In a keynote address to open the workshop, David Rolnick, McGill University and Mila–Quebec Artificial Intelligence Institute, provided an overview of how artificial intelligence (AI) tools can inform climate science, adaptation, and mitigation, providing overarching context for an audience with varying levels of familiarity and experience using AI in a climate science context; examined considerations around the environmental impact of AI; and discussed what areas of innovation are likely to be most useful in this space.
Rolnick described how AI tools can play a role in helping to facilitate climate action across many areas, including heavy industry,1 buildings and cities, land use, transportation, electricity systems, and societal adaptation, and highlighted some overarching themes illustrating AI’s relevance in these areas.2 One area where AI has strong potential is in turning large, unstructured climate data into information that is usable by decision makers, scientists, or other stakeholders. Rolnick gave the example of Climate
___________________
1 Heavy industry commonly refers to industries producing and/or requiring large-scale facilities and equipment and creating products intended for business or industrial use. Examples of heavy industry include petroleum refining, steelmaking, mining, and shipbuilding.
2 D. Rolnick, P.L. Donti, L.H. Kaack, K. Kochanski, A. Lacoste, K. Sankaran, A. Slavin Ross, et al., 2023, “Tackling Climate Change with Machine Learning,” ACM Computing Surveys 55(2): 1–96, https://doi.org/10.1145/3485128.
TRACE (Tracking Real-time Atmospheric Carbon Emissions)—a coalition of for-profit and nonprofit entities that monitor global greenhouse gas emissions—which is applying machine learning (ML) to remote-sensing data to gain real-time information on where and when emissions occur. Similarly, in disaster response, the United Nations and other agencies use AI to identify where flooding is occurring in real time and then apply this information to inform response efforts.
AI is also being used to improve operational and energy efficiency. For example, AI tools can help complex systems, such as heating, ventilation, and air conditioning systems or manufacturing processes, run more efficiently and thus use less energy and raw materials. AI is particularly useful for short-term predictions, or “nowcasting,” in the area of energy management. In electricity systems integrating solar and wind power, for example, AI approaches can be used to predict electric power supply and demand with highly accurate, fine-grained weather forecasts combined with energy use predictions. Getting these predictions right can help to avoid energy waste and support efficient operations while ensuring consumer needs are met.
Another opportunity for AI is in speeding up simulations. Traditional climate models can take months or even a year to run, but AI can sometimes speed up these processes by approximating some pieces of the model or, for shorter-term prediction, emulating the entire model. AI can also be used for downscaling (increasing the resolution of data from a simulation or real-time sensors, akin to sharpening a grainy picture). This can make time-intensive simulations more accessible, leading to meaningful action based on climate and weather data.
The energy demands to create and run AI models are large and growing, which has implications for climate-relevant areas, such as land use, resource extraction, energy generation, and greenhouse gas emissions. In thinking about the upsides—and potential downsides—of applying AI to address climate change, Rolnick said it is important to recognize that AI encompasses a range of very different algorithms. One way to conceptualize this is as a spectrum of animals representing different types of AI models. On one end are the “elephants,” which include large language models (LLMs) and generative AI systems, such as ChatGPT, Gemini, DeepSeek, and MidJourney. These models are highly energy intensive and rely on large-scale, centralized infrastructure. On the other end are the “bees,” which include algorithms for remote sensing, optimization, predictive modeling, and physics-informed ML. These tend to be smaller in scale,
distributed, and capable of running on more modest infrastructure, such as laptops or phones.
These “elephants” and “bees” are all AI models, but they are designed for different objectives, operate in different ways, and have different environmental impacts. Generative AI models are focused on producing text or images, whereas optimization and prediction algorithms are oriented toward classification, forecasting, or control. These distinctions influence not only the computational requirements and energy consumption but also the broader societal and environmental impacts. Rolnick said that the attention given to generative AI has created a perception that LLMs and chatbots are needed for almost every use, including forecasting floods or optimizing factory operations. However, for many climate issues, he said that elephants are not only excessive but also the wrong tool. Instead, he posited that most uses for AI fall into the bees category and suggested that the best strategy for avoiding unnecessary energy consumption is to maintain a few elephants for the tasks suited to them and a large population of bees for distributed, efficient, and context-specific uses. He added that it should not be assumed that an AI tool is even needed for every problem, noting that there are many situations in which conventional software engineering, search tools, or effective user interfaces are sufficient.
Rolnick pointed to some examples from his research group to illustrate why climate contexts often require beelike rather than elephant-like models. In one effort, they are using remote-sensing images to create agricultural maps that reflect how climate change affects productivity, which can inform governments on how to guide policy. For this goal, ChatGPT, DeepSeek, and similar large-scale generative AI approaches will not work; the project required an algorithm designed for the problem at hand.
The group developed Galileo, a self-supervised transformer designed specifically for remote sensing.3 They removed part of the data to train the transformer which then reconstructed the missing pieces, forcing the model to learn a meaningful representation that can be used for downstream tasks using very few labeled data. Galileo was used to create agricultural maps using self-supervised learning to combine data from multiple remote-sensing satellites and then used the signals themselves as supervision.
___________________
3 G. Tseng, A. Fuller, M. Reil, H. Herzog, P. Berkema, F. Bastani, J.R. Green, E. Shelhamer, H. Kerner, and D. Rolnick, 2025, “Galileo: Learning Global & Local Features of Many Remote Sensing Modalities,” pp. 60280–60300 in Proceedings of the 42nd International Conference on Machine Learning, https://arxiv.org/abs/2502.09356.
Galileo is more accurate and smaller than generative AI tools and requires fewer labeled data, which helps to overcome the dearth of labeled crop data around the globe.
This example demonstrates how focusing on user needs can drive impactful innovations. Tackling real-world user needs has led to cutting-edge solutions in areas such as out-of-distribution generalization, uncertainty quantification, explainability and interpretability, physical constraint satisfaction, multimodal data, limited label availability, and causality.4 In addition, tools that are built for one use case can end up being applicable in other appropriate situations; Rolnick noted that Galileo is being used to monitor the water stored in snowpacks because it works well for addressing common challenges encountered with remote-sensing data.
Rolnick also described how a global coalition of organizations is building scalable hardware, software, and algorithmic systems to monitor insect biodiversity.5 Automated camera traps are used to attract and photograph insects; images are then processed to identify species. The project uses ResNet50, a relatively standard AI model that is well suited to the job even though it is not necessarily cutting edge. On top of relatively simpler AI algorithms, the researchers have developed enhancements to solve particular problems. For example, a single camera trap in the cloud forest in Panama found 2,000 species of moths, 1,000 of which were new to science and undescribed. Since standard AI approaches cannot deal with data that are completely different from what has been seen before—a problem called “open-set recognition”—computer vision algorithms are required that can handle undescribed species.
These examples demonstrate that LLMs and chatbots are neither necessary nor suitable for all needs; while cutting-edge AI approaches are central to advance climate science and decision making, large-scale generative AI is usually not likely to be the solution. State-of-the-art AI is in both elephants and bees, Rolnick said, with a lot of different ways to build smarter AI to meet future needs. In addition, he emphasized that it is not enough to build algorithms; it may be essential to also develop the associated software and human capacity for the algorithms to be used effectively. Because of the global network of organizations deploying the systems he described, a lot of work was put into building the software stack so that ecologists, who
___________________
4 D. Rolnick, A. Aspuru-Guzik, S. Beery, B. Dilkina, P.L. Donti, M. Ghassemi, H. Kerner, et al., 2024, “Position: Application-Driven Innovation in Machine Learning,” pp. 42707–42718 in Proceedings of the 41st International Conference on Machine Learning, https://proceedings.mlr.press/v235/rolnick24a.html.
5 Y. Chen, N. Lang, B. Schmidt, A. Jain, Y. Basset, S. Beery, M. Larrivee, and D. Rolnick, 2025, “Open-Insect: Benchmarking Open-Set Recognition of Novel Species in Biodiversity Monitoring,” https://arxiv.org/pdf/2503.01691.
do not necessarily know how to run AI ML codes, can use them, he said, reiterating that impactful innovations are driven by thinking deeply about user needs.
For further resources, Rolnick pointed attendees to Climate Change AI,6 a nonprofit organization that catalyzes impactful work at the intersection of climate change and AI through hosting events, teaching classes, and offering funding to organizations. During a brief discussion following his talk, Rolnick addressed the concern that specialized open-source AI tools (“bees”) could be misused by nonexperts. He said that the greater risk lies in the misapplication of general-purpose models (“elephants”), where users may not realize they are getting wrong answers. He emphasized that AI should aim to scale expert judgment rather than replace it, since domain knowledge remains essential for interpreting results. He also suggested that the elephants and bees may eventually converge through a routing approach, in which chatbots provide the interface to steer users to the tools that will be useful to them, while specialized tools actually solve the hard problems.
___________________
6 See Climate Change AI, n.d., “About Climate Change AI,” htttps://www.climatechange.ai/about (accessed May 5, 2026).
This page intentionally left blank.