Models can produce out-of-distribution results if you push them. The results won't necessarily be good, but describing it as an average, suggesting that it'll a

Models can produce out-of-distribution results if you push them. The results won't necessarily be good, but describing it as an average, suggesting that it'll always be inside the some tight hull of the training set is misleading. Also note that averages can actually be highly unusual results if you're working with N-modal data, if you have a training set of horses and unicorns then you can get unusually short horns or other weirdness like horn-looking hair decorations. Just look at the avocado chair or dogball.