This study examines fundamental challenges in using data-driven models, especially neural networks, for simulating complex climate dynamics. While these models can often reproduce average climate behavior, they struggle to capture responses to external changes. The author, Fabrizio Falasca, shows that this limitation becomes especially pronounced when only partial observations are available, a common scenario in real-world climate systems. His findings highlight the importance of incorporating physically informed methods, like coarse-graining and stochastic parameterizations, to improve the accuracy and interpretability of neural climate emulators.