When Generative AI Can Predict Fish Communities – University of Montpellier

[Press Release] - Published on October 23, 2026 in the Press Room

A new study published on October 22 in the science journal Current Biology looks at a new way to predict which fish live together on coral reefs. The study was led by Noémie Coulon, a researcher at the Marbec Joint Research Unit, working with scientists from Lirmm. The researchers created a model inspired by the language models used in artificial intelligence (AI). They trained it using nearly 27,000 surveys of reef fish. Instead of predicting just one typical group of fish for a particular environment, the model can create several different groups that could live there. In about one-third of the surveys, the model produced clearly different kinds of fish communities. For example, some included fish linked to corals, while others included fish linked to large seaweeds called macroalgae. This method gives scientists a new way to understand how marine communities are organized and how different groups of species might live in the same environment.

A New Way to Predict Fish Communities

In the past, scientists usually used ecological models to predict where each species might live, one species at a time. For example, a model might estimate how likely it is to find a certain fish on a particular reef. This new study takes a different approach. It uses a model similar to those used for language to predict a whole community of fish living together.

The idea is a bit like writing a sentence. The model starts with information about the environment and gradually builds a fish community, just as a language model adds words to a sentence. As it adds each species, it considers the species it has already chosen and their biomass, meaning the total weight of those living organisms. This helps the model learn which species tend to appear together in the data.

Most importantly, the model does not give just one answer or one average fish community. Just as a language model can write several different sentences that all make sense, this model can create several different groups of fish that could live under the same environmental conditions.

The researchers trained the model using nearly 27,000 surveys of coral reefs. In the study, it also predicted fish communities more accurately than the other methods the researchers tested.

Exploring the Different Fish Communities That Could Live in the Same Environment

For each survey, the research team used the model to create 500 possible fish communities, all based on the same environmental conditions. When they studied these computer-generated communities, they found something interesting. In about one-third of the surveys, the communities could be separated into several clearly different groups. Each group had its own combination of fish species and biomass, or total weight of living organisms.

These groups were not simply different lists of fish. They also showed patterns in which species tended to live together. Some species were strongly linked to one particular type of community, while others appeared in several different types.

These patterns in the data fit with the idea that several different fish communities could exist in the same environment. However, the model cannot prove by itself that all these communities really exist in nature or that natural communities change from one type to another in this way. What it does offer is a new way for scientists to investigate this possibility using ecological data.

Why Does It Matter That Different Fish Communities Are Possible?

These different types of communities are not just slightly different versions of the same thing. They may be communities that work differently in nature.

For example, on Lord Howe Island, one possible reef community is linked to large seaweeds called macroalgae and areas where seagrass grows. Another possible community has more fish species that depend on hard, stony corals.

The difference is not just about which species are present. It can also change what matters most when trying to protect marine life. In this example, threatened species make up very different shares of the total fish biomass in the two possible communities. In the first community, threatened species account for as much as 10% of the biomass. In the second, they account for 3.6%.

By telling these different communities apart, scientists can get a clearer picture of the variety of fish communities that might exist. This also opens the door to studying the different jobs these communities do in their ecosystems.

A language-model method for predicting where species live reveals different possible stable states of ecological communities.