Integrating AI models into ecological research workflows: The case of terrestrial bioacoustics

Abstract Data collected by autonomous sensors, including camera traps and acoustic recorders, have enormous potential to generate new scientific insights in ecology and related fields. Modern machine learning and AI classification methods are critical to analysing these often immense data streams. Accordingly, considerable effort has been dedicated to building AI models that accurately detect and classify species and events of interest in these data. These AI models, however, form only one part of a larger research framework that is needed to answer ecological questions using sensor data. We argue that a deep understanding of this research context is required to develop and apply appropriate AI models that can support scientific advances in ecology and evolution. In this manuscript, we contextualize the use of AI methods in autonomous biodiversity surveys, focusing on terrestrial bioacoustics as a case study, by discussing six sequential areas that together form a research project: hardware, field deployments, data management, detection and classification using AI and related models, statistical analysis and ecological insight. For each area, we briefly highlight several ways in which decisions made in that area can constrain, support or interact with the development and application of AI models. We conclude with several suggestions for better development and integration of AI models into ecological research, including the need for additional research at the interface of AI models and statistical analysis, the question of achieving human‐level performance with AI models and the sources of future methodological advances in AI for ecology.