Yes, but so is telling if a photo contains a dog or understanding sentiment in a paragraph of text. Complexity isn't quite the issue, I think it is that there is a distinction between the type of reasoning which these models have learnt and that which is necessary for concrete mathematical reasoning.