeholders did not interact directly with the tool. Instead, they were exposed to a curated selection of responses generated by COAST-AId and presented by the research team during the project’s workshop. Following the presentation, participants were requested to evaluate the outputs and articulate their levels of agreement or disagreement through structured questionnaires and interviews. The evaluation was conducted based on recent frameworks commonly used for LLM benchmarking [54–56], which identify five key criteria: accuracy, relevance, clarity, completeness and conciseness. Accuracy assesses whether the LLM’s response is factually correct, aligned with verifiable data, or grounded in trusted external sources. In critical applications such as environmental risk management or financial services, incorrect information can lead to severe consequences. Relevance measures how well the response addresses the specific prompt. Outputs should be focused and aligned with the topic at hand [53]. For instance, when evaluating risks in the Veneto coastal area, responses must provide information pertinent to the location and associated hazards. Clarity refers to the ease with which the response can be understood by the user. A well-structured, easy-to-follow response ensures that even complex information is digestible. This is particularly important for decision-makers who need quick, actionable insights. Completeness ensures that all components of the prompt are addressed in the response. For multi-part questions or prompts requiring detailed lists (e.g., hazard identification, risk ownership), it is crucial that no critical information is omitted. Conciseness evaluates whether the LLM delivers the required information efficiently, without unnecessary verbosity. Responses should be brief yet comprehensive, avoiding excessive elaboration that may obscure important points. This indirect evaluation approach allowed for the identification of content gaps and inconsistencies, which were addressed through manual updates to the prompt structure and expansion of the input dataset. For instance, a key information related to saltwater intrusion was found absent, prompting the addition of further technical documentation to improve the model’s coverage and accuracy. The evaluation approach adopted differs from standard model validation practices, as the primary objective was to develop and qualitatively assess a prototype tool rather than to provide a cross-validation framework for numerical models. This methodological choice reflects the exploratory and interdisciplinary nature of COAST-AId, while future developments could include sensitivity analyses to assess the tool’s transferability to other regional contexts and the upscaling potential at EU-level. The insights gained from these evaluations were synthesised into the final COAST-AId results, contributing comprehensive information on risk identification, risk analysis, and policy analysis for the Veneto coastal area, which will be detailed in Section 4.3. 3.4. Veneto coastal multi-risk assessment and management The last step of the EUCRA approach involves determine the urgency and type of policy action [16]. When focusing on the coastal area of the Veneto region, the urgency to act was determined in two different ways. First, COAST-AId Fig. 3. Conceptual framework for coastal risk assessment in the Veneto region (Italy). M.K. Dal Barco et al. International Journal of Disaster Risk Reduction 130 (2025) 105818 7 outputs from the previous steps were integrated into the EUCRA matrix. Then, COAST-AId was employed one final time to independently calculate the urgency to act. Both results - those calculated by the EUCRA matrix and those determined by COAST-AId - were categorised into five levels of urgency (i.e., watching in brief, sustain current action, further investigation, more action needed, and urgent action needed). More details on these categories are provided in Supplementary Material SM1. This approach enabled a comparative analysis of the results produced by the EUCRA matrix and COAST-AId, presented in Section 4.4. 4. Results and discussion 4.1. COAST-AId development The development of COAST-AId was a comprehensive and iterative process, culminating in the identification and refinement of the final set of prompts, reported in Table 1. The initial application of the tool revealed challenges related to hallucinations, which were effectively mitigated through structured prompting techniques (outlined in Section 3.1) and multiple rounds of refinement. These improvements significantly enhanced the reliability and accuracy of responses, ensuring that the tool’s outputs aligned more closely with the input data and were more applicable for decision-making processes. A crucial factor in strengthening COAST-AId’s reliability was its co-development with local stakeholders, whose in-depth knowledge of territorial characteristics, challenges, and needs proved invaluable. Their expertise, as discussed in Section 4.3, played a pivotal role in refining the tool’s functionality, ensuring that the generated insights were both contextually relevant and scientifically sound. Given that LLMs are still in the early stages of their application, adopting structured prompting techniques and actively involving domain experts in the validation process remains essential. While these tools can streamline and expedite desk reviews, their inherent limitations must be acknowledged. Therefore, the iterative refinement of AI-assisted tools, such as COAST-AId, combined with expert engagement, ensures they serve as valuable instruments for analysing complex, multi-dimensional risks while maintaining scientific rigor. 4.2. Application of COAST-AId to the veneto region The proposed study aims to apply the EUCRA approach to the local context of the Veneto coast, harnessing the capabilities of the COAST-AId tool. The following sections provide a comprehensive breakdown of the outcomes from each step of the EUCRA approach (detailed in Section 3.2), with the corresponding COAST-AId outputs reported in Supplementary Material SM3. In particular, Section 4.2.1 covers risk identification, risk analysis is discussed in Section 4.2.2, and finally policy analysis is presented in Section 4.2.3. 4.2.1. Risk identification COAST-AId was employed to identify the major climate risks impacting the coastal area of the Veneto region, through a set of specific risk related prompts, encouraging the retrieval of relevant information from the documents available in the tool. This application provided a comprehensive understanding of the Veneto coastal system, culminating in the development of the conceptual framework presented in Fig. 3. The conceptual framework illustrates the interplay between climatic-related hazards and non-climatic risk drivers, which together shape the major climate-related risks identified in the case study. Coastal erosion is driven by an intricate combination of climate-related hazards, including storm surge, sea-level rise, and pluvial flooding [23]. These hazards, exacerbated by the increasing frequency of extreme weather events observed over recent decades [9] and projected to intensify in the future [57], collectively reshape shorelines, redistribute sediments, and destabilize coastal ecosystems [21]. When these hazards interact, their impacts often reveal the compounding nature of extreme events [9]. For instance, sea-level rise amplifies the effects of storm surges and wave impacts, whereas the combination of pluvial flooding and storm surge-induced coastal inundation accelerates sediment displacement and erosion [21,58]. Non-climatic risk drivers, such as sediment changes, shipping traffic and subsidence, further exacerbate this unstable situation by disrupting the physical processes that regulate sediment dynamics and shoreline stability. Although not directly linked to climate change, these factors (such as land subsidence and anthropogenic pressures) significantly increase the susceptibility of coastal system Table 2 COAST-AId outputs related to risk analysis, which includes risk severity across three different temporal timeframes (current, mid-term, long-term) and the confidence level for the Veneto coastal area. Risks Risk severity Confidence level Current Mid-term Long-term Coastal erosion Substantial Critical Catastrophic þþ Permanent inundation Limited Substantial Critical þ Coastal flooding Substantial Critical Catastrophic þþþ Saltwater intrusion Limited Substantial Critical þ M.K. Dal Barco et al. International Journal of Disaster Risk Reduction 130 (2025) 105818 8 to climatic hazards like wave action and storm events [59,60]. The combined effects of subsidence and sea-level rise are particularly concerning, as they increase relative sea-level changes and heighten the risk of coastal flooding and permanent inundation [61,62]. This dynamic is especially pronounced in the Veneto region, characterised by low-lying topography prone to flooding [58,63,64], dense coastal development that amplifies exposure [59,60], and reliance on artificial water management systems (e.g., MOSE system), which face escalating challenges from these rising hazards [65,66]. Saltwater intrusion arises from a feedback loop involving subsidence, rising sea-levels and temperatures, drought, and heatwaves, all of which amplify coastal risks [41,67]. Subsidence reduces land elevation, increasing vulnerability to saltwater intrusion during droughts and heat waves. Concurrently, rising temperatures and heat waves diminish freshwater availability, weakening natural hydrological defences [41,68,69]. Human activities, such as excessive groundwater extraction for irrigation and urban development, further exacerbate these risks by intensifying subsidence and depleting freshwater reserves [70,71] 4.2.2. Risk analysis Following the identification of the major risks affecting the coastal area of the Veneto region (i.e., coastal erosion, permanent flooding, coastal flooding, and saltwater intrusion), the COAST-AId tool was employed to conduct a detailed risk analysis through a set of specific prompts to retrieve the relevant information from the documents available in the tool. This analysis assessed both the severity of each coastal risk, and the confidence level associated with the supporting documentation. The resulting matrix (Table 2) offers a clear visual representation of the methodology detailed in Section 3.2.2. It enables the understanding and effective prioritization of the intensifying risks affecting the Veneto coast, while capturing the temporal evolution of these risks across current, midterm, and long-term scenarios. Coastal erosion, driven by rising sea levels, storm surges, and reduced by beach nourishment and dune restoration efforts [21,23], currently poses a Substantial risk. However, mid-term classification indicates an escalation to Critical, fuelled by accelerating sea-level rise and increasing storm intensity [22]. In the long term, unchecked erosion could lead to Catastrophic consequences for coastal infrastructure, ecosystems, and local communities. These findings align with regional studies, which emphasise the compounded impacts of geomorphological vulnerabilities and anthropogenic pressures, such as urbanization and tourism development [72–74]. A similar pattern is observed for coastal flooding, exacerbated by the Veneto coast’s peculiar geomorphology, characterized by lowlying littoral zones [31,32]. The situation is further strained by the existing flood defence infrastructures (e.g., MOSE system) which may require substantial upgrades to address the growing magnitude and frequency of extreme weather events [36,66,75]. The interplay of sea-level rise and land subsidence are making the Veneto coast increasingly vulnerable to permanent inundation [61,62]. While currently classified as a Limited risk, this coastal risk is projected to become Substantial in the mid-term and Critical in the long-term, particularly under high-emission scenarios such as RCP8.5 [22,61]. Similarly, saltwater intrusion, currently set as a Limited risk, is projected to intensify as rising sea levels diminish freshwater flow [70,71]. Mid-term and long-term scenarios elevate this risk to Substantial and Critical, respectively, posing significant threats to agriculture and freshwater availability, underscoring the need for proactive water management strategies to limit this risk [76]. Finally, as detailed in Section 3.2.2, the confidence level was directly derived from the input documents of the COAST-AId tool, serving as a reliable proxy of the EUCRA-defined confidence level. Specifically, COAST-AId quantified the number of documents associated with each risk identified in Section 4.2.1. The analysis revealed that coastal flooding had the highest number of related documents (16), followed by coastal erosion (10) and permanent inundation (5). Saltwater intrusion had the lowest level of documentation (4), reflecting a relatively recent – but growing – concern within the scientific community. 4.2.3. Policy analysis Major coastal risks identified for the Veneto region include coastal erosion, coastal flooding, permanent inundation, and saltwater intrusion, driven by both climatic and non-climatic drivers. To conduct a preliminary assessment of existing regional policies, strategies, and plans, the COAST-AId tool was employed through a set of specific prompts to retrieve relevant information with the aim of applying the EUCRA approach [16] to the Veneto coast. The policy analysis followed the EUCRA methodology, thus assessing risk ownership, policy horizon, and policy readiness, as illustrated in Table 3. The analysis considered all sectoral plans included in the SRACC, through which the Veneto Region intends to achieve strategic objectives to foster transformative adaptation. The assessment of policy readiness reveals varying levels of maturity across the different risks. Coastal erosion and saltwater intrusion policies are categorised as having a medium level of readiness, reflecting ongoing efforts to develop both grey and naturebased solutions, such as dune restoration and sediment management [72,77]. Similarly, readiness for permanent inundation is also rated as medium, with efforts focused on enhancing spatial planning, restoring floodplains, and strengthening critical infrastructure [66]. In contrast, policies addressing coastal flooding demonstrate an advanced level of readiness, bolstered by the MOSE system, Table 3 COAST-AId outputs for policy analysis, including risk ownership, policy horizon and policy readiness for the Veneto coastal area. Risks Risk ownership Policy horizon Policy readiness Coastal erosion Veneto Midto Long-term Medium Permanent inundation Veneto Long-term Medium Coastal flooding Veneto Long-term Advanced Saltwater intrusion Veneto Midto Long-term Medium M.K. Dal Barco et al. International Journal of Disaster Risk Reduction 130 (2025) 105818 9 [63] F. Antonioli, M. Anzidei, A. Amorosi, V. Lo Presti, G. Mastronuzzi, G. Deiana, G. De Falco, A. Fontana, G. Fontolan, S. Lisco, A. Marsico, M. Moretti, P.E. Orrù, G. M. Sannino, E. Serpelloni, A. Vecchio, Sea-level rise and potential drowning of the Italian coastal plains: flooding risk scenarios for 2100, Quat. Sci. Rev. 158 (2017) 29–43, https://doi.org/10.1016/j.quascirev.2016.12.021. [64] A.R. Scorzini, E. Frank, Flood damage curves: new insights from the 2010 flood in Veneto, Italy, Journal of Flood Risk Management 10 (2015) 381–392, https:// doi.org/10.1111/jfr3.12163. [65] L. Cavaleri, M. Bajo, F. Barbariol, M. Bastianini, A. Benetazzo, L. Bertotti, J. Chiggiato, C. Ferrarin, F. Trincardi, G. Umgiesser, The 2019 flooding of venice and its implications for future predictions, Oceanography (Wash. D. C.) 33 (1) (2020) 42–49. https://www.jstor.org/stable/26897834. [66] D. Faranda, M. Ginesta, T. Alberti, E. Coppola, M. Anzidei, Attributing venice acqua alta events to a changing climate and evaluating the efficacy of MoSE adaptation strategy, Npj Climate and Atmospheric Science 6 (1) (2023) 181, https://doi.org/10.1038/s41612-023-00513-0. [67] J. Luo, E. Straffelini, M. Bozzolan, Z. Zheng, P. Tarolli, Saltwater intrusion in the Po River Delta (Italy) during drought conditions: analyzing its spatio-temporal evolution and potential impact on agriculture, Int. Soil Water Conserv. Res. 12 (3) (2024) 714–725, https://doi.org/10.1016/j.iswcr.2023.09.009. [68] D. Bellafiore, C. Ferrarin, F. Maicu, G. Manf` e, G. Lorenzetti, G. Umgiesser, L. Zaggia, A.V. Levinson, Saltwater intrusion in a mediterranean Delta under a changing climate, J. Geophys. Res.: Oceans 126 (2) (2021), https://doi.org/10.1029/2020JC016437. [69] D. Bonaldo, D. Bellafiore, C. Ferrarin, R. Ferretti, A. Ricchi, L. Sangelantoni, M.L. Vitelletti, The summer 2022 drought: a taste of future climate for the Po valley (Italy)? Reg. Environ. Change 23 (1) (2023) 1, https://doi.org/10.1007/s10113-022-02004-z. [70] G. Pietro Beretta, J. Terrenghi, Groundwater flow in the venice lagoon and remediation of the Porto Marghera industrial area (Italy), Hydrogeol. J. 25 (3) (2017) 847–861, https://doi.org/10.1007/s10040-016-1517-5. [71] P. Tarolli, J. Luo, E. Straffelini, Y.-A. Liou, K.-A. Nguyen, R. Laurenti, R. Masin, D’Agostino Vicenzo, Saltwater intrusion and climate change impact on coastal agriculture, PLOS Water 2 (4) (2023) E0000121, https://doi.org/10.1371/journal.pwat.0000121. [72] G. Fontolan, A. Bezzi, D. Martinucci, S. Pillon, C. Popesso, GCV - Geodatabase Gestionale Delle Coste Venete, 2014. [73] A. Marsico, S. Lisco, V. Lo Presti, F. Antonioli, A. Amorosi, M. Anzidei, G. Deiana, G. De Falco, A. Fontana, G. Fontolan, M. Moretti, P.E. Orrú, E. Serpelloni, G. Sannino, A. Vecchio, A. Marsico, S. Lisco, V. Lo Presti, F. Antonioli, E. Serpelloni, Flooding scenario for four Italian coastal plains using three relative sea level rise models, 5647, https://doi.org/10.1080/17445647.2017.1415989, 2017. [74] J. Van Der Borg, Tourism and urban development: the case of venice, Italy, Tour. Recreat. Res. 17 (2) (1992) 46–56, https://doi.org/10.1080/ 02508281.1992.11014649. [75] M. Anzidei, T. Alberti, A. Vecchio, X. Loizidou, D. Orthodoxou, E. Serpelloni, A. Falciano, C. Ferrari, Sea level rise and extreme events along the Mediterranean coasts: the case of venice and the awareness of local population, stakeholders and policy makers, Rendiconti Lincei. Sci. Fis. Nat. 35 (2024) 359–370, https:// doi.org/10.1007/s12210-024-01236-x. [76] I. Abd-Elaty, S. Straface, A. Kuriqi, Sustainable saltwater intrusion management in coastal aquifers under climatic changes for humid and hyper-arid regions, Ecol. Eng. 171 (2021) 106382, https://doi.org/10.1016/j.ecoleng.2021.106382. [77] S. Del Vecchio, S.K. Sharma, M. Pavan, A.T.R. Acosta, G. Bacchetta, F. de Bello, M. Isermann, R. Michalet, G. Buffa, Within-species variation of seed traits of dune engineering species across a European climatic gradient, Front. Plant Sci. 13 (2022), https://doi.org/10.3389/fpls.2022.978205. [78] P. Lionello, R.J. Nicholls, G. Umgiesser, D. Zanchettin, Venice flooding and sea level: past evolution, present issues, and future projections (introduction to the special issue), Nat. Hazards Earth Syst. Sci. 21 (8) (2021) 2633–2641, https://doi.org/10.5194/nhess-21-2633-2021. [79] C. Ferrarin, A. Valentini, M. Vodopivec, D. Klaric, G. Massaro, M. Bajo, E. Carraro, Integrated sea storm management strategy: the 29 October 2018 event in the Adriatic Sea, Nat. Hazards Earth Syst. Sci. 20 (1) (2020) 73–93. [80] G. Fontolan, S. Pillon, F.D. Quadri, A. Bezzi, Sediment Storage at Tidal Inlets in Northern Adriatic Lagoons : Ebb-Tidal Delta Morphodynamics , Conservation and Sand Use Strategies, vol. 75, 2007, pp. 261–277, https://doi.org/10.1016/j.ecss.2007.02.029. [81] D. Thulke, Y. Gao, P. Pelser, R. Brune, R. Jalota, F. Fok, M. Ramos, I. van Wyk, A. Nasir, H. Goldstein, T. Tragemann, K. Nguyen, A. Fowler, A. Stanco, J. Gabriel, J. Taylor, D. Moro, E. Tsymbalov, J. de Waal, D. Erasmus, ClimateGPT: towards AI synthesizing interdisciplinary research on climate change, Computer Science - Machine Learning (2024). http://arxiv.org/abs/2401.09646. [82] D. Kalla, N. Smith, S. Kuraku, F. Samaah, Study and analysis of chat GPT and its impact on different fields of study, International Journal of Innovative Science and Research Technology 8 (3) (2023) 827–833. https://ssrn.com/abstract=4402499. [83] S.A. Vaghefi, D. Stammbach, V. Muccione, J. Bingler, J. Ni, M. Kraus, S. Allen, C. Colesanti-Senni, T. Wekhof, T. Schimanski, G. Gostlow, T. Yu, Q. Wang, N. Webersinke, C. Huggel, M. Leippold, ChatClimate: grounding conversational AI in climate science, Commun. Earth Environ. 4 (1) (2023) 480, https://doi. org/10.1038/s43247-023-01084-x. [84] L. Reimann, A.T. Vafeidis, S. Brown, J. Hinkel, R.S.J. Tol, Mediterranean UNESCO world heritage at risk from coastal flooding and erosion due to sea-level rise, Nat. Commun. 9 (1) (2018), https://doi.org/10.1038/s41467-018-06645-9. [85] G.R. Tomasicchio, G. Salvadori, L. Lusito, A. Francone, A. Saponieri, E. Leone, S. De Bartolo, A statistical analysis of the occurrences of critical waves and water levels for the management of the operativity of the MoSE system in the venice lagoon, Stoch. Environ. Res. Risk Assess. 36 (9) (2022) 2549–2560, https://doi. org/10.1007/s00477-021-02133-7. [86] E.A. Hern´ andez-Delgado, The emerging threats of climate change on tropical coastal ecosystem services, public health, local economies and livelihood sustainability of small islands: cumulative impacts and synergies, Mar. Pollut. Bull. 101 (1) (2015) 5–28, https://doi.org/10.1016/j.marpolbul.2015.09.018. [87] M. Antonellini, P. Mollema, B. Giambastiani, K. Bishop, L. Caruso, A. Minchio, L. Pellegrini, M. Sabia, E. Ulazzi, G. Gabbianelli, Salt water intrusion in the coastal aquifer of the southern Po Plain, Italy, Hydrogeol. J. 16 (8) (2008) 1541–1556, https://doi.org/10.1007/s10040-008-0319-9. [88] C. Da Lio, E. Carol, E. Kruse, P. Teatini, L. Tosi, Saltwater contamination in the managed low-lying farmland of the venice coast, Italy: an assessment of vulnerability, Sci. Total Environ. 533 (2015) 356–369, https://doi.org/10.1016/j.scitotenv.2015.07.013. [89] I. Lovrinovi´ c, A. Bergamasco, V. Srzi´ c, C. Cavallina, D. Holjevi´ c, S. Donnici, J. Erceg, L. Zaggia, L. Tosi, Groundwater monitoring systems to understand sea water intrusion dynamics in the mediterranean: the neretva Valley and the Southern venice coastal aquifers case studies, Water 13 (4) (2021) 561, https://doi. org/10.3390/w13040561. [90] H. Nazarnia, M. Nazarnia, H. Sarmasti, W.O. Wills, A systematic review of civil and environmental infrastructures for coastal adaptation to Sea level rise, Civil Engineering Journal 6 (7) (2020) 1375–1399, https://doi.org/10.28991/cej-2020-03091555. [91] V. Giannini, A. Bellucci, S. Torresan, Sharing skills and needs between providers and users of climate information to create climate services: lessons from the Northern Adriatic case study, Earth Perspectives 3 (1) (2016) 1, https://doi.org/10.1186/s40322-016-0033-z. [92] B.C. O’Leary, L.E. Wood, C. Cornet, C.M. Roberts, C. Fonseca, Practitioner insights on challenges and options for advancing blue Nature-based solutions, Mar. Pol. 163 (2024) 106104, https://doi.org/10.1016/j.marpol.2024.106104. [93] R. ˇ Saki´ c Trogrli´ c, K. Reiter, R.L. Ciurean, S. Gottardo, S. Torresan, A.S. Daloz, L. Ma, N. Padr´ on Fumero, S. Tatman, S. Hochrainer-Stigler, M.C. de Ruiter, J. Schlumberger, R. Harris, S. Garcia-Gonzalez, M. García-Vaquero, T.L.F. Ar´ evalo, R. Hernandez-Martin, J. Mendoza-Jimenez, D.M. Ferrario, P.J. Ward, Challenges in assessing and managing multi-hazard risks: a European stakeholders perspective, Environ. Sci. Pol. 157 (August 2023) (2024), https://doi.org/ 10.1016/j.envsci.2024.103774. M.K. Dal Barco et al. International Journal of Disaster Risk Reduction 130 (2025) 105818 16