COPENHAGEN, July 21 — In the quiet corridors of the Global Biodiversity Information Facility, the digital backbone of the world’s natural history is quietly confronting a 21st-century spectre: the contamination of citizen-science data by artificial intelligence. Against this backdrop, three of the largest repositories of species observations — iNaturalist, eBird, and GBIF itself — have each charted a different course to protect the integrity of the records that underpin conservation and research.
The stakes are high: if the data become unreliable, the scientific conclusions drawn from them risk being built on sand. GBIF is an international network funded by governments, and its secretariat is located in Copenhagen.
It acts as a central clearinghouse for biodiversity data aggregated from thousands of contributors worldwide. In many ways, it is the quiet librarian of planetary life — but that library is now being flooded with material of uncertain provenance. iNaturalist, operated jointly by the California Academy of Sciences and the National Geographic Society, relies on a machine learning identification tool and a community of volunteer verifiers. It formally banned AI-generated images in April 2023, and has since been removing them manually while developing automated classifiers to flag future fakes.
The decision was swift, but the backlog is significant. eBird, run by the Cornell Lab of Ornithology at Cornell University, takes a different approach. Its data are used in over 200 peer-reviewed publications each year.
The platform relies on human reviewers to assess observations, and has not yet reported widespread problems with AI contamination. Still, it created a ‘Photo Quality’ filter in 2021 to help flag suspicious images — a pre-emptive measure that now looks prescient.
Guidelines, not enforcement
GBIF itself released data quality guidelines in 2024 that recommend all records include metadata about AI editing. The phrase “recommend” is the operative word: the guidelines are voluntary, and compliance depends on the policies of contributing institutions. What is at stake is the long-term integrity of these databases.
Without consistent and transparent policies across all organisations, the scientific community risks building an increasingly unreliable picture of global biodiversity — precisely when that picture is most needed for conservation planning. For the wider picture, the European perspective is instructive.
Many of the oldest natural history collections in the world are now digitising their holdings, creating parallel digital archives. These institutional collections, unlike crowd-sourced platforms, are less vulnerable to AI contamination because they are curated by experts and physically verifiable. Yet they cannot match the scale and geographic breadth of citizen-science databases.
The two systems are complementary, but only if the citizen-science side remains trustworthy. The divergence in responses — iNaturalist’s ban, eBird’s cautious filter, GBIF’s soft guidelines — reflects a field still feeling its way.
In the corridors of Copenhagen, the conversation now is about standardisation. What is missing is a common framework that every platform can adopt, rather than a patchwork of local rules. The next step, many argue, is a binding code of practice for AI-generated content in biodiversity data.
Until then, the fight to keep the world’s biological archives clean will depend on the vigilance of volunteers, the algorithms of curators, and the willingness of governments to fund the kind of metadata infrastructure that makes transparency possible. The natural world deserves nothing less.
