Why genome-wide associations with cognitive ability measures are probably spurious