Safety evaluation of pharmaceuticals does not end when they enter the market. Because clinical trials involve use only under limited conditions, new safety information must continue to be collected and assessed from Real-world Use after market introduction. The institutionalized framework for this approach is Pharmacovigilance, which provides an important reference model for structuring AI Vigilance.

In Pharmacovigilance, information such as Adverse Events reported after market introduction is collected and assessed, and new Signals are detected across multiple sources of Evidence. Where necessary, this leads to Regulatory Reporting, revisions to product labeling, changes in conditions of use, or additional Safety Measures. Continuous Monitoring then continues, returning new Evidence to further Assessment. Rather than functioning as a response to isolated incidents, this process operates as Continuous Safety Governance structured under systems such as GVP.

AI presents a similar problem. Pre-deployment Testing and Evaluation cannot anticipate every issue that may emerge when an AI System is used in the real world. As users, Context, connected Systems, Data, Models, Policy, and the broader social environment change, Risks and Patterns that were not identified at Deployment may emerge.

Various mechanisms already respond to AI-related problems, including Provider Monitoring, Incident Reporting, Audit, Standards, Government Guidance, Voluntary Actions, and Litigation. However, when these mechanisms exist independently, they do not necessarily form a coherent Feedback Loop that continuously collects and assesses Evidence from Real-world Use, detects Signals across multiple events, and returns the findings to Governance improvement.

AI Vigilance provides a way to connect these Fragmented Responses into Continuous Vigilance. This does not mean directly transferring pharmaceutical governance systems to AI. What AI Vigilance draws from Pharmacovigilance is the structural principle of not relying solely on pre-market evaluation, but continuously observing Real-world Use, detecting emerging Risks from Evidence, and connecting those findings to appropriate Governance Action.

AI Governance therefore should not end at Deployment. Continuous observation and evaluation must continue after Deployment. Execution Evidence, Incidents, Deviation, Anomaly, Repetition, Shift, and Emerging Pattern are continuously observed, with relevant findings connected to appropriate Evaluation or Escalation. Under the appropriate Authority and Process, these findings can then lead to the review or update of Policy, Process, Constraints, or other Governance elements.

AI Vigilance is not a mechanism for evaluating an AI system once and declaring it “safe.” It is a mechanism for observing a continuously changing Real-world Environment, detecting change through Evidence, and keeping Governance continuously operational.