What Happens to the Safety Case When Artificial Intelligence Enters Pharmacovigilance? An Umbrella Review from Intake to Signal Action

Artificial intelligence is entering pharmacovigilance through case intake, information extraction, coding support, duplicate detection, case linkage, signal detection, prioritization, and communication. These applications may improve processing capacity, yet the safety case depends on more than technical accuracy. To synthesize review-level evidence on how artificial intelligence affects pharmacovigilance from intake to signal action and to identify where evidence supports, qualifies, or leaves unresolved claims of safety benefit. An umbrella review of eligible systematic and scoping reviews published from 2017 through 2026 was conducted using a pathway-based framework. Reviews were eligible when they reported transparent search methods, explicit criteria, identifiable included studies, and findings mappable to at least one pharmacovigilance stage. Data were extracted at review level. Methodological quality, relevance, validation maturity, implementation status, discordance, and primary-study overlap were assessed using design-appropriate tools. Findings were synthesized without pooling incomparable outcomes. Evidence was concentrated in retrospective detection, prediction, text extraction, and data-processing tasks. Reviews repeatedly described heterogeneous data sources, labels, reference standards, metrics, and validation practices. External validation, prospective workflow evaluation, calibration, comparative effectiveness, and downstream assessment of communication or regulatory action were uncommon. Duplicate reporting, noisy intake sources, representativeness, automation bias, and repeated use of the same primary studies could weaken apparent confidence. Artificial intelligence may strengthen selected pharmacovigilance operations, but review-level evidence does not establish an end-to-end, deployment-ready safety case. Confidence depends on traceable data, independent validation, workflow evaluation, overlap-aware synthesis, human accountability, and explicit separation of technical performance from medication-safety benefit in routine practice.