AI Struggles to Identify Colombian Wildlife

A study found GPT-5.2 identified only 9.5% of vertebrate photographs correctly at species level.

Updated on Nov. 18, 2026 in Environmental

Bold flat-color editorial illustration of an unlabeled camera trap on a mossy tree trunk among broad tropical leaves, evoking AI's struggle with wildlife identification.
Researchers found GPT-5.2 correctly identified only 9.5% of 10,000 Colombian vertebrate photographs at species level, raising accuracy concerns for conservation work. AI Illustration. Upload story photo >

Researchers tested GPT-5.2 on 10,000 iNaturalist photographs of vertebrates recorded in Colombia, published in the Biodiversity Data Journal. The model correctly identified 9.5% of the images at species level, with the model required to name a family, genus and species for every image.

Why it matters

The results raise questions about whether general-purpose AI is ready to support conservation work that depends on accurate species identification. Misidentifications, such as a marsupial mistaken for another animal, could affect ecological monitoring.

The benchmark used 2,000 images per group across amphibians, birds, fishes, mammals and reptiles, with mammals and reptiles reaching 17%-19% species-level matches while amphibians, birds and fishes fell below 5%. The extraction does not specify whether results would change with different prompts or models.

The players

GPT-5.2

A general-purpose large language model from OpenAI whose image recognition was tested in this study.

iNaturalist

A citizen science platform where users upload nature photographs, which provided the research-grade images used in the benchmark.

The details

The model was told only that each animal was seen in Colombia and was required to answer for every image without the option to decline. Species with more publicly available photographs were generally identified more accurately, while recently described species were harder to recognize.

Timeline

  1. The Porcupine Coendou vossi was recognized as a distinct species in 2025.

  2. The Phys.org article reporting the study was published on November 18, 2026.

Big Picture

The study measures general-purpose AI against iNaturalist's research-grade species labels, which are the community-verified standard used by many biodiversity datasets. The results suggest that such labels remain difficult for current general models to reproduce at species level.

Researchers and volunteers who rely on AI tools to sort wildlife photographs may need to verify identifications rather than accept them automatically. Conservation projects that use camera traps should expect that species-level labels from general models could be unreliable for many animal groups.

The takeaway

General-purpose AI can be useful for broad tasks, but species-level wildlife identification still requires human expert review before results inform conservation decisions.

Further reading

Explore more Environmental coverage to see how AI tools are being tested against real-world ecological data.

Source note: This article includes information reported by Phys.