SpeciesNet supports ecological estimates with some species missed

MegaDetector and SpeciesNet [1]

MegaDetector and SpeciesNet. [1]

MegaDetector 5a; SpeciesNet checkpoint not reported [1]

Google and MegaDetector contributors [1]

MegaDetector located animals. SpeciesNet classified image crops. Statistical occupancy models estimated ecological relationships from the resulting labels. [1]

MegaDetector 5a detections at confidence 0.15 or above were passed to SpeciesNet. Researchers applied regional species lists, taxonomic adjustments, sequence smoothing, and six combinations of classification thresholds and sequence cleaning. The SpeciesNet checkpoint is unspecified. [1]

Camera photographs became animal crops and species identifications, which researchers used to estimate occupancy and relationships with environmental variables. [1]

The evaluated AI workflow did not require routine human review of each image. Researchers designed the processing and ecological analysis. No automated conservation action is documented. [1]

Researchers processed images with MegaDetector and SpeciesNet, prepared six AI datasets, fitted occupancy models, and compared those models with models based on expert identifications. [1]

Researchers supplied regional filters and taxonomic adjustments and assessed the statistical output. Expert identifications provided the comparator. The authors recommend human review for some rare-species and biodiversity tasks. [1]

The study evaluates AI-generated species labels as inputs to ecological inference. Montana's expert comparison already included MegaDetector sorting. [1]

MegaDetector located animals and SpeciesNet supplied identifications for ecological models. Researchers configured and analyzed the workflow. The comparison tests analytical usefulness, not a measured improvement in conservation outcomes or net labor time. [1]

AIBD-0584-outcome-research [1]

Thornton, Morris, and colleagues compared ecological occupancy models derived from automated SpeciesNet identifications with models derived from expert identifications. The study used 638 cameras in Washington, 263 in Montana, and 326 in Guatemala. MegaDetector 5a located animals, and SpeciesNet classified the crops, with regional species lists, taxonomic adjustments, and sequence processing. Across six AI configurations, mean absolute differences in occupancy probabilities were 0.04-0.05 in Washington, 0.05-0.07 in Montana, and 0.03-0.06 in Guatemala. These comparisons cover species present in both datasets; they are not image-accuracy percentages. Some individual species differed substantially, with a maximum difference of 0.40 in Montana. All 94 mountain-goat detections in Montana were classified as domestic goats and omitted from the AI occupancy models. Montana's expert workflow already used MegaDetector for initial sorting. The study demonstrates automated ecological analysis with broadly similar results for shared species, while preserving a need for human review in some research tasks. It does not measure net labor savings, changed management decisions, or wildlife outcomes. Google developer Dan Morris coauthored the study. The author summary was first dated March 17, 2026; the full paper appeared on May 6. [1]

Improved Knowledge Discovery [1]

Reported research result with explicit limitations. [1]

Participating research organizations. Wider conservation benefits remain unmeasured. [1]

The participating ecological research teams. Benefits to monitoring organizations and conservation managers are prospective unless subsequent use is documented. [1]

The study does not count people who benefited. Cameras, images, species, and detection events are study units. [1]

Agreement varied by region and species. Shared-species comparisons do not cover species missing from the AI datasets. Rare or difficult species may require human review. [1]

The study does not establish how long the benefit persisted after the evaluation. [1]

The result concerns the intended use of AI to support wildlife-image analysis. [1]

Broad agreement coexisted with substantial errors for some species. All 94 Montana mountain-goat detections were labeled domestic goats and omitted from AI occupancy models. Some species appeared falsely in AI outputs. [1]

AIBD-0584-evaluation-thornton-2026 [1]

Comparative ecological analysis of the same regional image datasets using expert labels and six AI processing configurations. The comparator is not uniformly free of AI. [1]

The authors compared occupancy probabilities, detection probabilities, species-environment relationships, precision, and spatial projections from fitted models. [1]

Washington: 638 cameras and about 1.8 million images. Montana: 263 cameras and about 1.7 million images. Guatemala: 326 cameras and about 280,000 images. [1]

Mean absolute occupancy-probability differences were 0.04-0.05 in Washington, 0.05-0.07 in Montana, and 0.03-0.06 in Guatemala for species shared by the AI and expert datasets. [1]

Ecological models fitted to expert identifications of the same images; Montana experts already used MegaDetector for initial sorting. [1]

Mean occupancy estimates were broadly similar for shared species. Individual differences reached 0.40 in Montana, and some species were omitted or added by classification errors. [1]

The comparison demonstrates an automated route to useful ecological estimates for many shared species. It does not quantify net time saved or improved conservation decisions relative to the existing workflow. [1]

Comparisons of model estimates cover shared species. Rare-species omissions remain outside those comparisons. Regional filtering and sequence processing affect results. Montana's expert workflow already used AI. The exact SpeciesNet checkpoint, some regional inputs, and the complete supplement were unavailable in this review. [1]

The study includes 638 Washington cameras, 263 Montana cameras, and 326 Guatemala cameras. These are sampling units, not beneficiary counts or a denominator for conservation success. [1]

Zenodo version 1.0.1 contains Washington CSV inputs and scripts for all regions. Referenced Montana and Guatemala inputs are absent. The public Washington image subset has a different denominator. The complete supplement was inaccessible. [1] [2]

The study does not report a monetary benefit or net-cost estimate. [1]

Not reported [1]

Not reported [1]

Not reported [1]

Not reported [1]

Not reported [1]

Not reported [1]

Not reported [1]

Not reported [1]

Not reported [1]

Not reported [1]

Washington State University and wildlife research partners supplied regional data and analysis. Dan Morris of Google contributed to conception, analysis, and writing. Google developed SpeciesNet. [1]

Washington and Montana in the United States, and the Maya Biosphere Reserve in Guatemala. [1]

Not assessed

Camera datasets, MegaDetector, SpeciesNet, regional species information, and statistical analysis in R were required. A complete historical SpeciesNet checkpoint and all regional data inputs are not preserved in the inspected archive. [1] [2]

Specific cross-border transfers of study data or model outputs are not documented in the reviewed sources. [1]

Not assessed

The system must retain the species that matter to the intended ecological question. The exact SpeciesNet checkpoint and processing settings must be recoverable. Regional input data and analysis scripts must be available, with appropriate protections for sensitive wildlife locations. [1]

Not assessed

Not assessed

The full article and versioned code-and-data archive are public. The inspected archive does not contain all regional inputs, and the complete supplement was inaccessible. [1] [2]

The study does not document a subsequent institutional or conservation-management decision. [1]

Not reported [1]

Not reported [1]

Not reported [1]

Two retained sources have stable identifiers beneath this record.

Thornton et al., sections 2.1-2.6, Table 1, sections 3.1 and 4, and disclosures. Zenodo version 1.0.1 supplies the inspected scripts and partial data. [1] [2]

The retained sources are in English. [1] [2]

The original peer-reviewed study includes Dan Morris, a Google employee and contributor to SpeciesNet. It is research with developer participation. The code-and-data archive belongs to the same work and is not independent confirmation. [1] [2]

The paper supplies the comparative results. Its archive and author summary belong to the same work and do not provide independent confirmation. [1] [2]

The full paper and archive version 1.0.1 were inspected. Some regional inputs are absent, and the full supplementary document could not be retrieved. [1] [2]

Researchers used MegaDetector and SpeciesNet to produce ecological estimates similar to expert analysis for many species, while some species were missed or misclassified. [1]

The original study supports the comparative result. The second source adds methods or reproducibility information. [1]

The original comparison supports useful ecological estimates for many shared species and documents exceptions. Developer participation, unavailable materials, and the absence of measured downstream outcomes remain explicit. [1]

The exact SpeciesNet checkpoint, missing regional CSV inputs, complete supplementary document, subsequent adoption, actual labor savings, and conservation decisions remain unresolved. [1]

The documented research benefit for the participating organizations. [1]

Organizational depth 2: a demonstrated analytical capability for shared species. Societal depth is unassessed because a conservation outcome is not documented.

The study demonstrates an automated analytical capability with broadly similar occupancy estimates for shared species. Rare-species omissions and setting-specific differences limit its use. No changed conservation decision or wildlife outcome is documented.

The record distinguishes demonstrated analytical results from estimated or unmeasured downstream benefits. [1]

AIBD depth scale and the Trustible benefit taxonomy in the current local catalog, reviewed October 25, 2026.

AIBD reviewed this assessment on October 25, 2026.

Expert review may remain preferable for rare species and biodiversity inventories. Targeted human review may resolve problematic classifications without reviewing every image. [1]

The study includes developer participation. The comparative research design and limitations remain explicit.

Comparative research study. [1]

The other new conservation record concerns a separate evaluation that also uses MegaDetector. It is not a replication of this result.

Environment and conservation [1]

SpeciesNet identifies wildlife in images. The study tested whether automated identifications could support ecological occupancy models without routine human verification of each image. [1]

Images were collected in 2016-2021 across the three regions. An author account was first dated March 17, 2026; the full study appeared May 6. The original March text was not recovered. [1] [2]

Research evaluation using field-collected camera images. [1]

The record has a content fingerprint that changes when its source data change.

The study DOI is 10.1111/1365-2664.70370. The record links both archive versions and the public Washington image subset. [1] [2]

Not assessed

Not assessed

The record and its source assessment were added on October 25, 2026.