Well-intentioned initiatives hinder understanding biodiversity conservation: an essay on a recent deep-learning image classifier for Amazonian fishes

The identification of fish species by non-specialists remains a constant challenge for biodiversity management. In this regard, Robillard et al. developed a machine learning computer vision model to identify Amazonian fish at the genus level, with an accuracy of 97.9%. Their model aimed to facilitate fish identification by non-specialists, allowing them to contribute to collecting and sharing data for biodiversity management. However, when tested with a different set of fish pictures, their classifier was unable to accurately identify fish photographs, resulting in 82% of misidentification, and did not outperform what would be expected by chance, indicating that it is not suitable for the accurate identification of taxa in its current form. The results underscore the need for a balanced approach, combining automated tools with expert taxonomic input for accurate conservation decisions, emphasizing caution in relying solely on Artificial Intelligence methods. While acknowledging the potential of the model, we recommend restricting its application primarily to larger fish of commercial interest or scenarios where conservation decisions are less directly affected by the model's identifications.