Millions of people worldwide, likely without realizing it, are capturing rare and valuable data that can help researchers assess trends in ecosystem health. Such data, however, are hidden in an ocean of comments on free online platforms like eBird and iNaturalist used by birdwatchers and other nature enthusiasts.
A new study led by the University of Michigan has shown that artificial intelligence can help researchers identify comments with pertinent data, then extract and categorize the information to build robust datasets. In this proof-of-concept study, the team found evidence of non-native plant species creeping into ecosystems along America's West Coast that might have been otherwise missed.
"The potential here is really big," said Hengxing Zou , the lead author of the study who performed the work as a postdoctoral scholar in the U-M Institute for Global Change Biology . "The approach is really versatile and we have a lot of data out there basically waiting to be mined."
The work was supported in part by federal funding from the U.S. National Science Foundation and the U.S. Department of Agriculture and is published in the Proceedings of the National Academy of Sciences.
In its study, the team was specifically interested in isolating comments that captured interactions between different species. While advances in technology have given researchers access to richer data on individuals and groups of a single species in the wild, interactions between different species are a different beast.
"In many cases, species interactions require direct observation and that's really hard to come by simply because you can't keep a lot of people in the field monitoring the same patch of an ecosystem for 24 hours a day, seven days a week," Zou said. "But it's very important data to have. It's connected to ecosystem function, ecosystem stability and a lot of the things we care about, especially in a changing global environment."
The team has now shown there is a new way to access data on interactions. Instead of a team of trained ecologists observing a location nonstop, they can use observations that have already been made by a community of hundreds of millions of nature lovers.
"This approach gives us something ecologists have long needed: a way to examine species interactions across broad spatial and temporal scales," said Kai Zhu , a senior author and associate professor in the U-M School for Environment and Sustainability, or SEAS . "AI can help turn scattered observations into ecological datasets that reveal how biological communities are responding to global change."
Putting the 'natural' in natural language processing
To prove the viability of its approach, the team used two case studies of species interactions. One focused on plant-pollinator interactions between the West Coast lady butterfly that pollinates plants along the west coast of the United States and Mexico using observations recorded on iNaturalist. The other looked at various interactions between Michigan birds found in the comments of observations recorded on eBird.
On these platforms, observations are uploaded with crucial metadata, such as when and where an observation was made. The next step for the team, then, was identifying which observations also had a comment that could provide insights on interactions.
For the plant-pollinator interaction, that was as straightforward as the West Coast lady landing on a flower. Bird interactions come in a wider array, including eating or being eaten by another species, cooperative behaviors like flocking and competitive behaviors like fighting over a spot at a bird feeder.
"I was surprised, and delighted, by the wide range of natural history observations that members of the public noted," said Brian Weeks , a senior author and associate professor in SEAS. "In addition to how exciting these observations are as a source for scientific data, a highlight of this for me was getting glimpses of the joy that observing birds can bring to people."
While humans could readily identify such interactions in comments, even when ecological vocabulary wasn't used, there were millions of observations to sift through. The team thought that AI tools designed to process natural or human language, like Claude and ChatGPT, would not only be able to quickly identify which observations had comments, but also which comments documented interactions.
Working with undergraduate researchers from ecology and computer science, the team crafted prompts that resulted in both Claude and ChatGPT extracting and classifying interactions with good accuracy and precision, Zou said.
To illustrate the time-saving afforded by the LLMs, the team first had ChatGPT isolate comments from eBird data that contained interactions. Then Olivia Stein, an undergraduate research assistant, took a random sampling of nearly 500 of those comments to classify which type of interaction they contained. The team then had Claude perform the same task. What took Stein eight hours, Claude completed in six minutes.
Zou understands that there are controversies surrounding the use of AI in research, but said this project is an example of AI being used to extend human effort, rather than replace it.
"I personally think it's kind of a rustic way of using AI in this booming age where people are using it to solve Millennium Prize math problems," Zou said. "It'd be almost impossible for humans to go through the millions of comments that are out there. We're using the large language models to go through and generate data sets, instead of analyzing them or replacing any of the creative work that scientists are doing."
While it is new technology that allows researchers to efficiently tap into an underutilized data source, the observations themselves reflect the founding spirit of ecology, Zou said.
"It does speak to this tradition of people going out there, observing nature and recording whatever they think is interesting," Zou said. "I think there's definitely a lot more information that could be recovered from these kinds of comments, in addition to species interactions."
Researchers from the University of California Santa Cruz, University of California Davis, Georgia Institute of Technology, American Bird Conservancy, Notre Dame University, Santa Fe Institute and Michigan State University also contributed to the research.
Additional funding for the study was provided by the David and Lucile Packard Foundation, Alfred P. Sloan Foundation, Institute for Global Change Biology at SEAS, U-M Undergraduate Research Opportunity Program, and MSU's Ecology, Evolution, and Behavior Program. Advanced Research Computing at U-M also provided computational resources and services.
Proceedings of the National Academy of Sciences
Large language models unlock the ecology of species interactions
8-Oct-2026