It's pretty unlikely that a person would think of this on the spot, and such a person would be almost indistinguishable from someone who had already seen the qu

It's pretty unlikely that a person would think of this on the spot, and such a person would be almost indistinguishable from someone who had already seen the question. On the other hand, easier but more complicated questions tend not to have single right answer, so it's harder to judge the candidate objectively. At one interview, a lot of questions involved implementing problems that didn't require any trick (at least to people with sufficient knowledge of CS) but were fairly complex (10-20 lines of pseudocode). I think this is the right approach because sufficiently many moderately-hard questions will produce a bell curve, while questions with a "trick" provide little information.