Here's a different hypothesis that also accounts for opinions reverting in the direction of the original uneducated position. Suppose "uneducated" and "contrarian" opinion are two independent random (e.g. normal) variables with the same mean representing the truth (but maybe higher variance for "uneducated"); and suppose what you call "meta-contrarian" opinion is just the truth. Then if you start from "contrarian" it's more likely that "meta-contrarian" opinion will be in the direction of "uneducated" than in the opposite direction, simply because "uneducated" contains nonzero information about where the truth is. I think you can also see this as a kind of regression to the mean.