[Someone]

Short way to describe the problem: we want to build systems that detect causation, but statistical models can only detect correlation.

[wongarsu]

That's not entirely true: it's hard to show causation, but with enough data you can. If A correlates with B you know that either A causes B, B causes A, some C causes both A and B, or the correlation is a coincidence. If you have the data to rule out 3 of those the remaining possibility is the causation.

[Someone]

So, how do you, for example, rule out “some C causes both A and B“, if you may not even know of the existence of C?

More importantly, the only way to really show causation is by positing a mechanism.