Abstract
False positives in wildlife detection datasets can substantially bias ecological inference, including diel activity pattern estimation from time-stamped detections. This problem is intensifying with widespread use of automated classifiers for camera-trap images and bioacoustic recordings, where full human verification is rarely feasible. Here, we present a Bayesian framework that provides a false-positive-aware alternative to commonly used pooled kernel density estimation workflows for diel activity. The model uses a subset of human-verified records to estimate (i) θ, the dataset-average true-positive rate among apparent detections, and (ii) time-of-day distributions describing the relative frequency of true detections and false positives across the 24-h cycle. We implement two error structures: a parsimonious uniform false-positive model and a flexible "skewed" model that learns a false-positive diel distribution to accommodate temporally clustered misclassifications. Simulations across multiple activity shapes, false-positive rates, and verification intensities showed that the approach can recover true diel patterns with limited verification effort and reduces bias relative to uncorrected estimates. We illustrate the method on an empirical dataset, where accounting for false positives yielded meaningfully different activity inferences than standard pooled approaches. Finally, we discuss extensions that allow θ to vary with covariates or random effects (e.g., site/night) and highlight opportunities to integrate diel activity correction with modern false-positive occupancy and abundance models.