The recent preprint 'Earth Science Foundation Models: From Perception to Reasoning and Discovery' by Xiangyu Zhao and colleagues (arXiv, 2026) introduces a transformative approach to Earth science through AI foundation models (Earth FMs). These models integrate diverse data sources—such as satellite imagery, gridded reanalysis data, geophysical observations, and domain-specific texts—to advance tasks from basic environmental perception to complex scientific reasoning and discovery. The authors frame their review across two dimensions: depth (evolution from perception to agentic workflows) and breadth (applications across Earth’s systems like atmosphere, hydrosphere, and biosphere). While the paper offers a comprehensive roadmap with over 200 datasets and benchmarks, it stops short of addressing real-world implementation challenges and societal implications, which are critical given the escalating environmental crises.

Earth FMs represent a significant leap in climate modeling by enabling multimodal data fusion, a capability that traditional models often lack. For instance, combining real-time satellite imagery with historical climate data could improve the accuracy of extreme weather predictions, a pressing need as events like hurricanes and heatwaves intensify due to climate change. However, the preprint’s focus on technical advancements overlooks the accessibility of these tools for policymakers and local communities, who are often on the front lines of climate impacts. This gap mirrors a broader pattern in AI research: while innovation accelerates, deployment and equity considerations lag behind.

Contextually, this work aligns with a surge in AI applications for environmental solutions, such as Google’s DeepMind using machine learning for wind energy optimization (Nature, 2019) and NOAA’s adoption of AI for weather forecasting (Bulletin of the American Meteorological Society, 2021). Yet, unlike these applied efforts, Zhao et al.’s review remains theoretical, missing a discussion on how Earth FMs could integrate with existing systems or address data biases—issues that have historically plagued climate models, often underrepresenting vulnerable regions. For example, datasets from the Global South are frequently incomplete, risking skewed predictions that could exacerbate environmental injustice.

Another underexplored angle is the sustainability of Earth FMs themselves. Training large AI models consumes significant energy, a concern raised in studies like Strubell et al. (2019, arXiv) on AI’s carbon footprint. Given that Earth FMs aim to combat climate change, their environmental cost must be weighed—a point absent from the preprint. Furthermore, the transition to 'agentic and embodied Earth intelligence,' as proposed by the authors, raises ethical questions about autonomous decision-making in environmental contexts. Could AI-driven agents prioritize certain regions or crises over others based on flawed data or programming? This potential for unintended consequences deserves scrutiny.

Looking ahead, Earth FMs could revolutionize not just prediction but also policy simulation, helping governments test climate interventions before implementation. However, without addressing scalability, data equity, and public trust, these models risk becoming academic exercises rather than actionable tools. The methodology of the preprint is a literature review with no primary data or experiments, limiting its claims to conceptual insights rather than empirical validation. As a non-peer-reviewed work, its findings await rigorous scrutiny. Still, with over 200 cited datasets, it provides a robust starting point for future research.

Ultimately, Earth FMs connect to a larger narrative of data-driven environmental solutions, often underrepresented in popular media, which tends to focus on crisis rather than innovation. By bridging this gap, we can foster a more proactive approach to the climate crisis, ensuring technology serves as a partner, not a panacea.