There's a very interesting point in the thread below which would resolve the question - we could use chess puzzles! If there is a mismatch between performance on puzzles with live results, it would be pretty solid proof that the effect is real. It doesn't have to be full on AI.