The Cornell Lab of Ornithology's Merlin app — which identifies birds by their sound using machine learning — will soon upload its identifications directly into eBird, the Lab's citizen-science database holding more than 2 billion observation records. The eBird mobile app will gain the ability to send audio recordings and AI-generated identifications into that central repository, linking a mass-market acoustic tool to the world's largest structured bird-observation system for the first time.

The scale matters operationally. Merlin has passed 40 million downloads across 240 countries; between December 2025 and mid-2026, it added roughly 7 million installs. Britain alone accounts for close to 2 million users. Since 2021, Merlin has used machine learning to match birdsong to species, returning a species guess alongside an image — a capability that made structured bird data collection accessible to casual observers who lack formal training.

eBird, launched by Cornell Lab in 2002 and run by its Center for Avian Population Studies, serves as the backbone for global bird monitoring. Researchers, conservationists, and policymakers use its records to track population trends across continents. Routing Merlin identifications into eBird near real-time would significantly expand the geographic and temporal coverage of eBird's data — especially in regions where trained surveyors are scarce.

The Accuracy Question

The integration comes with a methodological caveat the scientific community has not resolved. The European Bird Census Council explicitly recommends against using Merlin in official breeding bird surveys. The EBCC has also formed a monitoring group to set common standards for automated acoustic bird identification across Europe — a structural response to the proliferation of AI tools and their inconsistent data quality.

The concern is not purely theoretical. Prof. Richard Gregory of the RSPB has noted publicly that Merlin identified his dachshund as a mallard — a concrete illustration of how ambient noise and non-bird sounds can confuse the classifier. As of July 2026, Merlin can identify 2,066 species, covering most birds in the US, Canada, and Europe, but its coverage becomes sparser in the tropics and parts of Asia where training data is limited.

Jessie Barry, who leads the Merlin project at Cornell Lab, and Marshall J. Iliff, eBird's Project Leader, are managing this integration. The Lab has not publicly specified what quality filters or metadata labels will be applied to Merlin-sourced records before they enter eBird's main dataset. This distinction — whether AI-generated records mix with human-verified data or remain in a separate analytical category — will determine how much scientific weight future studies can place on them.

The Conservation Stakes

The broader context here goes beyond technical logistics. The British Trust for Ornithology estimates the UK bird population has dropped by more than 70 million individuals over the past 50 years. Globally, bird population declines are among the most reliable early warning signs of ecosystem collapse. Denser monitoring data — even imperfect data — can expose local population crashes before they show up in surveys conducted years apart.

The tension crystallized by this integration is familiar to any field absorbing large-scale automated sensing: volume versus verifiability. Traditional eBird records carry observer accountability; a human submitter can be questioned, their record challenged, their site revisited. An ML-generated submission is reproducible only if the original audio is preserved and the model version is recorded. Whether Cornell Lab's system retains this provenance chain at scale is the engineering question that will determine whether this integration strengthens the scientific record or just inflates it.

A free Cornell Lab account already works across both Merlin and eBird, so the authentication infrastructure exists. The remaining decisions — how to flag AI records, whether to separate them into tiers, how to log provenance — are where the scientific value of this integration will actually be determined.