We know that the training set is large and non public, and that answer keys are generally available and ML likes to output exact answers when it's been trained on the questions.

It's significantly more likely that a model outputting the right answer was trained in that answer than not