Preparing for a future of interconnected shocks

Our global future is projected to be marked by turbulence and a changing landscape of shocks like floods, fires, conflicts, and disease outbreaks. Can AI contribute to our analysis of such risks, and help identify novel solutions?

Introduction

The Anthropocene is characterized by a changing landscape of shocks, which are broadly defined as sudden events with noticeable impact (e.g., floods, fires, conflicts, disease outbreaks). As Delannoy et al. (2025)¹ show, not only have shocks become more numerous but also more co-occurring, especially between 1970 and 2000, meaning that multiple types of shocks increasingly struck the same country within the same year. Shocks also appear to revolve around the climate–conflicts–technology nexus, which suggests ramping up interconnection and potential lock-in effects between these sectors.² At the same time, shocks are difficult to unpack due to their non-linear features, resulting in changes that are difficult or even impossible to reverse (e.g., tipping points, regime shifts, trap situations) with large repercussions for people.³

Our knowledge about such shocks has expanded over time, as reflected in the gradual uptake of the multi-hazards and compound events literatures, which aim to describe how multiple drivers or hazards combine and contribute to societal or environmental risk.^4,5 However, multiple challenges remain for this knowledge to translate into practice. For instance, most shock databases suffer from significant limitations: non-standardized spatial and temporal classifications,⁶ underreporting of events,⁷ and gaps in impact estimates—particularly in low-income countries.⁸ Other issues lie for instance in the lack of understanding of causal relationships, or the difficulty in modeling uncertainty, whether that applies to single shocks or shocks’ interrelationships.⁹ A final challenge remains in the decision-making stage, as most existing approaches to risk management are designed for single shocks and single sectors,¹⁰ thus often failing to grapple with multiple co-occurring shocks. Artificial intelligence (AI) can in this context be a way forward to tackling some issues, but is not without limitations.

AI for Shocks Data

AI can be used for three distinct data-related processes in the context of interconnected shocks. First, AI can improve the reporting of shocks. For instance, convolutional neural processes (ConvNPs) have successfully been used to suggest where to place sensors to measure air temperature anomalies while minimizing uncertainty.^11,12 AI can also be used for extracting shocks data, notably from textual content. Sodoge et al. (2023)¹³ for example, applied several natural language processing (NLP) methods to automatize the detection of drought impacts from newspaper articles in Germany between 2000 and 2021. In the same vein, Li et al.⁶ created a database of climate extreme impacts (WikiImpacts), based on information extraction from textual content from Wikipedia, combined with a large language model (LLM)-based text classifier (ChatGPT 4o) for information selection. This effort allowed for the assemblage of essential data from over 22,000 disasters from 1900 to the present, and for benchmarking it against existing databases. The results show that extraction accuracy varies for different types of disasters, which suggests that AI is useful for enabling meta-comparison of databases. However, these methods did not outperform previously existing databases and still face coverage biases (e.g., an earthquake’s impacts can be under reported or even overlooked in news if it is elections week).

Second, AI can help break down broad or complex data into more detailed, local information, so-called downscaling. This is important because shocks like floods, heatwaves, or food shortages often have very different impacts at the local level. By using AI to zoom in on specific regions or communities, researchers can better understand who is most at risk and suggest more accurate responses. Lin et al.¹⁴ exemplified this by downscaling high-resolution daily near-surface meteorological variables over East Asia, using convolutional neural networks (CNNs). Mobin and Kamrujjaman,¹⁵ as another notable example, downscaled epidemiological time series data through their proposed Stochastic Bayesian Downscaling (SBD) algorithm, inspired by machine learning (ML) methods. However, while promising, AI-based downscaling techniques face challenges in ensuring accuracy and consistency across different variables and resolutions, especially in data-scarce environments.^16,17

Last, AI can be used to correct data, for instance by learning systematic biases between forecasts and observations, and correcting the probability distribution. A wide spectrum of approaches coexists, ranging from support vector machine (SVM)¹⁸ to ridge regression¹⁹ and random forest,²⁰ showing the high flexibility and applicability of AI data correction. However, such methods are mostly used in Earth observation contexts, where ML-based models may introduce statistical biases. Such bias stems from the fact that they are often trained on datasets calibrated with data from resource-rich regions, where the majority of weather stations are located. Another issue is that AI techniques can struggle to capture “extreme” events or shocks (those in the tails of the distribution) because the models are not explicitly penalized for failing to capture outliers.²¹ As such, a model can be good at predicting averages but perform poorly at detecting extremes.

AI-Assisted Predictive Modeling

AI is for now mostly used to model climatic, ecological, and economic shocks, yet in different contexts. For climatic shocks, deep learning (DL) architectures such as transformers, graph neural networks (GNNs), and physics-informed neural networks have a) transformed weather forecasting, and b) enabled the modeling of complex physical systems and long-range dependencies. Prominent DL models are employed across domains such as flood forecasting, air quality prediction, and drought assessment, with GNNs and transformers demonstrating superior performance, scalability, and interpretability in spatiotemporal forecasting tasks.^17,21

Analyses of ecological shocks also benefit from advances in ML methods for improved modeling. For instance, Ahvo et al. (2023)²² modeled the effects of agricultural input shocks (e.g., an abrupt export ban on fertilizers that affects the importing country) using a random forest algorithm, with considerable variation between crop–climate bin combinations, but overall good or very good model performance. Modeling contributions also explored the propagation of ecological shocks, notably in disease outbreaks, using DL methodologies like long short-term memory algorithm.²³

For economic shocks, a strong focus is given to modeling financial shocks, especially through advanced ML techniques such as random forest and extremely randomized trees.²⁴ However, the usefulness of such techniques depends on the underlying economic theory, as models based on neoclassical economics often overlook financial instability, whereas post-Keynesian approaches offer a more suitable framework for analyzing financial shocks, for instance.²⁵

For disaster-related shocks, the literature analysis conducted for this report indicates that natural disaster management studies predominantly rely on classic ML algorithms, such as random forests, SVMs, and ensemble methods. Studies on DL techniques are increasing, including CNNs and recurrent neural models for tasks like flood forecasting, risk mapping, and event detection.

However, most hazard prediction applications remain single-hazard focused, or implicitly incorporate multi-hazard interactions through the integration of diverse static and dynamic variables. In other words, existing approaches focus on the “what,” “when,” and “where” questions, not on the “why,” “what if,” and “how confident.”²¹

AI-Assisted Decision-Making

AI is rarely used openly in decision-making for interconnected shocks. Although such uses have shown potential to improve real-time responses to extreme weather²⁶ or to enhance crisis communication,²⁷ its use in crisis decision-making is limited by institutional and ethical challenges. Indeed, such decision-making processes often happen under high pressure, with incomplete or unreliable information and high levels of uncertainty, and involve difficult moral choices. These conditions make it hard to trust or rely on AI systems.²⁸ However, AI can still play a valuable supporting role in specific contexts, particularly where cognitive biases, urgency, and data overload might limit human performance. For example, while doctors may not fully trust AI in medicine, it’s used to screen large volumes of medical images to identify potential cancers²⁹ and prioritize which cases need human review. Similarly, AI-based forecasting models are deployed in response to pandemic outbreaks, before widespread testing becomes available, to guide early containment strategies and resource allocation.³⁰

Potential and Limitations

AI holds promise for shocks-related context, and especially for developing early warning systems (EWSs) for conflicts,³¹ ecological shocks,^32,33 and their intersections.³⁴ EWSs such as the Violence & Impacts Early Warning System (VIEWS)³⁵ have demonstrated the capacity to forecast political violence up to three years in advance,³⁵ while AI tools for climate hazard detection now offer probabilistic assessments to support agricultural resilience.³⁶

Yet, AI-enabled EWSs are only as reliable as the susceptibility layers that underpin them. Susceptibility captures the conditions that turn a hazard into harm: exposure to the shock, sensitivity of livelihoods and ecosystems, and the capacity of people and institutions to cope or adapt.3 It is therefore affected by trade and other forms of connectivity that transmit disturbances, unequal access to economic resources and information leading to harm on marginalized groups, and by the capacities of governments to respond. Recent AI pipelines can blend satellite imagery, mobility data, and economic networks to sketch these patterns,^37,38 but the source data still lean toward well-monitored, affluent regions and may reproduce historical social bias.³⁹

As a way forward, Reichstein et al. (2025)⁴⁰ propose shifting toward causal AI models and decadal early warnings to avoid misleading short-term predictions. They also call for strict adherence to the FATES principles (Fairness, Accountability, Transparency, Ethics, and Sustainability), including open access to training data and source code to ensure replicability. Still, major obstacles remain: institutional inertia, ethical dilemmas, infrastructure limitations, and the broader difficulty of embedding AI analysis into decision-making processes already under stress by urgency, uncertainty, and moral complexity.

THEME BOX 1

Foundation modeling in climate and sustainability science

Authors: David Montero, Miguel Mahecha

What are foundation models?

Foundation models are large-scale AI systems trained on vast datasets to serve as general-purpose tools across multiple tasks. Instead of training separate models for each problem, a foundation model learns broad representations, allowing
adaptation to diverse applications (Fig. 5). These models, typically transformer-based, process large volumes of data, capturing patterns that can be fine-tuned for specific tasks.⁴³

How Can Foundation Models Help in Climate and Sustainability?

Climate and sustainability sciences deal with complex, data-intensive challenges that benefit from AI-driven pattern recognition.⁴⁴ Foundation models trained on decades of ecological ground data, Earth observation (EO) data from space, and climate records over land, ocean, and in the atmosphere shall generalize across modalities, variables, and scales. This is important, given the coupled dynamics of Earth System processes,⁴⁵ for enabling consistent cross-domain AI applications. Examples of foundation model applications for sustainability include: forecasting of multiple variables of interest (e.g., weather-related variables), classification of land-surface properties (e.g., land cover, land use, their change, and objects), environmental disaster detection, and biodiversity monitoring. By leveraging pre-trained knowledge, foundation models require less data and computational effort for specialized tasks.⁴⁶

Existing Climate and Sustainability Foundation Models

Several foundation models have emerged to support climate and sustainability sciences.⁴⁷ In EO, models such as Clay,⁴⁸ Prithvi-100M⁴⁹ (IBM and NASA), and SpectralGPT⁵⁰ leverage state-of-theart AI models trained on satellite data, including Landsat and Sentinel archives. These models enable applications like land cover classification, disaster monitoring, and environmental change detection. Beyond EO, climate-focused foundation models are also gaining traction. Models such as ClimaX⁵¹ (Microsoft), Aurora⁵² (Microsoft), and Prithvi WxC⁵³ (IBM and NASA) are trained on diverse climate datasets and reanalysis data. Furthermore, the AlphaEarth (Google and Google DeepMind) model was trained on a combination of multiple sources, including EO and climate datasets, as well as geophysical parameters and text embeddings.⁵⁴ These models support downscaling, climate projections, and probably extreme event forecasting, although the latter poses particular challenges.⁵⁵

Potential of Foundation Models

Foundation models optimize AI model development through transfer learning, requiring a minimal set of training data for novel tasks. Their ability to capture multi-scale patterns, both spatial and temporal, across Earth Systems allows the unification of traditionally separate tasks, such as climate scenario downscaling and extreme weather forecasting.⁵³ This capability also extends to bridging gaps between observational data and numerical simulations, allowing more comprehensive Earth System analyses. Additionally, numerical predictions are computationally expensive,⁴⁷ and foundation models offer a more efficient alternative by significantly reducing inference time while maintaining competitive accuracy.⁴⁶

Challenges and Limitations

Despite their promise, foundation models require vast computational resources for training, making them costly to develop. Additional challenges are the heterogeneity and distributed nature of Earth System data. Models trained on limited or unbalanced datasets may struggle to generalize, and purely data-driven approachesrisk producing outputs that violate known physical laws. Ensuring physical consistency is particularly important for scientific credibility and decision-making. This can be addressed by incorporating physical knowledge into model design or combining data-driven and physics-based approaches. Data-related biases (e.g., clustered sensor networks for in-situ data or sensor degradations in satellite data) further complicate generalization, especially when models encounter conditions outside their training distribution, such as “record shattering” climate extreme events.^55,56 These risks are amplified in regions with sparse observations or limited data infrastructure, potentially leading to lower model accuracy and under representation. Interpretability remains a concern, as complex deep learning architectures function as black boxes, making it difficult to validate predictions. Bias in training data can also lead to uneven model performance, necessitating careful evaluation and governance.⁴³

Equity and Access Considerations

While foundation models offer powerful capabilities across climate and sustainability sciences, they also present challenges related to access and usability. Running or adapting these models often requires significant computational infrastructure and technical expertise, which may not be readily available to all stakeholders, including smaller research teams, local agencies, or community organizations. Moreover, effectively interpreting foundation model outputs frequently demands domain-specific and AI expertise. Making foundation models more broadly accessible will require efforts such as simplified model interfaces, improved documentation, and dedicated efforts to support users in building the skills needed to apply foundation models effectively.