Generative AI can help scale climate services to meet growing demand. This approach is illustrated here with two prototype AI systems designed to enhance access to climate data, support decision-making, and improve efficiency. Key concerns are identified around responsibility, quality, and trustworthiness of AI outputs. Emerging best practices in system development are linked to core principles of salience, credibility, and legitimacy, to ensure alignment with user needs and societal values.Motivation: Can AI help to scale climate services?Decades of climate change research have created a huge volume of knowledge and expertise1, yet its effective use to inform climate adaptation is challenged by a bottleneck of human labour. Climate services—the provision and use of climate change knowledge to support decision-making—are difficult to scale. Increasing unmet demand for climate-related skills in the labour market2 suggests there are too few climate specialists to provide in-depth guidance to the many people and organisations who must respond to climate change in the coming years. While climate mitigation and decarbonisation efforts continue, “committed warming” and observed emissions trajectories indicate around 2.5–3 °C of warming by 21003. Adaptation is made more challenging by the need for localised, context-dependent, and impact-specific strategies4.Meanwhile, new generative AI technologies present the intriguing possibility of partial automation of climate services, potentially removing barriers to access and improving decision-making. While conventional software tools (e.g. data platforms, decision-support tools) can help meet demand, generative AI might offer enhanced ease of use, greater autonomy and/or automation capacity, and the ability to connect diverse data formats across domains. Yet such technologies are unproven and might inadvertently weaken or disrupt the effective use of climate science if they are not deployed in a responsible manner (i.e. aligned with common responsible innovation principles of anticipation, reflection, engagement and action5). The climate services context brings domain-specific risks around AI systems that (e.g.) misinterpret climate science, confuse models or emissions scenarios, or do not track uncertainty. Here we consider how generative AI might help deliver climate services at scale, and the need for such methods to be trustworthy, equitable, and integrated with human workflows. We identify autonomy, flexibility and uncertainty as shared features of generative AI tools in the climate services context, as well as some emerging best practices to improve system quality.Examples of existing AI tools for accessing climate information“Generative AI” refers to the ability of some kinds of machine learning system, most commonly large language models (LLMs, very large neural network models trained on enormous text datasets), to create multimodal content (e.g. text, images, video, audio) that is fast becoming indistinguishable from human-created content. Such models are the basis for popular AI applications like chatbots, image-generators, virtual assistants and analytical tools, which are beginning to appear in the climate services domain (e.g. refs. 6,7,8,9). Subject to critical evaluation, such tools might be integrated with climate services workflows to help synthesise information and support human decision-making.The AI tools so far reported in the climate services literature (e.g. refs. 6,7,8,9) have diverse purposes, datasets, and approaches to LLM integration. As an illustration from a UK perspective, here we consider two such applications from our own work. Figure 1 shows schematics for a chatbot designed to communicate information from the UK Climate Projections (UKCP18) (Fig. 1a) and for a tool for generating climate service “recipes” based on a large evidence base of climate models and adaptation solutions (Fig. 1b). Both tools adopt an “agentic” approach (i.e. modular systems which decompose the larger task into smaller sub-tasks that are each handled by semi-autonomous LLM-based agents). Both also use “human-in-the-loop” mechanisms for resolution of ambiguities. Both were co-created with end users at the UK Met Office, with users involved in shaping and evaluating the systems.Fig. 1: Schematic diagrams showing workflows for illustrative examples of agentic AI tools for climate services.Full size imagea Multimodal chatbot and data analysis tool for UKCP. 1. User interacts with chat interface and data workspace. 2. Chat interface provides multimodal outputs and data visualisation. 3. Data workspace holds extracted datasets, analysis/visualisations, and chat history. 4. RAG agent supports multimodal question-answering based on UKCP archive. 5. Data agent enables data extraction and generates code to perform analysis/visualisation. b Agentic workflow for generative climate service recipes. 1. User provides input: templated task/query and project description. 2. Reasoning LLM is prompted to create a “naïve” draft recipe. 3. Knowledge graph collates climate service case studies and climate models/datasets. 4. Multiple LLM agents (supervisor & specialists) augment/improve naïve recipe with specific climate models/datasets/procedures, constrained by knowledge graph. 5. Output is an actionable “final” recipe provided back to the human user.UKCP18 was produced by the UK Met Office to provide robust climate projections and aid planning and decision-making for the UK until 2100. It consists of scientific reports (containing text, images, tables and graphs), large numerical datasets of climate model outputs (in gridded/tabular form), and various ancillary content (e.g. visual summaries, documentation). Despite many applications, many users highlight difficulties in applying UKCP18 outputs without technical support10. Our experimental chatbot (Fig. 1a) seeks to overcome this usage barrier by using natural language interactions to provide user-friendly and fast access to UKCP18 data/information, with question-answering, data retrieval, and data analysis functionalities. It uses retrieval-augmented generation (RAG)8 to ground LLM responses in the UKCP18 knowledge base and reduce the risks of inaccuracy or hallucination (generating a false or unfounded answer that is not based on real information11). The RAG tool is multimodal and can answer with relevant images/graphs or datasets as well as text. Human-in-the-loop validation is used when the system detects an ambiguous query, especially during data selection tasks where precision on spatial location, time period and climate variables is needed. The chatbot also supports interactive analysis of raw climate model data, generating on-the-fly code that enables users to (in natural language) select different climate variables, locations and time periods for visualisation, comparison, and statistical analyses. Many users previously required input from a (human) climate consultant to access these functions.Our second example tool (Fig. 1b) is an agentic system designed to automatically generate climate service recipes (i.e. detailed workflows for application of climate science in a given scenario or context). This tool aims to help human climate service providers to easily identify a method for using climate data in a particular client context, reducing the time and cost of delivery. The tool combines LLM reasoning with contextual knowledge from a curated knowledge graph12 that maps relationships between climate variables, emission scenarios, indices, hazards, sectors, and key datasets (e.g. CORDEX, CMIP5, UKCP18). The agentic workflow begins with a description of a climate scenario; human-in-the-loop confirmation is used to augment the initial description and complete a scenario template for further processing. The architecture involves a memory-enabled supervisor agent that orchestrates multiple specialised agents to plan workflows, create draft recipes, query the knowledge graph, source external (non-climate) domain knowledge, and summarise the final output. This multi-agent approach effectively generates “recipes” that specify appropriate methodologies and datasets, with clear traceability back to the knowledge graph to encourage user trust. The approach shows good potential to increase efficiency of climate service design across multiple hazards and domains, enabling human consultants to deliver climate services more rapidly.Emerging best practices for trustworthy AI developmentFrom the two examples above, and many others in the literature (e.g. refs. 6,7,8,9,11,13), we can identify three characteristic features of generative AI systems that are particularly relevant to the climate services context. Firstly, autonomy; the LLM components, and therefore the wider agentic systems, are given an amount of freedom in how they respond to a query or complete a task. Second, flexibility; these tools can respond effectively to a wide range of inputs and generalise to novel situations. Third, uncertainty; the inherent complexity and stochasticity of LLMs makes derived systems somewhat opaque, and it is hard to be sure what answer they will give to a given query. These three attributes give these systems their ability to generate useful and diverse outputs, but also introduce unpredictability and risk of inaccuracy; these are undesirable features in the context of climate services, where precise use of climate science, terminology, and data/models is essential.Much effort in system development is expended on constraining such tools to reduce uncertainty and avoid undesirable behaviours. The examples above, and others in the literature, have done this by: (1) careful prompt engineering to steer the behaviour of LLMs, by both users in their queries and developers in internal system prompts; (2) controlled access to data and resources, such as RAG systems that force answers to be generated only from a given database; (3) embedding transparency in system behaviour using visible reasoning steps and traceability of responses; (4) using human-in-the-loop confirmation to resolve ambiguities; and (5) rigorous evaluation of outputs using both quantitative and qualitative metrics. Implementing these practices requires careful consideration. For example, human-in-the-loop mechanisms can only scale when the human is the end user (who is always present) rather than a developer (who is no longer involved at point of delivery). Development should also adapt to the target user audience; tools for public use might require different standards to those designed for experts. Yet ultimately, the three core features of autonomy, flexibility, and uncertainty are simultaneously the strength and the potential weakness of generative AI tools. Good engineering can improve performance and consistency, but the inherent uncertainties of such technologies suggest that careful human steering (of both development and usage) will be needed to achieve trustworthy and high-quality outcomes. Users and developers both share responsibility for the accurate generation and careful interpretation of AI system outputs.AI for climate services should prioritise salience, credibility and legitimacyBeyond implementation, questions remain about whether creation of automated tools for climate services is desirable or responsible. Responsible innovation principles (e.g. the AREA framework5) should be followed in developing any new AI technology, but the climate domain has particular features. Climate services are delivered in complex contexts currently requiring expert human judgement; uncritical adoption of AI-generated solutions and loss of human oversight might have serious consequences. Emerging standards that seek to regularise and improve quality of climate service delivery (e.g. ref. 14) might be disrupted or undermined by integration of AI tools. To explore these issues, we held a UK-based stakeholder workshop (Box 1) that identified additional risks around lack of accountability, inequality of access, and value alignment (Table 1). These views, while exploratory and drawn from a UK-only sample, align with previously reported concerns in the climate services literature (e.g. ref. 15).Here we suggest some guiding principles for AI deployment in climate services (Fig. 2), framed around the foundational climate service concepts of salience, credibility and legitimacy16. To ensure salience, AI tools should provide climate information specific to user context and decision scenario. Furthermore, to date there has been little exploration of whether such tools are useful; future behavioural experiments under controlled conditions could evaluate whether their use improves or worsens decision-making outcomes17. A perspective of continuous improvement and user engagement should be adopted to ensure ongoing alignment with community/stakeholder needs.Fig. 2Full size imageRisk factors and suggested good practices for generative AI tools for climate services.Considering credibility, the central concerns for any new AI tool for climate services should be trust and reliability, with robust evaluation (including co-evaluation by stakeholders) to ensure outputs are accurate and consistent with established climate science. Both quantitative and qualitative evaluation approaches are valuable, while benchmarks and tests should be tailored to the specific context (ideally with engagement from stakeholders). AI tools for climate services should adopt emerging best practices for creating “trustworthy” AI systems, prioritising transparency and interpretability to avoid undesirable “black boxes” and build user confidence18.To aid legitimacy, AI tools for climate services should seek to mitigate concerns about social, cultural or geographic biases, and meet common standards of political/procedural fairness. Participatory design and co-creation can help increase legitimacy by ensuring that tools reflect user needs and local contexts16. To aid transparency and accountability, tools should have clear governance, follow good software development practices (e.g. version control, data updating), and give clear information about ownership, construction, inputs, and underlying models, as well as responsible usage/interpretation and liability guidelines for outputs.Box 1 Exploring concerns around AI for climate servicesA workshop held at the Met Office (Exeter, UK) in February 2024 explored AI’s impact on climate services, focusing on legitimacy, credibility, and saliency16. Thirty-four UK-based climate scientists, service providers, and AI practitioners were selected to represent the climate services community. All participants completed a pre-workshop survey to identify: (i) potential challenges and risks associated with integrating AI into climate services; and (ii) potential solutions or pathways for addressing these challenges. Then 24 participants attended in person from academia (n = 14), government (n = 4), and commercial sectors (n = 6). In facilitated discussions, the group considered both opportunities and risks of AI, mainly generative tools like LLMs, and proposed strategies to address potential challenges. Workshop findings were later collated from recurring themes identified across survey responses and discussion outcomes. While the UK-based group and qualitative approach limit the generality of the findings (e.g. across geographies), the workshop produced useful exploratory findings from a representative group.Identified risks (Table 1) echoed those previously reported in climate services literature (e.g. ref. 15): concerns about fairness, transparency, trustworthiness, and value alignment, especially regarding which values are reflected in AI outputs. AI-specific risks included misappropriation of information, model biases, lack of accountability (of both commercial providers of LLMs and developers of AI-based climate services applications), and the potential to worsen inequalities in access to climate services (cf. the “digital divide”). Solutions such as codes of practice, accreditation, co-production, and interdisciplinarity were suggested, aligned with established best practices, showing that while generative AI introduces new issues, it also reinforces proven approaches (see Fig. 2).Table 1 Selected quotes from workshop participants illustrating concerns about AI tools in climate servicesFull size tableConclusion: Responsible AI can help deliver climate servicesIf the concerns outlined above can be addressed, generative AI technologies offer substantial potential benefits for delivering climate services at the scale needed to support the local, national and international actions needed for climate adaptation and resilience. Specialised tools could assist human practitioners to deliver projects faster, cheaper and more consistently, thereby increasing accessibility, relevance and quality for end beneficiaries. AI tools and applications could be deployed at scale via the Web, in most languages and geographic regions, with minimal cost at point of use. This would help increase access for people and organisations where language and/or resource constraints currently prevent effective uptake of climate knowledge. However, achieving these potential benefits will require careful deployment to avoid worsening digital divides (e.g. by alignment with inclusive design standards19) and ensure usage costs do not become prohibitive (e.g. by use of open source models). It is also important that deployment of AI for climate services does not worsen climate change; systems might be deployed on data centres using renewable energy sources and use ‘green prompting’20 to reduce their energy use and associated greenhouse gas emissions. Expanding our collective ability to apply climate science to real world adaptation challenges might enable people everywhere to make better decisions, as climate impacts start to affect their lives.Data availabilityNo datasets were generated or analysed during the current study.ReferencesIPCC. Climate Change 2023: Synthesis Report. Contribution of Working Groups I, II and III to the Sixth Assessment Report of the Intergovernmental Panel on Climate Change. (Core Writing Team, Lee, H. & Romero, J. eds) (IPCC, 2023).World Economic Forum. The Future of Jobs Report 2025. (World Economic Forum, Geneva, 2025).United Nations Environment Programme. Emissions Gap Report 2025: Off Target – Continued Collective Inaction Puts Global Temperature Goal at Risk (United Nations Environment Programme, 2025).New, M. et al. Decision-making options for managing risk. In Climate Change 2022: Impacts, Adaptation and Vulnerability (eds Pörtner, H.-O. et al.) 2539–2654 (Cambridge University Press, 2022).Stilgoe, J., Owen, R. & Macnaghten, P. Developing a framework for responsible innovation. Res. Policy 42, 1568–1580 (2013).Article Google Scholar Vaghefi, S. A. et al. ChatClimate: Grounding conversational AI in climate science. Commun. Earth Environ. 4, 480 (2023).Article Google Scholar Thulke, D. et al. ClimateGPT: Towards AI synthesizing interdisciplinary research on climate change. ArXiv preprint at https://arxiv.org/abs/2401.09646 (2024).Biswas, A. et al. Evaluating retrieval augmented generation to communicate UK climate change information. In Proc. 2nd Workshop on Natural Language Processing Meets Climate Change (ClimateNLP 2025), 126–141 (Association for Computational Linguistics, 2025).Kuznetsov, I. et al. Transforming climate services with LLMs and multi-source data integration. npj Clim. Action 4, 97 (2025).Google Scholar Mittal, N., Fung, F., McSweeney, C. F. & Lowe, J. A. Identifying user requirements to inform national climate projections in the UK. Clim. Serv. 40, 100613 (2025).Article Google Scholar Farquhar, S. et al. Detecting hallucinations in large language models using semantic entropy. Nature 630, 625–630 (2024).Article Google Scholar Ceriani, M. et al. Domain knowledge graph instantiation and evaluation. Zenodo https://doi.org/10.5281/zenodo.14783009 (2025).Park, J. S. et al. Generative agents: interactive simulacra of human behavior. In Proc. 36th Annual ACM Symposium on User Interface Software and Technology (UIST ‘23), 1–22 (ACM, 2023).Doblas-Reyes, F. J. et al. Standardisation of equitable climate services by supporting a community of practice. Clim. Serv. 36, 100520 (2024).Article Google Scholar Findlater, K. et al. Climate services promise better decisions but mainly focus on better data. Nat. Clim. Chang. 11, 731–737 (2021).Article Google Scholar Cash, D. et al. Salience, credibility, legitimacy and boundaries: linking research, assessment and decision making. SSRN https://doi.org/10.2139/ssrn.372280 (2002).Hauser, O. P., Light, M., Shelmerdine, L. & Blumenau, J. Why evaluating the impact of AI needs to start now. Nature 643, 910–912 (2025).Article Google Scholar Kaur, K., Uslu, S., Rittichier, K. J. & Durresi, A. Trustworthy artificial intelligence: a review. ACM Comput. Surv. 55, 39 (2022).Google Scholar UNDP - Chief Digital Office. Inclusive by Design: Accelerating Digital Transformation for the Global Goals (Policy Brief, 2022).Podder, S., Date, H. & Murthy, S. Green prompt engineering for sustainable generative AI. Environ. Sci. Ecotechnol. 30, 100684 (2026).Article Google Scholar Download referencesAcknowledgementsH.W. and H.X. acknowledge funding from the UKRI AI Hub on Collective Intelligence (EPSRC Grant No. EP/Y028392/1). A.B. is funded by the UKRI AI CDT on Environmental Intelligence. N.M., F.F. and A.A. acknowledge funding from the HACID project (EU grant ID: 101070588). No funder had any influence on the content of this work.Author informationAuthors and AffiliationsCentre for Environmental Intelligence, University of Exeter, Exeter, UKHywel T. P. Williams, Arjun Biswas & Hailun X. XieMet Office, Exeter, UKAnrijs Abele, Fai Fung, Jason A. Lowe & Neha MittalCentre for Climate Research, Singapore, SingaporeFai FungUniversity of Bristol, Bristol, UKFai FungPriestley Centre, University of Leeds, Leeds, UKJason A. LoweAuthorsHywel T. P. WilliamsView author publicationsSearch author on:PubMed Google ScholarAnrijs AbeleView author publicationsSearch author on:PubMed Google ScholarArjun BiswasView author publicationsSearch author on:PubMed Google ScholarFai FungView author publicationsSearch author on:PubMed Google ScholarJason A. LoweView author publicationsSearch author on:PubMed Google ScholarNeha MittalView author publicationsSearch author on:PubMed Google ScholarHailun X. XieView author publicationsSearch author on:PubMed Google ScholarContributionsAll authors helped shape the ideas and material presented in this manuscript. H.W. wrote the initial draft and then all authors commented and edited the manuscript. A.B. contributed content on the chatbot tool. A.A. and H.X. contributed content on the climate services tool. N.M. and F.F. contributed content on the ethics/governance workshop. J.L. contributed content on the climate services context. All authors reviewed and approved the manuscript.Corresponding authorCorrespondence to Hywel T. P. Williams.Ethics declarationsCompeting interestsThe authors declare no competing interests.Additional informationPublisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.Rights and permissionsOpen Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/.Reprints and permissionsAbout this article