Will Chapman et al. investigate an important challenge in developing physically consistent AI weather and climate emulators. The study shows that enforcing exact water-budget conservation during training can inadvertently allow precipitation biases to grow, even when the final corrected output appears physically perfect. By introducing a revised training strategy that supervises the raw model predictions while penalizing budget imbalances, the authors restore stable learning and demonstrate that exact budget closure alone is not sufficient to ensure physically meaningful AI models.