Measuring Whether Explainable AI Improves Nuclear Fault Diagnosis: A Human-Centred Evaluation Framework

Explainable artificial intelligence (XAI) is increasingly proposed for safety-critical fault diagnosis, yet a technically plausible explanation does not necessarily improve the decisions made by the people who receive it. This paper develops a controlled evaluation framework to test whether explanations improve nuclear power plant fault diagnosis beyond an AI system’s prediction alone. The proposed study compares an AI-only condition with an AI-plus-explanation condition while holding the underlying model, diagnostic cases, and prediction outputs constant. A simulated nuclear plant fault dataset is proposed as the test environment, with cases spanning normal operation and representative abnormal conditions. Participants would identify the most likely fault and indicate confidence after receiving either a prediction alone or a prediction with a concise feature-attribution explanation. Primary outcomes are human diagnostic accuracy, false-alarm rate, and time to diagnosis; secondary outcomes include explanation usefulness, confidence calibration, and perceived interpretability. The protocol uses counterbalanced task assignment, repeated cases across conditions where feasible, and statistical analysis that accounts for participant and case variability. Importantly, this manuscript reports a study design rather than completed experiments; no empirical performance gains are claimed. The framework addresses a practical gap between evaluating explanation quality and measuring whether explanations change diagnostic behaviour. It provides a reproducible basis for deciding when XAI should be included in nuclear monitoring interfaces and what evidence should be required before operational use.