TL;DR.

Why?

1. Heuristic approximations often confer significant knowable performance improvements that come at a cost of only approximately performing the original processing. Often, the approximate algorithms do not have a closed-form specification of their behavior; rather, one just understands that it is somehow similar to the behavior of the unoptimized system. This introduces an ill-understood delta in the system’s behavior.
2. The nature of these intelligence algorithms does not seem to lend itself to simple closed form understandings. The specifics of what they will do tend to be the result of a complex interaction beween data and heuristic and greedy learning.