New study maps where AI species identification fails

General purpose AI use in species identification. Can it be useful for cosnervation?

Trap camera images
Getting your Trinity Audio player ready...

A small marsupial photographed by a camera trap in Colombia was identified by an AI model as a mammal whose known range does not include the country, even though the model had been told the animal was seen in Colombia. A new study in the Biodiversity Data Journal uses this kind of error as a starting point to ask whether general-purpose AI is ready for conservation work. 

The marsupial, Metachirus myosuros, was recorded in the Morrocoy protected areas in Montes de María, where the conservation organisation Sembrandopaz has been monitoring wildlife. The AI proposed Bassariscus sumichrasti, a species whose known range does not include Colombia. And the expert Dr Juan Fernando Díaz Nieto of EAFIT, confirmed the correct identification to be Metachirus myosuros.

Metachirus myosuros
Camera-trap image recorded during biodiversity monitoring in the Morrocoy reserves in May 2026. A general-purpose multimodal AI system proposed Bassariscus sumichrasti despite receiving the Colombian locality and a description of the animal’s locomotion, posture, facial structure, tail proportions, and possible colour pattern. Expert review identified the animal as Metachirus myosuros. Dr Juan Fernando Díaz Nieto, Universidad EAFIT supported the identification. Photo credit to Fundación Sembrandopaz.

Two other cases from the same fieldwork showed different problems. The armadillo Cabassous centralis has been known to science for decades, yet few verified photographs of it are publicly available.

Armadillo Cabassous centralis
Armadillo Cabassous centralis. Credit to Fundación Sembrandopaz.

The porcupine Coendou vossi, recognised as a distinct species only in 2025, was given an older, superseded name that still circulates in public resources.

Porcupine Coendou vossi
Porcupine Coendou vossi. Credit to Fundación Sembrandopaz.

“If a rare animal is confused with a more familiar one, its presence might go unnoticed,” said study author Gustavo Nicolas Paez Salamanca.

If AI assigns an incorrect species name or suggests an animal outside its natural range, that error could enter biodiversity records and affect how researchers interpret species distributions.

Testing 10,000 photographs

To test identification accuracy in a systemic way, the study evaluated OpenAI’s GPT-5.2 on 10,000 research-grade iNaturalist photographs of vertebrate species recorded in Colombia: amphibians, birds, fishes, mammals and reptiles, with 2,000 images per group. The model was told only that the animal was seen in Colombia and was required to name a family, genus and species for every image, with no option to decline.

Under a benchmark requiring species-level predictions, GPT-5.2 correctly identified 9.5% of these images, or roughly 1 in 10.

Performance varied widely by group, with exact species-level match rates of about 17-19% for mammals and reptiles and under 5% for amphibians, birds and fishes.

– said Paez Salamanca

It turns out, the errors were not random. Species with more publicly available photographs were generally identified more accurately, while recently described species were harder to recognise. Tests with other AI systems showed similar general patterns.

“This suggests that improving wildlife identification is not simply a matter of choosing a different AI application,” Paez Salamanca said. “The challenges are also connected to how much information exists about each species and how that information changes as scientific knowledge advances.”

Implementing AI could present a problem that goes well beyond species identification. Musters and de Snoo (2026) make a similar point in a recent paper on predicting ecological change with deep learning: conservation AI may need an entirely different kind of data, not just more of it. See our earlier blog here.

A role for AI, with safeguards

infographic
Visual poster presenting the key takeaways from the study. Infographic generated with ChatGPT-5.5 and provided by Gustavo Nicolas Paez Salamanca. The photos in the poster are from the trap cameras by Fundación Sembrandopaz.

The study does not argue for abandoning AI. It proposes using AI as a first step, with candidate identifications checked against current classifications and known ranges. Uncertain or high-stakes records should go to specialists, and AI tools should be allowed to say when an image can’t be reliably identified.

AI can help researchers process large numbers of wildlife photograph. But its greatest value for conservation comes when automated predictions are combined with reliable biodiversity information and scientific expertise.

“Although the research focused on Colombian animals, the challenges it identifies are relevant to biodiversity monitoring in other regions as well,” he concluded.

Original source:

Paez Salamanca GN (2026) Is conservation AI ready? Quantifying the challenges of automated species identification in Colombia. Biodiversity Data Journal 14: e208137. https://doi.org/10.3897/BDJ.14.e208137