I'm not sure how effective an AI would really be on this. Distinguishing from a pattern on top of the dice vs a pattern on the side/front seems like it would actually produce a lot of false positives.
I don't know - with some edge detection I think you could distinguish between the faces of each die, and some heuristics would probably get you the upwards-facing one. Another thing that may be useful is that in each image the only face that is fully visible for all dice is the upward one (who knows if that is just for this example though).
I think if you iterated over enough pictures of the dice from the CAPTCHAs along with the correct sums, you could easily train a model to solve it.