Reef fish do not necessarily form one predictable community for a given set of environmental conditions. Researchers at the University of Montpellier have trained an AI model to generate different plausible fish communities from the same environmental information. Instead of announcing one average result, the system can produce alternative combinations of species and their relative abundance. That sounds like a small change in modelling, but it could be important for conservation: an average community may hide very different possibilities for the living system that ecologists are trying to understand.
The University of Montpellier's report describes research published in Current Biology on 22 October 2026. Led by Noémie Coulon, the team used nearly 27,000 reef fish surveys to develop a model inspired by the way language systems construct sequences. The researchers emphasise an essential limitation. Generating several plausible communities does not prove that all of them exist in the same place in nature, or that a real reef would move between them. The finding supplies hypotheses and a new analytical method, not a detailed prediction of the fate of any individual reef.
AI Reef Fish Communities Are More Than Species Lists
Traditional models often ask whether a particular fish species is likely to be found in a location. That is a useful question, but it may treat each species separately and then combine the predictions into a single picture of a community. The problem is that fish do not exist independently of one another. They share habitat, encounter competitors and form relationships that may affect which combinations are plausible. Predicting every species in isolation can overlook some of those patterns.
The Montpellier method borrows an idea from language models, though it is not writing prose about the ocean. A language model can produce a sequence by considering what has already appeared. Here the system progressively assembles species and estimated biomass while considering environmental context and earlier choices in the same generated community. In everyday terms, the model is asking not just whether a species could live there, but what other species might plausibly be present alongside it.
Biomass matters because a list of species alone cannot describe the balance of a community. Two reefs might host similar fish yet differ greatly in the amount of living material represented by each kind. Those differences could affect ecological functions and the significance of conservation decisions. The model attempts to predict both occurrence and biomass. According to the researchers, it performed better on those measures than the comparison approaches they evaluated. That is a result within the study's testing framework, not proof of superiority over every conceivable ecological model.
One Environment, Several Possible Outcomes
The team generated 500 communities for each survey context in its analysis. In roughly one-third of the situations examined, the generated results formed more than one distinct group rather than clustering around a single typical answer. Some groups contained different associations of fish species, including patterns linked with coral habitats or macroalgae. The important phrase is 'generated results'. These are outputs of a model trained on observed surveys, not hundreds of separate future reefs that the scientists actually visited.
This can be understood through a familiar analogy. Imagine asking for a realistic description of a village given its size, climate and location. One plausible answer might include a particular combination of shops and public spaces; another might assemble a different but still coherent combination. Averaging them could describe a village that exists nowhere. Ecological communities are much more complicated than villages, but the problem of combining distinct patterns into a misleading average is similar. A model that can retain alternatives may be more informative than one producing only the middle of every distribution.
The university gives an example involving reef communities around Lord Howe Island. One generated configuration was linked more closely to macroalgae and seagrass, while another included more fish associated with stony corals. Their estimated representation of threatened species also differed. Such contrasts show why the alternative groups are not merely cosmetic variations. Different mixes could imply different ecological functions or conservation priorities. They do not, however, establish that one specific reef will change from one group to the other after a particular event.
The full scientific interpretation depends on whether those modelled groups correspond to meaningful states in nature. That is a difficult question. Environmental measurements cannot capture every local process, and different species may respond to unmeasured conditions or historical events. A model could create coherent statistical clusters without proving they are dynamically stable ecosystems. That is why the researchers treat their results as evidence consistent with alternative ecological regimes, rather than declaring that the computer has discovered all possible futures.
Why Predicting a Single Average Can Mislead
Conservation decisions often require an understanding of what might happen if conditions change. An average prediction can be useful for broad trends, but it may conceal distinct sets of organisms with different vulnerabilities. If some plausible communities are associated with coral and others with seaweed, those differences could change how an ecologist evaluates habitat condition or plans further surveys. The model opens a way to explore these possibilities, subject to validation against actual ecological change.
An average is not wrong simply because variation exists. It becomes less informative when the distribution contains separate clusters that do not resemble the mean. The researchers are effectively providing a way to investigate the shape of uncertainty. Rather than asking only 'what is the predicted community?', they can also ask 'how many different coherent answers does the model produce for this environment?' That second question helps identify places where a single headline number may be an inadequate description of what the research knows.
It is tempting to describe this as AI showing a reef's alternative futures. The phrase is memorable, but it must be kept within a clear limit. The model is trained on observations and generates statistically plausible communities under specified conditions. It does not simulate every ecological mechanism governing how one real reef changes over years. Forecasting a transition would require further work, including independent observations, an understanding of timing and evidence about what might cause the system to move between states.
The study also raises questions about generalisation. A model can learn patterns within a large collection of reef surveys, but environmental conditions differ around the world. Species that occur together in one region may never encounter each other elsewhere. Sampling methods, local habitats and the quality of environmental data matter. Further evaluation would be needed before treating a system trained on these records as a dependable guide to every marine ecosystem. A successful comparison within the original study is a starting point for scrutiny, not the endpoint.
The Researchers Borrowed an Idea From Language, Not Its Certainty
The similarity to language modelling lies in generating a structured sequence. In writing, the choice of the next word depends on earlier words. In the ecological application, adding a species depends partly on species already placed in the simulated community. That makes it possible to represent some of the relationships that simpler independent predictions might miss. The system does not actually understand fish society in a human sense. It learns patterns from the observations and the modelling choices made by its designers.
That may seem like a specialist distinction, but it matters whenever AI is presented as discovering a law of nature. Models can reveal meaningful statistical relationships and point to tests that scientists had not previously considered. They can also capture artefacts of how data were collected. The task for ecologists is to decide which patterns survive contact with independent evidence. That requires ecological expertise, careful validation and sometimes entirely new field observations.
The authors' work involves scientists from Montpellier's marine and computer-science research groups, bringing together knowledge of reef ecology and computational methods. The lead researcher's first-hand description explains the same core finding and its limits, including why alternative simulated communities may help examine ecosystems. The result was published as peer-reviewed research, which is meaningful, but peer review does not turn a statistical pattern into a guaranteed outcome for a reef. It means the work has entered the scientific record for examination, testing and possible refinement.
What Conservation Teams Could Eventually Gain
A model that flags multiple plausible community structures might help researchers decide where to gather more evidence. If the outputs suggest sharply different sets of fish under similar environmental measurements, scientists could ask whether an unmeasured factor explains the difference. The technique might also encourage closer study of habitat features associated with coral or macroalgal communities. Those are reasonable research uses, not claims that the tool already gives park managers an exact decision about which intervention to make.
Conservation involves more than counting fish. Habitat condition, fishing pressure, human activity and the interactions between species all matter. LiveAIWire's coverage of AI identifying hidden fishing vessels addresses one human pressure on marine environments, while AI mapping changes on land and water concerns a separate observation problem. The Montpellier study does not replace those forms of evidence. It offers a way to think about the living community itself, potentially adding nuance when other information is available.
There is a useful caution against overconfidence here. A numerical model often looks authoritative because it produces a clear answer. In ecology, a clear single answer may be less faithful to reality than a range of possibilities. Generative methods can make that uncertainty visible, but they can also create many impressive-looking results that are difficult to verify. The burden is on the researchers to demonstrate why the different generated groups are meaningful and where the method stops being reliable.
A reader does not have to understand the architecture of the model to grasp the consequence. The creatures inhabiting a reef form communities, and the same broad conditions may be associated with more than one plausible combination. Recognising that possibility could lead to better questions about biodiversity and conservation. Other marine AI research on orca detection shows how technology can help observe living systems; the Montpellier project asks a more abstract question about how their members fit together. Both rely on real observations, and both require care when interpreting an automated answer.
The most valuable contribution may be a change in how scientists ask the question. Instead of assuming that a reef has one expected fish community and treating all differences as noise, they can examine whether several coherent patterns deserve attention. The model cannot tell us what will happen to a particular reef next year. It can help expose the limits of treating a complex ecosystem as if it had only one sensible answer. For conservation science, acknowledging that uncertainty may be more useful than pretending it has disappeared.
About the Author
Stuart Kerr is Technology Correspondent at LiveAIWire, covering artificial intelligence, cybersecurity and the social impact of emerging technology. LiveAIWire is an independent, human-led technology publication using AI-assisted research, editorial production and original AI-assisted editorial illustrations under his direction.