Abstract
This narrative review examines where artificial intelligence (AI) improves aquaculture environmental monitoring relative to conventional methods, how validation designs shape the interpretation of reported gains, and what evidence is needed for routine farm use. A thematic synthesis, supplemented by a targeted OpenAlex literature update, compares baselines, validation designs, and measured outcomes across water-quality monitoring, harmful algal bloom surveillance, contaminant sensing, and environmental management. In a flow-through farm, a neural time-series model improved imputation of influent dissolved oxygen (DO), while linear interpolation performed best for tank DO. In a separate six-pond evaluation within one farm, physics-guided transfer learning reduced the mean absolute error of nighttime DO forecasts by 35.5% relative to a standard long short-term memory model. In a 45-day pilot in one koi pond, integrated predictive control used 26.86% less energy than manual operation. These findings support task-specific use of AI for monitoring and decision support. Broader deployment of automated control requires independent farm validation and prospective assessment of water quality and resource use.