Artificial intelligence systems are rapidly entering healthcare, finance, public services and other safety‑critical domains. Yet current engineering practices remain model‑centric, focusing on accuracy, robustness, fairness and interpretability.
These properties are necessary but insufficient once the AI system is deployed, because it operates within a constantly evolving socio‑technical environment characterized by distribution shifts, institutional constraints, human feedback loops, privacy requirements and interactions among multiple AI agents.
To address this gap we propose AI Deployment Accountability Engineering (ADAE), which treats accountability as a deployment‑layer property rather than a property of an individual model. ADAE aims to provide continuous, measurable and actionable monitoring of whether a deployed system stays within acceptable risk limits, to discover the contexts in which failures arise, to attribute responsibility across technical and human components, to translate technical failures into downstream consequences, and to support timely intervention.
The research agenda of ADAE is built around four inter‑connected pillars:
- Structured discovery of context‑dependent failure modes;
- Privacy‑preserving accountability measurement;
- System‑level risk analysis for agentic AI;
- Translation of technical failures into operational and institutional risks.
The overarching goal is to establish foundational principles, mathematical tools and system architectures that enable accountable AI deployment in safety‑critical applications.
Review: ADAE shifts the focus from isolated models to the whole deployed ecosystem, offering a systematic path toward sustained safety. Realizing this vision will require interdisciplinary collaboration and scalable monitoring solutions.