AI Confuses Wild Animals: Study Reveals Weaknesses in Automatic Species Identification

Edited by: Svitlana Velhush

In the nature reserves of Colombia, a camera trap captured a small marsupial. Having received the photo and information about the area, the artificial intelligence confidently identified it as a species that does not occur in the country. Experts, however, determined the actual inhabitant — Metachirus myosuros. This case became the starting point for a new study showing how far general-purpose AI models are from reliable work with wildlife.

The researcher Gustavo Nicolás Páez Salamanca tested the GPT-5.2 model on ten thousand photographs of vertebrates from Colombia uploaded to iNaturalist. The model was supposed to name the family, genus and species for each image, knowing only the country of the photograph and having no right to decline to answer. The exact result at the species level was only 9,5 percent — roughly one in ten photographs.

Accuracy varied greatly across animal groups. In mammals and reptiles it reached 17–19 percent, while in amphibians, birds and fish it dropped below five. Errors occurred more often with species for which there are few public photographs, and with those described by science recently, for example Coendou vossi, which was distinguished as a separate species only in 2025.

Such mistakes are not merely a technical trifle. If a rare animal is confused with a more familiar one, its presence may be missed. An incorrect name or an indication of a species outside its natural range can distort data on the distribution of animals and affect plans for the protection of territories.

The study does not call for abandoning AI. On the contrary, the author proposes using it as a first filter: checking automatic suggestions against current classifications and known ranges, and passing questionable or important records on to specialists. Models should be allowed to acknowledge uncertainty instead of being required to answer.

Although the work focuses on Colombian fauna, the problems it identifies are characteristic of biodiversity monitoring in other regions. The reliability of automatic identification depends directly on the volume and quality of available data on species.

For AI to genuinely help conserve nature, its predictions must be combined with verified scientific information and the expertise of field biologists.

2 Views

Sources

  • When AI misidentifies wildlife: New study maps where species identification fails

  • phys.org

Comments

Read more articles on this topic:

Did you find an error or inaccuracy?We will consider your comments as soon as possible.