Let's suppose that as part of an experiment, 100 genetically similar people are offered a choice: accept a red hat or accept a blue hat. They may, statistically, have a preference toward one color or the other. Now, if you only saw one of these people, and you saw that they accepted a red hat, would you assume that red hats or blue hats were more commonly chosen among the other 99? It's only one data point, but gun to your head, you would probably assume that among the other 99, red hats were more popular.
Here's where it gets weird. What if you were one of the 100, and you chose the red hat? Would you make the same assumption? I argue you probably could. Even if you chose it for some obscure reason, knowing that reason led you to choose the red hat means that others might pick it for the same reason. But if you picked the blue hat, then it would similarly seem more likely that blue hats were probably more likely overall.
Now, knowing this, what if you wanted to influence what hats the others picked? If you pick the red hat, then you can logically conclude that red hats are more likely to be popular. If you pick the blue hat, then you can logically conclude that blue hats are more likely to be popular. Based on this choice, you can change the expected outcome of events you don't participate in. Suppose everyone read this thought experiment before participating. If you pick the red hat to influence the group, it increases the expected likelihood that the majority picked the red hat to influence the group.
I suspect that a possible answer to this paradox is that picking the hat doesn't give you any new information: even before you pick it, you could guess your own thought process and then make your best guess as to what the other 99 pick. Then that would be the most accurate guess you could make regardless of what color you choose. I don't know if this answer would hold up empirically, but it's the only solution that makes sense to me.