Comment by Catloafdev

Frankly, I don't buy this difficulty argument.

They know which model was used to come up with that particular idea.

A text search over the corpus of user data used in the training set can only take so long.