My take is that it is not entirely disimilar from some AI research I have seen where applying conditions to simulations that are not real actually trains the model better in some ways in addition to training with real data. I remember they were using a robotic arm simulation with different gravity, friction, and other physical constants.

It makes sense to me that if you saw a small version of an animal but never saw a large one you might dream of a big one. Then if you ever saw one your brain would not have to have a brand new thought at something it may have thought impossible. I think dreams make our brains more flexible than is possible with just real data.