OXFORD, July 21 — A modeling study estimates that without proactive labeling of all AI-generated content, the contamination rate of synthetic images in biodiversity databases could double every two years. The figure comes as database managers race to deploy detection tools, policy changes and community guidelines against a rising tide of machine‑generated wildlife photos. The timeline of countermeasures is now well documented.
The citizen‑science platform iNaturalist banned AI‑generated images outright in April 2023 and began using a custom‑trained ResNet‑152 classifier to detect synthetic uploads. By early 2024 that detector had reached an accuracy rate on known fakes. eBird, the massive bird‑observation database, implemented a “photo quality” filter as early as 2021 that flags images for human review; in 2024 it added a metadata check that looks for signatures left by AI editing software.
In February 2024, the Global Biodiversity Information Facility updated its data quality guidelines to require publishers to declare if an image has been AI‑modified. The Cornell Lab also released a public tool, the “Bird Photo ID Checker,” in June 2024 that helps users verify whether an image matches known species morphology.
Reactive Measures and a Call for Standards
The record shows these steps are largely reactive. The research, which modeled the trajectory of contamination, underscores a core gap: no system yet mandates proactive, universal labeling of every AI‑generated image at the point of upload. The consequence, the model suggests, is a compounding problem — without such labeling the rate of fake images infiltrating real‑world datasets could accelerate on a short cycle.
Recognizing the need for shared detection standards, a loose coalition formed in 2023. That group, which includes participants from several platforms and research institutions, has been exploring cross‑platform collaborations and harmonized detection protocols.
The group is still in the phase of comparing technical approaches rather than implementing uniform rules.
What to Watch Next
The future of database integrity in biology will likely turn on two questions. First, whether the detectors now deployed — such as iNaturalist’s ResNet‑152 classifier or eBird’s combined image‑quality and metadata filters — can keep pace with rapidly improving generative models. The accuracy figure from early 2024 is already a snapshot in time; performance against newer synthetic images remains unverified in public benchmarks.
Second, whether the coalition of platforms and researchers can evolve into a permanent governance structure that mandates proactive labeling. The model provides a stark timeline: every two years of inaction doubles the rate of contamination.
The group, now three years old, has yet to produce a binding standard, but its existence signals that the major repositories recognize the scale of the threat. For now, the record shows a patchwork of tools and policies. The innovation is real: a custom neural network checking every upload, a metadata scanner identifying editing signatures, a public tool helping birders verify a photo’s credibility.
The gap remains in the absence of a system‑wide, proactive label that tells the user — and every downstream database — that an image is entirely synthetic. The model suggests that gap will not close on its own.
