Artificial intelligence is increasingly positioned to accelerate pharmacovigilance by extracting adverse-event information, classifying cases, ranking candidate signals, and directing human attention. Yet these functions are often evaluated primarily through discrimination, retrieval, or benchmark accuracy, even though operational pharmacovigilance decisions depend on a different question: whether an output can be interpreted reliably in the population, data source, time period, and decision context in which it is used. This Current Opinion argues that calibration should therefore precede automation as the governing design principle for consequential AI-assisted pharmacovigilance. Calibration is interpreted broadly but precisely: not merely as post-hoc probability adjustment, but as alignment among model output, event prevalence, data-generating conditions, evidential maturity, clinical seriousness, and the human action triggered by that output. Rare-event imbalance, dataset shift, heterogeneous information sources, changing reporting practices, and reviewer interaction can all disrupt this alignment. We propose that AI should progress from classification or ranking to increasingly autonomous workflow roles only when its outputs have been calibrated for their intended interpretation and when escalation pathways preserve human review where uncertainty remains consequential. This approach reframes automation as an outcome of demonstrated calibration rather than as the default objective of pharmacovigilance AI development.
IJPRAS
