For any question the number of incorrect answers is larger than the number of correct answers.
This is a fundamental problem, constrained by energy costs, and one that will only be exacerbated as training datasets becomes more and more tainted by generated content.