> The latter requires complex cognitive faculties not present in a mathematical model of the English language.

The training is on images and text, and the output is text. But nothing constrains the model in the middle of those two ends to only model based on language features, as opposed to world features discoverable from the language and images.