Can AI learn to understand nature? Scientists set out a roadmap for the future of ecology

Scientists have collaborated to develop a roadmap for the future of ecology with the help of artificial intelligence (AI). This international team of experts from ecology, conservation, and AI fields outlines how new AI models could revolutionize our understanding of ecosystems and predict their responses to environmental changes.

A significant challenge in ecology is making sense of the vast complexity of nature, given the diverse and varied data collected over the years. To address this, researchers propose developing ecological foundation models (ecoFMs), which are AI systems capable of learning from different types of ecological information and applying that knowledge to numerous ecological questions.

These foundation models, akin to large language models like ChatGPT, have the potential to transform various fields, including ecology. Dr. Rachel Parkinson, from Queen Mary University of London, emphasizes the unprecedented opportunity to merge different ecological knowledge using AI. While AI is already aiding in tasks like identifying animals in photographs, recognizing wildlife sounds, and mapping species distributions, current tools are primarily designed for specific tasks rather than providing a comprehensive understanding of ecological systems.

The proposed ecoFMs could move beyond this by learning about species interactions, how these relations depend on environmental conditions, and how changes impact broader ecological communities. An exemplary application could involve predicting which flowers an understudied pollinator might visit, based on combined data on plants, pollinators, including genetic relationships, physical characteristics, geographic distributions, and known interactions.

To achieve this, the roadmap outlines the necessary scientific and technical steps, including integrating diverse ecological data, incorporating ecological theory, developing new AI architectures, and evaluating prediction reliability. However, overcoming these challenges involves more than just technical innovation. The authors highlight that ecological data are often incomplete and biased, with many ecological relationships remaining poorly understood.

Therefore, the models need to acknowledge these limitations, quantify uncertainties, and be developed responsibly alongside the communities producing and using ecological knowledge. The authors stress the importance of developing transparent, reliable, and ethically responsible models that enhance our understanding of nature without reinforcing existing gaps and biases.

The full article, "Towards ecologically meaningful foundation models," was published in Nature Machine Intelligence and can be accessed via the DOI: 10.1038/s42256-026-01317-6.