AI finds the ecological relationships hidden in birdwatchers’ notes

“Mixed with bohemians.” In an eBird comment examined by researchers, that meant Bohemian Waxwings. Three casual words recorded which birds were together, adding a relationship to a record of a sighting.[1]

A PNAS Perspective published on 8 October explores how language models can extract such interactions from bird and butterfly observations.[1]

The ecological distinction is substantial. Imagine two surveys that list the same birds and plants. In one, a bird feeds on a plant’s fruit; in the other, the two merely occupy the same patch of ground. Identical species lists can describe different ecological activity. Interaction datasets preserve those connections: who eats whom, visits which flower or acts as a host. Global Biotic Interactions, or GloBI, already makes records of these relationships searchable across openly shared datasets.[2]

The information between the names

A fine-tuned GPT-4o model screened 632,169 Michigan comments from January 2019 to January 2024, flagging 29,345 as containing interactions. In a separate check of 485 flagged comments, Claude Sonnet 4 agreed with human interaction labels in 395 cases: 81.4%.[1]

Results varied by setting. Among 143 interaction-containing comments in a largely Chinese-language Taipei sample, the model correctly identified and classified 88, or 61.5%. The authors linked errors partly to difficulty recognising Chinese bird names.[1]

Useful extraction therefore depends on preserving meaning. “Together” might describe proximity; an account of one bird driving another away records an action. Turning both into the same generic connection would discard the detail that makes the observation interesting. A database needs to retain the kind of interaction and its original evidence so a later researcher can ask a different question of it.[2]

In the butterfly case, comments about the West Coast Lady yielded 20 plant taxa absent from the comparison list in GloBI. Such records can direct attention towards relationships an existing index has missed.[1]

For a gardener or an ecologist, this opens a different question: which plants are being used, and by whom? The visitors to a patch of flowers could change while every species still appears somewhere in the local records. Relationship data gives that change something concrete to investigate.[2]

A field note with a second life

For someone watching wildlife, the contribution remains refreshingly ordinary. eBird’s documentation guidance encourages notes made in the field, including behaviour, alongside identification details and photographs or recordings where possible. Written descriptions remain useful when no camera is available. The platform already encourages documentation beyond the unusual sightings that trigger its automatic requests for supporting evidence.[3]

A useful description names the participants and says what happened. In a hypothetical field note, “bird at flowers” leaves the action ambiguous; naming the bird and plant and describing it taking nectar gives a later reader something more definite to work with. Uncertainty belongs in the note too: an unidentified plant should remain unidentified until someone has evidence to name it.[3]

The lasting contribution begins with that act of attention. A few words written while an encounter is fresh can preserve an event that a photograph, a count or a list of names leaves unexplained.[3]

Report an error or suggest a correction