All of the ai climate models, such as the one by Microsoft mentioned, are trained on model outputs not data. They simply are learning the model that was already developed. The same is true for the Google model which is trained on era5 reanalysis which is a mix of modeled and real data. From my perspective the main benefits of these neural network replacements is that they can, once trained, run on desktop hardware and are much less onerous to implement and iterate with. Their power lies in their ability to be used easily and efficiently, not in that they are able to discern physics from raw data.