Birdwatchers who log sightings on eBird often add a quick note. “Mixed with bohemians,” one wrote, meaning a flock that included Bohemian Waxwings. Notes like that record which birds flock or feed together, or gang up on a predator. Ecologists struggle to collect that kind of detail because nobody can watch one patch of ground around the clock.
A University of Michigan team pointed AI at those notes. A GPT-4o model, tuned on 544 hand-sorted comments from Texas, screened more than 18 million Michigan sightings logged from 2019 to early 2024. Of the 632,169 that came with comments, it flagged 29,345 that may describe birds interacting.
The clever bit
The clever bit is the side-by-side test. Undergraduate researcher Olivia Stein hand-sorted 485 of the flagged notes by type of encounter, which took her about eight hours. Claude, given the same rulebook and no practice examples, matched her on 395 of them in minutes. When birds ganged up on a possible predator, it caught every case.
Then the team tried a butterfly, the West Coast lady. Of 8,999 sightings in a global biodiversity database, 996 had comments in English or Spanish. The AI found 299 describing interactions, including 100 flower visits, and those visits named plants missing from a major database’s list for the butterfly. Many were not native to its range.
I was surprised, and delighted, by the wide range of natural history observations that members of the public noted.
— Brian Weeks, associate professor, University of Michigan School for Environment and Sustainability, via earth.com
Hooray
Hooray: quick notes from people who simply love watching birds and butterflies are becoming wildlife data scientists can use, and AI handles the first read in minutes.
What comes next
People stay in the loop to fix names and catch misses. Lead author Heng-Xing Zou says there is “definitely a lot more information that could be recovered from these kinds of comments,” and the team’s code and data are public for others to try.