Artificial intelligence is increasingly used to improve forecasting, optimization, control, and decision support in energy systems, yet much of the literature remains dominated by accuracy-centric evaluation. This critical review argues that such evaluation is insufficient for sustainable energy transitions because AI models also impose computational, environmental, operational, and governance costs across their lifecycle. Rather than descriptively summarizing prior work, the paper comparatively examines major groups of studies, including AI for renewable-energy forecasting, smart-grid optimization, energy management, Green AI, explainable AI, and trustworthy AI, to identify where their assumptions, metrics, and deployment priorities converge or conflict. The review shows that existing research often treats model performance, computational efficiency, carbon footprint, explainability, robustness, privacy, and accountability as separate concerns, leaving a gap in integrated lifecycle assessment. To address this gap, the paper proposes a lifecycle computational framework for sustainable AI in energy systems, covering data acquisition, model design, training, inference, deployment, monitoring, retraining, and governance. The framework supports multi-dimensional evaluation of AI systems in terms of accuracy, resource efficiency, carbon impact, transparency, reliability, and real-world deployability, offering a pathway from merely accurate AI toward sustainable, explainable, and trustworthy AI for energy applications.
