This doesn't sound like it would be unrealistic. I think this can be partially solved by two systems: one that paints and one that finds the best painting to match a current event (in the form of a topic). The latter system could be trained on a set of paintings and corresponding topics or meanings, and then uses this training to match newly generated paintings to topics/meanings.