Predictive Safety in the Era of Artificial Intelligence: Promises, Limits, and the Case for an Enterprise Lifecycle Evidence Ecosystem

Drug safety remains central to patient benefit, as maximizing the value of beneficial therapies requires recognition, appropriate characterization, and effective mitigation of treatment-related adverse drug reactions (ADRs). These challenges, particularly with the advent of artificial intelligence (AI) tools, have increased interest in predictive safety as a lifecycle scientific capability that seeks to anticipate plausible harm in a timely fashion and translate evolving evidence into more informed decisions across acquisitions, clinical development, and postmarketing use. In this article, predictive safety is used primarily to mean product-level and population- or subgroup-level anticipation of plausible treatment-related harm, rather than an autonomous patient-level clinical decision-making approach. Recent advances in AI, human genetics, mechanistic modeling, translational biomarkers, and real-world data have strengthened the scientific basis for this approach. However, predictive safety should never be viewed as a promise to eliminate ADRs or as a substitute for clinical judgment. Its value would be in improving prospective ADR characterization, supporting portfolio prioritization, and enabling timely and more targeted mitigation. In some settings, notably prospective genotype-based screening before exposure, it might prevent the reaction from occurring. This article argues for the development of a governed AI-supported predictive safety ecosystem. It outlines the reasons predictive safety is needed and the context in which it is most likely to add value. It also discusses principal implementation risks, including fragmented data, unstable phenotypes, model drift, transportability failure, and misuse of probabilistic outputs, together with practical mitigation strategies and leading indicators for early governance response.