A new roadmap for ecological foundation models asks how AI might connect fragmented knowledge of the living world. The social question is equally important: whether those systems will strengthen field research and local stewardship, or make decisions appear more certain than the evidence allows
An environmental decision often begins with an uneven view of the place it concerns. One habitat may have years of surveys, another a few photographs, and a third extensive local knowledge that has never entered a standardised database. A computer can make those differences less visible by presenting every place on the same map. The resulting consistency may be useful, but it can also conceal how differently the conclusions were reached. As artificial intelligence enters conservation, the central question is therefore not only what it can predict. It is how societies will decide which predictions deserve influence.
On 6 November 2026, Queen Mary University of London announced a new Perspective on ecological foundation models, published in Nature Machine Intelligence. The team proposes connecting information such as genetic sequences, satellite images, wildlife recordings and species observations to investigate relationships within ecosystems. This is a research roadmap, not the release of a universally validated ecological forecasting system. Its ambition is to move beyond isolated identification tasks towards models that can help researchers examine interactions between organisms and their environments.
That ambition has consequences beyond the research laboratory. Conservation organisations, land managers, communities and public institutions all make choices with incomplete ecological evidence. A system that helps them combine knowledge could improve the questions they ask and direct scarce observation towards important gaps. A system treated as an authority could also concentrate influence among those able to build, interpret or purchase it. The difference will depend on scientific validation, but also on the institutions through which information becomes a reason to act.
The living world cannot be reduced to a collection of labels
Identifying a bird in a photograph and understanding the role of that bird in a changing habitat are different achievements. The first concerns a relatively bounded observation. The second involves relationships across time and space: food, competitors, breeding conditions and the effects of human activity, among other factors. Connecting these dimensions may help conservation move beyond a catalogue of presences. Yet a model that recognises regularities in recorded observations does not automatically explain the processes that produced them or predict how they will respond to an unfamiliar disturbance.
The university’s account illustrates the proposed approach with plants and pollinators, suggesting that diverse evidence might help predict interactions for an understudied pollinator. It also emphasises uncertainty, uneven data and collaboration with communities producing ecological knowledge. These are aims and acknowledged challenges, not demonstrated conservation outcomes. Read in that light, the roadmap is valuable because it makes the conditions of progress part of the research question. It does not ask users to assume that assembling more data will by itself produce an adequate understanding of nature.
An illustrative restoration project makes the distinction concrete. A model might suggest a relationship between a flowering plant and a local insect. That could guide a field visit or help a researcher identify a plausible comparison. It would not establish that introducing the plant will restore a damaged habitat. The timing of flowering, the surrounding landscape and other species might alter the result. This is a hypothetical example, not a finding reported by the new Perspective. It shows how a useful prediction can support investigation without becoming a complete prescription.
What has been recorded is not the same as what exists
The challenge is familiar to biodiversity data specialists. GBIF’s guidance on data access and quality explains that its records come from varied sources, including museum collections and citizen science, and that not every dataset is suitable for every use. It highlights unequal sampling across regions and taxonomic groups, as well as identification problems that may require expert knowledge. These are properties of the evidence available to any analytical method, not a defect unique to AI. More powerful modelling makes the need to recognise them more consequential.
Consider an illustrative landscape with frequent observations near accessible paths and few records elsewhere. A model could learn a pattern partly shaped by where people travel, rather than only by where species live. A blank area might indicate a genuine absence, inadequate observation or records stored in another form. These possibilities imply different decisions. The temptation to interpret a smooth map as a complete picture is especially strong when the system supplies precise-looking boundaries. Good presentation should make the limits of observation easier to see, rather than hide them behind visual confidence.
GBIF’s fit-for-purpose data guidance also distinguishes accuracy from precision and stresses that processing mitigates quality problems rather than eliminating them. A coordinate can be consistently formatted yet refer to the wrong location; an observation can be broadly correct but too imprecise for a local decision. This distinction matters when AI combines records collected at different scales. The question is not whether a dataset has passed a generic cleanliness check. It is whether the evidence can support the particular inference someone wants to make.
A prediction can redistribute attention and resources
Conservation involves choices about where to send people, which threats to investigate and how to allocate limited funding. An AI-generated assessment could influence all three. If it helps identify places that deserve closer observation, it can make fieldwork more strategic. If it becomes a shortcut for withholding attention from poorly documented places, it may deepen existing gaps. The same technology can therefore support discovery or create a self-reinforcing picture of importance, depending on how its outputs are used.
The possibility of feedback deserves particular care. A predicted hotspot might attract more surveys, yielding more records and strengthening its apparent significance in later analyses. A site assigned a lower priority might remain poorly observed. This is a plausible institutional mechanism to investigate, not a measured outcome of the proposed ecological models. Its importance is that data collection and decision-making are connected. The evidence entering the next generation of a system may partly reflect the priorities established by the previous one.
An organisation could respond by reserving some monitoring capacity for uncertain or underrepresented areas, rather than using every resource to confirm the strongest predictions. It could compare model recommendations with independent surveys and record why staff departed from them. Such practices would treat AI as one contributor to an adaptive research programme. They would also make disagreement informative: a field observation that challenges a model becomes an opportunity to improve understanding, rather than an inconvenience to be removed from a tidy dashboard.
Local knowledge needs a role beyond supplying data
The people who live and work in a landscape may notice changes that are difficult to capture in a standard observation form. They may know when a waterway behaves unusually, which seasonal pattern has shifted or why a previously recorded site is no longer accessible. This does not make every account equally applicable to every scientific question. It does mean that formal data and situated experience can reveal different aspects of the same place. A conservation system that recognises only information already formatted for computation risks excluding knowledge that could challenge its assumptions.
Participation should therefore extend beyond asking communities to contribute records. People affected by a proposed decision need a way to understand what the model is being used to infer and how that inference may change management. They should be able to question a description, explain missing context and see how disagreement is handled. A consultation that occurs after priorities have effectively been fixed has limited influence. Involving knowledge holders earlier can change the research question itself, including what counts as a useful or harmful outcome.
There are also decisions about access. Ecological information can be sensitive when it identifies a vulnerable species or a location that could be exploited. Knowledge may have been shared for one purpose without a clear understanding that it could support another. Research teams and institutions should make the terms of collection and reuse understandable and preserve appropriate routes for review. These are governance implications of building connected knowledge systems, not allegations about the authors of the new roadmap. Their relevance grows as information is combined and becomes easier to apply beyond its original setting.
Better forecasts do not decide what a society values
Even a highly reliable ecological forecast would leave choices open. A landscape can support biodiversity, livelihoods, cultural practices and other public needs, with tensions between them. An algorithm may help describe likely consequences under a set of assumptions. It cannot derive a universally acceptable balance between competing interests simply by improving its predictive accuracy. Presenting a preferred management option as the inevitable result of science can hide the value judgements that entered through the question, the indicators selected or the alternatives excluded.
The distinction should be visible in how an institution communicates a result. An estimate that a particular species may decline is different from a decision about land use, compensation or enforcement. The first requires evidence and scientific scrutiny; the second also requires legitimate authority and an accountable process. Keeping those stages connected but distinguishable can improve public discussion. It allows people to challenge an assumption without being portrayed as rejecting evidence, and to accept a forecast while debating the appropriate response.
Uncertainty is part of that discussion rather than a reason to postpone every action. Conservation decisions often must be made before complete knowledge is available. What matters is whether decision-makers understand how uncertain the estimate is, what the consequences of error might be and whether the proposed action can be revised. A system that helps compare those possibilities could be socially useful. A system that disguises uncertainty to sound decisive could weaken the very judgement it was introduced to support.
The measure of progress is stronger stewardship
The new ecological roadmap creates a timely opportunity to ask what success should look like. A useful programme would evaluate more than identification accuracy or performance on a convenient collection of records. It would examine whether models remain informative in places and periods different from those used in development, whether they reveal important gaps and whether their uncertainty helps practitioners choose appropriate next steps. Crucially, it would measure the value of the combined practice: field observation, local expertise, modelling and institutional response.
Resources also matter. A conservation organisation adopting AI must consider the continuing work of maintaining data, checking outputs, training staff and preserving access to the underlying evidence. A visually impressive prototype may be difficult to sustain if only its original developers can explain it. Conversely, a modest system that is well documented, locally usable and open to correction may become a more durable contribution. Those are implementation possibilities, not conclusions established by the Perspective. They identify the questions that should accompany the next stage of research.
AI could help people connect fragments of ecological knowledge that have long been difficult to study together. That is a substantial scientific opportunity. Its wider promise will be realised when it strengthens society’s ability to observe, question and care for the living world. The relevant achievement is not a machine that appears to know nature on our behalf. It is a better relationship between evidence and stewardship, in which unfamiliar findings prompt investigation, missing knowledge remains visible and the people responsible for a place retain a meaningful role in deciding its future.
Image: Honey bee on a flower by Kapteinkos, CC BY-SA 4.0, via Wikimedia Commons. Illustrative photograph of a plant–pollinator encounter, not a result or validation of the proposed models.