A self-monitoring explainable AI decision-support architecture for pre-publication content decisions in public-sector communication

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

AI-based decision-support systems often achieve high predictive performance but remain difficult to use when their outputs are not explainable, monitored, or connected to human action. This paper develops a self-monitoring explainable AI decision-support architecture for pre-publication communication decisions, in which monitoring generates diagnostic signals while resulting actions remain under human control. The architecture combines five layers: pre-decision input construction, XGBoost-based prediction, SHAP-based explanation, residual monitoring, and a knowledge/action layer for human-in-the-loop use. It is evaluated on 10,728 pre-publication content decisions from 16 official public-sector Facebook channels using chronology-preserving validation with train-only preprocessing. Decision-utility analyses show that the model supports ordinal differentiation between lower- and higher-potential posts. The model primarily exploits the recent communication context of the channel, while structural and format-related characteristics provide additional decision-relevant information. Temporal stress testing and residual monitoring show why the system should not be deployed as a static prediction model, but as a monitored decision-support architecture. Thus, the paper contributes a design pattern for transparent, traceable, and human-overseen AI use in dynamic organisational environments, inspired by selected EU AI Act principles.

• Scalable and Explainable AI Architectures for Intelligent Decision-Support Systems: Systematic Literature Review Protocol

  Zenodo (CERN European Organization for Nuclear Research) 2026 · 0 citations