This paper explores the critical intersection of Artificial Intelligence (AI) and public climate services (CS). While AI offers transformative potential to make climate data more accessible and actionable, the authors warn that a growing dependency on private, non-European infrastructure—such as proprietary cloud platforms and foundation models—threatens the public mandate of climate safety and societal resilience.
TL;DR
The paper analyzes the risks and opportunities of integrating AI into public climate services using a PESTLE framework. It highlights a dangerous "mismatch" where public safety-critical services increasingly rely on privately controlled, non-sovereign AI infrastructure, potentially undermining democratic oversight and long-term equity.
What Problem Does This Paper Address?
Climate services (CS) are essential tools that translate complex climate data into actionable insights for sectors like agriculture, energy, and disaster management. Traditionally, these have been the domain of public national meteorological services. However, the "AI revolution" is shifting the landscape.
The authors identify a core tension: CS are a "public good" intended to protect life and property, yet the AI capabilities required for the next generation of these services (GPUs, cloud hosting, and Large Language Models) are owned and governed by a handful of private, mostly US-based "hyperscalers." This creates a strategic dependency. If a private provider changes its terms of service or a proprietary model produces biased results, the public's ability to respond to climate risks could be compromised.
Methods
The researchers employed a multi-step qualitative assessment:
- PESTLE Framework: They categorized potential impacts into six dimensions: Political, Economic, Social, Technological, Legal, and Environmental.
- Consensus-Based Elicitation: Four independent experts (assessors) evaluated 72 distinct items across three layers of climate services: Data, Model, and Infrastructure/Quality Assurance (QA).
- Scoring System: Items were rated on a scale from -4 (Existential Risk) to +4 (Transformative Opportunity).
The study also distinguished between Upstream providers (who generate foundational data, like Copernicus) and Downstream providers (who tailor that data for specific end-users).
Datasets
As a policy and governance analysis, this paper does not introduce a new machine-learning training dataset. Instead, it references and analyzes the ecosystem surrounding existing public data repositories, primarily the Copernicus Climate Change Service (C3S) and the Climate Data Store. The authors also discuss the potential of citizen science and participatory data gathering to fill gaps in the Global South.
Key Insights and Ideas
- The Infrastructure Mismatch: We are moving from an "epistemic gap" (how to understand data) to an "infrastructure gap" (who owns the tools to process it). While upstream data is largely open, the downstream tools used to interpret it are increasingly proprietary.
- Political Ambivalence: Political factors scored the highest for both risk and opportunity. AI could either strengthen "digital sovereignty" through European-led initiatives or lead to "mission drift" where security concerns limit data openness.
- Economic "Lock-in": Downstream services are the only area where risks currently outweigh opportunities. Reliance on proprietary APIs creates high "switching costs," weakening the bargaining power of public institutions.
- The "Valley of Death": AI systems are cheap to prototype but expensive to maintain reliably in real-world decision-making. This gap often prevents promising public-interest AI tools from becoming operational.
Key Results
The authors used a specific scoring table to quantify these qualitative expert opinions:
- Technological Optimism: Technological opportunities (+2.96) generally outscored technological risks (-2.33), particularly in the downstream market where AI can spur innovation in bespoke services.
- Upstream vs. Downstream: Risks tend to intensify as services move closer to the end-user (downstream). For example, biased AI models can lead to "maladaptation"—investing in the wrong types of climate resilience based on flawed automated advice.
- Reliability: The study found moderate agreement among experts ($\alpha = 0.25$), suggesting that while there is a consensus on the direction of these trends, the exact magnitude of the threat is still a subject of intense debate within the field.
Limitations and Future Work
The authors acknowledge that their study is explorative and based on a small pool of four assessors. They suggest that future work should involve a much broader community of stakeholders to refine these scores.
A major unresolved issue is legal liability: who is responsible when an AI-driven climate service provides "wrong" advice that leads to financial loss or physical harm? The authors argue that as long as infrastructure and interfaces are privately owned, accountability remains dangerously fragmented. They call for a "pan-European, decentralised public climate service" built on sovereign AI infrastructure to mitigate these risks.
Summary updated