Let's say you have 100 programmers. 20 of them "care". Of the 100, 5 use emacs. P(emacs) = 0.05. If all emacs users care, then P(emacs|care) = 0.25. If I wanted to model this phenomenon, I shouldn't ignore the correlation just because it's not perfectly causal.
The fallacy if any is assuming likeliness implies implication. This is commonly known as stereotyping, but it's perhaps more clearly described as generalizing a group stereotype to the entire group.
"stereotyping" is a floating signifier, and arguably in common usage doesn't merely signify hasty generalization.
(And that's also why you care about causation even if you are only looking for prediction, because you don't want your proxies to change their meaning as soon as you act on them.)
It is indeed full of correlation that can unmask interesting facts if studied further. And the GP complaint is meaningless if viewed from that angle, so I do really think he was talkinga bout causation.