Ecologists have records of species sightings, satellite images, wildlife sounds, genetic sequences and measurements of the environment. Each can answer a different question. A new research perspective asks whether future AI systems could bring those kinds of information together to reveal relationships that are difficult to see in isolation.
The idea is called an ecological foundation model. It is a research roadmap, not a finished system that can already explain every ecosystem. The authors argue that useful models would need to account for missing data, ecological context and uncertainty, as well as involve the people who collect and use ecological knowledge.
The perspective was published in Nature Machine Intelligence on November 6, 2026. The University of Oxford’s account explains that it grew from a workshop involving ecologists, conservation practitioners and AI researchers. The paper sets out what the authors think would be needed to make such models scientifically meaningful.
Why is ecological information so hard to combine?
A satellite can show a forest canopy changing over a large area. A field researcher can identify a plant or insect at a specific place. A microphone can record animal calls through the night. A genetic sample may indicate that a species was present in a stream even when nobody saw it.
These observations are valuable precisely because they see different aspects of nature. They also differ in resolution, timing and reliability. A satellite image cannot automatically tell a researcher which insect visited a flower. A single field observation cannot establish what happens across a whole continent.
Ecological relationships are also conditional. Two species may coexist in one region and never encounter each other in another. A plant may flower at a different time after an unusually warm season. A model that ignores place, season and uncertainty can produce a plausible-looking relationship that does not hold in the field.
The proposed models would try to learn across these data types rather than treating each as a completely separate problem. That ambition is attractive, but combining data is not the same as understanding why a relationship exists.
What is a foundation model?
In AI, a foundation model is generally a system trained on a broad body of data that can be adapted to more than one task. Language models are one familiar example, but the idea is not limited to text. A model might learn patterns in images, sound, measurements or a combination of sources.
An ecological foundation model, as envisioned in the perspective, would aim to learn representations useful across a range of ecological questions. Instead of building a separate system from scratch for every species or habitat, researchers could potentially adapt a broader model to a particular task.
That does not make the model a universal scientific authority. Its usefulness would depend on what it was trained on, which questions it is asked, whether its assumptions are reasonable and how its predictions perform against independent observations.
A pollinator example
The Oxford account offers an example involving plants and pollinators. Suppose scientists have information about the geographic ranges, physical characteristics and known interactions of some insects and flowers. A model might use these data to suggest which flowers an understudied pollinator could visit.
Such a suggestion could help researchers decide where to look. It would not, by itself, prove that the insect visits a particular flower or pollinates it effectively. Visiting a flower and transferring enough pollen to influence reproduction are different observations.
This difference is central to the potential value of AI in science. A model can help generate a hypothesis or prioritize limited fieldwork. Field observation and experiments are still needed to test the prediction and understand its biological significance.
The problem of missing observations
Nature is not sampled evenly. Some countries, habitats and charismatic animals receive far more research attention than others. A species that appears frequently in a dataset may be common, easy to observe, or simply well studied. Those possibilities are not interchangeable.
If a model treats the absence of a record as proof that an interaction never occurred, it may reproduce the gaps in the data. If it learns mostly from one climate or region, it may perform poorly in another. Combining several datasets can help, but it can also combine their biases.
Researchers therefore need to document how data were collected and distinguish unknown from absent wherever possible. They also need tests that challenge the model in new places, seasons or species, rather than only asking it to predict records that closely resemble its training examples.
Prediction is useful, but explanation is a different task
An AI system may predict that a species is likely to occur in a location because it found a pattern in temperature, vegetation and earlier sightings. That can be helpful. Yet a successful prediction does not automatically show which variable caused the species to be there.
The distinction matters when conservation decisions involve interventions. If a model predicts a decline, managers need to know whether protecting habitat, changing water use or addressing another pressure could make a difference. A pattern learned from past data may also fail when a new environmental condition appears.
The authors argue for models grounded in ecological knowledge and theory. In practice, that means predictions should be evaluated against what is known about organisms and environments, while remaining open to results that challenge existing assumptions.
Who gets to use the data?
Ecological data are often collected by researchers, local communities, conservation organizations and public agencies. Their knowledge and interests are not simply an input to a computer system. Decisions about access, interpretation and application can affect the people and places represented.
The perspective emphasizes collaboration with those who generate and use the knowledge. Practical questions include whether data can be shared responsibly, how sensitive locations are protected and how a model communicates uncertainty to someone making a decision.
A map showing an endangered species’ likely location, for example, could help conservation work but may also require careful handling. The right approach depends on the data and the people involved; technical accuracy alone is not the whole question.
What would count as progress?
A useful next step would be a model that makes a specific, testable prediction and performs well on independent observations. Researchers could compare it with simpler statistical approaches and existing ecological methods, including cases where those alternatives may be more transparent or require less data.
It would also matter to know where the model fails. A system that gives an uncertainty estimate and identifies unfamiliar conditions may be more useful than one that produces confident answers everywhere. Reproducible evaluation across regions and species would help reveal whether the approach generalizes.
For now, the new paper is best understood as a proposal for a research program. Its interdisciplinary appeal is real: AI may help connect observations across scales. The work of turning those connections into reliable ecological knowledge still depends on careful measurement, theory and testing.
Explore the book: How to Do Hard Things: The Discomfort Trap
Be First to Comment