AI Ethics and Policy Column - The Accountability Vacuum: Agentic AI in High-Stakes Domains

Agentic AI is no longer a research artifact. These systems are embedded in live infrastructure, clinical workflows, and legal processes: deleting databases, advising patients, and generating legal documents. Wrong outputs cannot simply be taken back. The central question has shifted from whether these systems are ready to make consequential decisions to who is accountable when they get those decisions wrong. The standard apparatus of fault-finding (identifying an actor, establishing a duty, connecting a breach to an injury) was built for a world in which agents are human, decisions are sequential, and causation is legible. Agentic AI systems violate all three assumptions simultaneously. Their actions emerge from the interaction of training data, optimization objectives, architectural choices, and real-time inputs in ways that no individual actor fully controls. When something goes wrong, the failure is everyone's and no one's.