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

Figure illustrating the different types of information available to ecological researchers and the challenges of integrating them. Credit to Kelly Cassella at MadeBold Design

A new Perspective article led by Dr Rachel Parkinson of Queen Mary University of London, published in Nature Machine Intelligence on 6 November, sets out a roadmap for developing artificial intelligence (AI) models that could transform how scientists study the natural world.
The international team, bringing together experts in ecology, conservation and AI, proposes a way to tackle one of ecology's biggest challenges: making sense of the extraordinary complexity of nature.

Scientists have collected enormous amounts of information about the natural world, from genetic sequences and satellite imagery to wildlife recordings, species observations and decades of ecological research. But these data come in vastly different forms and at different scales, making it difficult to combine them to understand how organisms interact with each other and their environments.

The researchers propose developing ecological foundation models (ecoFMs): AI systems that could learn from these diverse sources of information and apply that knowledge to a wide range of ecological questions.

Foundation models underpin many of today's most advanced AI technologies, including large language models such as ChatGPT. Trained on vast amounts of information, they can learn patterns and relationships that can be applied to new tasks. While these approaches have transformed fields ranging from language processing to weather forecasting and molecular biology, their potential in ecology remains largely untapped.

Dr Rachel Parkinson, Lecturer in Computational Ecotoxicology and Neuroethology at Queen Mary University of London and corresponding author of the paper, said:

"We have an extraordinary opportunity to use AI to bring together different kinds of ecological knowledge in ways that haven't previously been possible.

"Our vision is to develop models that can learn the relationships connecting organisms and their environments, uncover patterns we might otherwise miss, and ultimately help us better understand and predict how ecosystems respond to change."

From identifying species to understanding ecosystems
AI is already helping ecologists identify animals in photographs, recognise wildlife sounds and map species distributions. However, most existing tools are designed for specific tasks, rather than developing a broader understanding of ecological systems.

Ecological foundation models could go further, learning how species interact, how those relationships depend on environmental conditions, and how changes might affect wider ecological communities.

For example, a model could combine information about plants and pollinators, including their genetic relationships, physical characteristics, geographic distributions and known interactions, to predict which flowers an understudied pollinator might visit.

Similar approaches could help scientists discover undocumented species interactions, fill gaps in biodiversity knowledge and forecast how ecosystems might respond to climate change, habitat loss and other environmental pressures.

A roadmap for a new generation of AI
The Perspective paper, Towards ecologically meaningful foundation models outlines the scientific and technical steps needed to make such models possible, from integrating diverse ecological data and incorporating ecological theory to developing new AI architectures and evaluating the reliability of predictions.

The Perspective emerged from a multidisciplinary collaboration involving ecologists, conservation practitioners and AI researchers, including Schmidt AI in Science Fellows. The collaboration grew from a Schmidt Sciences-funded workshop at the University of Oxford in September 2025, co-organised by Dr Parkinson and Dr Amy Hinsley (Schmidt AI in Science Fellow, Oxford).

The authors emphasise that developing these systems will require more than technical innovation. Ecological data are often incomplete and biased towards well-studied species and regions, while many ecological relationships remain poorly understood. Models will need to recognise these limitations, quantify uncertainty and be developed responsibly alongside the communities that produce and use ecological knowledge.

Dr Parkinson added:
"Ecology poses some genuinely difficult challenges for AI because ecosystems are dynamic, interconnected and highly dependent on context. Developing models that can cope with this complexity will require new approaches in AI, grounded in ecological knowledge and theory.

"But the challenges aren't just technical. We need to ensure these models don't simply reinforce existing gaps and biases in our knowledge of the natural world. Our hope is that this roadmap will bring ecologists, conservation scientists and AI researchers together to build tools that are not only scientifically powerful, but also transparent, reliable and developed responsibly."

Read the paper:  Gardiner, R. J., Cerbone, H., et al. (2026). Towards ecologically meaningful foundation models. Nature Machine Intelligence. Perspective. Published 6 November 2026.

Sign up to hear more about Queen Mary's research.