Listening to the City: Evaluating BirdNET Reliability and Bias in Urban Bird Community Assessments

BirdNET is increasingly used for passive acoustic monitoring (PAM), yet its reliability in urban environments that have high levels of anthropogenic noise remains poorly understood. We evaluated BirdNET's performance in two Italian cities, Turin and Florence, by analysing over 4300 10‐min audio recordings (approximately 723 h of audio in total) and manually validating up to 500 detections per species. In total, BirdNET detected 203 species in Florence and 192 in Turin, many of which were shared between the two cities. Using BirdNET Analyzer, we quantified true and false positives, characterized misclassification types, compared classifier performance between seasonally filtered and full‐year species lists, and estimated species‐specific confidence thresholds. BirdNET overestimated species richness, and a substantial proportion of detected species could not be confirmed during manual review: only 48 species in Turin and 65 in Florence were validated as genuinely present, corresponding to 58% and 52% of detected species being false detections, respectively. Applying seasonally filtered species lists improved precision relative to the full‐year list in both cities (by 11.2% in Turin and 8.2% in Florence). Most misclassifications resulted from biophonies (confusion with other bird species), followed by anthropophonies, such as traffic noise and human voices. Despite these errors, overall classification performance was good, with segment‐level precision reaching 0.704 in Turin and 0.780 in Florence. Optimal confidence thresholds varied widely among species and differed between cities, reflecting local noise conditions and community composition and highlighting the need for site‐specific calibration. Our results indicate that although BirdNET can effectively support urban biodiversity monitoring, its performance is highly context‐ and species‐dependent. Applying ecological filters and calibrating species‐specific confidence thresholds are essential to reduce misclassifications and improve the reliability of automated acoustic data in complex urban soundscapes.