Journal of Modern Technology and Engineering

Electricity-theft detection (ETD) is increasingly delegated to deep neural networks that report near-ceiling accuracy on a few shared smart-meter benchmarks. A high benchmark score, however, certifies discrimination on one test set; it does not certify that a utility can deploy the detector or that a regulator can defend the intervention it triggers. Drawing on the trustworthy-artificial-intelligence frameworks that energy regulators are beginning to adopt, we organize trustworthiness into six dimensions of a deployed detector: reliability under distribution shift, calibration, explainability, fairness, privacy, and auditability. ETD serves as a case study that makes each dimension concrete and high-stakes, because severe class imbalance, confirmed-only labels, single-utility benchmarks, non-stationary metering, and a consumer-facing intervention each stress a different dimension that a single accuracy number cannot see. We find that the field’s reporting substantiates the accuracy axis and is largely silent on the other six. Because a deployed detector is trustworthy only to the extent of its weakest dimension rather than its best benchmark number, a single reported score cannot express trust; we close with a short, per-dimension trustworthiness profile that a study could report alongside accuracy.