If "information is a measure of surprise", then you don't want every prediction to be wrong. You want every prediction to have a 50% chance of being wrong. Half of your experiments should be failures.
I don't understand this logic at all. At those odds, you're just guessing like flipping a coin. An inferred prediction should have a higher success rate. We look at the data we do know, we see what holes are there, and then make predictions based on all of our previous knowledge. The fact that the previous knowledge allows us to make more accurate predictions shows we have a better understanding of the subject than just random coin flips.