I've also experienced this, to an extent, but on qualitative topics the goodness of an answer - beyond basic requirements like being parseable and then plausible - is difficult to evaluate.

They can certainly produce good-sounding answers, but as to the goodness of the advice they contain, YMMV.