Scientists publish roadmap for AI that studies ecosystems, not just individual species

Identifying an animal is one task; predicting how its relationships change with its environment is another. A Perspective published November 6, 2026, in Nature Machine Intelligence sets out a roadmap for AI that could tackle the second problem. The proposal connects diverse ecological data with ecological theory—not a newly released model or demonstrated forecasting breakthrough.

The paper, “Towards ecologically meaningful foundation models,” is led by Rachel Parkinson of Queen Mary University of London. Its authors propose ecological foundation models, or ecoFMs: systems trained on large collections of different kinds of ecological information, with knowledge that could be applied across multiple research questions rather than one narrow task.

From recognizing wildlife to predicting relationships

AI already helps ecologists identify animals in photographs, recognize wildlife sounds and map where species occur. Those tools largely address specific tasks. The proposed alternative would learn relationships among organisms and their surroundings, including how interactions between species depend on environmental conditions and how changes could affect a wider community.

The authors illustrate the idea with plants and pollinators. A model could combine genetic relationships, physical traits, geographic distributions and known interactions to predict which flowers a poorly studied pollinator might visit. That is a proposed use, not a reported successful experiment. The ambition is to infer missing relationships, rather than simply attach names to observations.

The obstacle is not a simple shortage of information. Ecology draws on genetic sequences, satellite images, wildlife recordings, species observations and decades of research. These sources describe nature in different forms and at different scales. The paper argues that fragmented data and theory leave existing models struggling to work beyond particular systems or scales.

Ecological theory belongs inside the model

The roadmap does not treat combining datasets as sufficient. The authors want models informed by ecological theory, linking observed patterns to the processes behind them. Their intended payoff is knowledge that transfers across ecological specialties and supports forecasting, rather than a system that works only in the setting where it learned.

The development agenda

• Data infrastructure: bring ecological information in different formats, resolutions and contexts into a common framework.
• Model design: develop AI architectures that can combine those inputs while incorporating ecological knowledge and theory.
• Evaluation and governance: assess prediction reliability and develop systems responsibly with ecological practitioners and other relevant communities.

The collaboration grew from a September 2025 workshop at the University of Oxford, co-organized by Parkinson and Amy Hinsley. It brought together ecologists, conservation practitioners and AI researchers. That mix reflects the roadmap's stated requirements: advances in machine learning alongside ecological knowledge and participation from people who produce and use it.

Missing knowledge can become model bias

In Queen Mary's account published by Phys.org, the authors warn that ecological data are incomplete and biased toward well-studied species and regions. Many relationships also remain poorly understood. A model intended to fill gaps could instead reinforce them, Parkinson cautions. The authors say systems must quantify uncertainty and recognize the limits of their underlying knowledge.

That challenge extends beyond data coverage. Parkinson describes ecosystems as dynamic, interconnected and highly dependent on context. The research agenda therefore asks AI to handle both varied inputs and relationships that change with conditions. Forecasting responses to climate change, habitat loss and other pressures is among the proposed applications; establishing reliable predictions is part of the work still ahead.