Beyond black boxes: using explainable causal artificial intelligence to separate signal from noise in pharmacovigilance

Artificial intelligence (AI), particularly machine learning (ML), is increasingly influencing pharmacovigilance (PV) by improving case triage and signal detection. Several studies have reported encouraging performance, with high F1 scores and alignment with expert assessments, suggesting that AI tools can help prioritize reports and identify potential safety issues faster than manual review. However, integrating these tools into PV raises concerns. Most models are designed for prediction, not explanation, and operate as "black boxes," offering limited insight into how decisions are made. This lack of transparency may undermine trust and clinical utility, especially in a domain where causality is central. Traditional ML relies on correlational patterns and may amplify biases inherent in spontaneous reporting systems, such as under-reporting, missing data, and confounding. Recent developments in explainable AI (XAI) and causal AI aim to address these issues by offering more interpretable and causally meaningful outputs, but their use in PV remains limited. These methods face challenges, including the need for robust data, the difficulty of defining ground truth for adverse drug reactions (ADRs), and the lack of standard validation frameworks. In this commentary, we explore the promise and pitfalls of AI in PV and argue for a shift toward causally informed, interpretable models grounded in epidemiological reasoning. We identify four priorities: incorporating causal inference into AI workflows; developing benchmark datasets to support transparent evaluation; ensuring model outputs align with clinical and regulatory logic; and upholding rigorous validation standards. The goal is not to replace expert judgment, but to enhance it with tools that are more transparent, reliable, and capable of separating true signals from noise. Moving toward explainable and causally robust AI is essential to ensure that its application in pharmacovigilance is both scientifically credible and ethically sound.