On 6 November 2026, a team led by Dr Rachel Parkinson of Queen Mary University of London published a Perspective in Nature Machine Intelligence titled “Towards ecologically meaningful foundation models” (DOI 10.1038/s42256-026-01317-6). It sets out a roadmap for building “ecological foundation models” (ecoFMs): AI models trained on many kinds of ecological data at once, rather than on a single narrow task.
What happened
According to Queen Mary’s release, Parkinson is lead and corresponding author, and the citation is given as Gardiner, R. J., Cerbone, H., et al. The release says the collaboration grew out of a Schmidt Sciences-funded workshop held at the University of Oxford in September 2025, co-organised by Parkinson and Dr Amy Hinsley, a Schmidt AI in Science Fellow at Oxford, and that the author team included Schmidt AI in Science Fellows. Phys.org carried the same story.
According to the Queen Mary release, the paper argues that most AI used in ecology today is built for narrow tasks such as identifying species in photographs or recognising wildlife sounds. Foundation models, the authors propose, could instead learn from varied sources, including genetic data, satellite imagery, wildlife recordings and species observations, and apply that knowledge to many ecological questions. The release gives an example: predicting which flowers an understudied pollinator might visit, using genetic relationships, physical traits, geographic ranges and known interactions.
The roadmap covers four areas, according to the release: integrating diverse data, building in ecological theory, designing new AI architectures, and evaluating how reliable predictions are. Potential uses described in the release include uncovering undocumented species interactions, filling biodiversity knowledge gaps, and forecasting how ecosystems respond to climate change and habitat loss.
The caveats the authors raise
Per the release and Phys.org, ecological data are incomplete and skewed toward well-studied species and places, so models must quantify uncertainty and be developed responsibly with the communities that produce and use ecological knowledge. Phys.org and the release quote Parkinson: “We need to ensure these models don’t simply reinforce existing gaps and biases in our knowledge of the natural world.” The Queen Mary release also quotes her saying ecology poses difficult challenges for AI because ecosystems are dynamic and interconnected.
This is a Perspective: a proposal and research agenda, not a report of a deployed or validated model. As reported, the roadmap describes what such models could do, not results that have been demonstrated.
What it means for researchers, universities and research offices
Researchers. The roadmap signals that evaluation and uncertainty quantification are treated as core design requirements for AI in ecology, not afterthoughts. Groups that hold ecological datasets may find their data, and its documented gaps and biases, becoming inputs to cross-disciplinary model building.
Universities and research offices. This is an editorial interpretation rather than something the paper states. Models trained across genetic, imagery, acoustic and observation data draw on many data owners and licences, so research data management, provenance records and clear reuse terms become practical prerequisites. See CASRAI’s coverage of research data management for the underlying practice. The Schmidt Sciences-funded workshop origin is also a reminder that funder-backed fellowship networks can seed multi-institution collaborations that later need agreements on data sharing and authorship.
Because the proposed models would depend on data from communities and regions that are underrepresented, offices supporting this work should expect questions about responsible and equitable data use alongside the technical ones.
Sources
- “Towards ecologically meaningful foundation models,” Perspective, Nature Machine Intelligence, 6 November 2026, https://www.nature.com/articles/s42256-026-01317-6 (DOI 10.1038/s42256-026-01317-6).
- Queen Mary University of London, “Can AI learn to understand nature? Scientists set out a roadmap for the future of ecology,” 6 November 2026, qmul.ac.uk. Read 12 November 2026.
- Phys.org, 6 November 2026, phys.org. Read 12 November 2026.
