They're testing understanding for the learning set, but applying a correlation to estimate understanding on the test set.

What is scientific causation but correlation that works to an acceptable quantity of Sigma?

Causation also implies well sequential causality in some way and reproducibility. A subtle but crucial difference.

If most people on the battlefield are covered in their own blood will soon die that doesn't mean gathering a gallon of a person's own blood through several donations and then throwing it at them at a Gettysburg musuem will kill them.

As Feynman discussed in "The Character of Physical Law", there is more to it than that, but I think one can say that the empirical evidence underpinning all science can be characterized as correlation that passes various statistical tests.

Causal studies need to control for confounding variables, for one. There are all sorts of spurious non causal correlations with low p values.