There are two actions, A and B, and there are two observable types of people L and R.  Everybody is the same in the following sense:  for any single individual either A or B is the optimal action but which one it is depends on an unknown state of the world.

But in another sense people are heterogeneous.  It is common knowledge that in the state of the world where A is best for people of type L then B is best for people of type R.  And in the other state its the other way around.  Each person observes a private signal that contains information about the state of the world.

Acting in isolation everybody would do exactly the same thing:  pick the action that is best according to their belief (based on the private signal) about the state of the world.  But now embed this in a model of social learning.  People make their choices in sequence and each observes the choices made by people who went before.

Standard herding logic tells us that L’s and R’s will polarize and choose the opposite action even if they get it completely wrong (with L’s choosing the action that is best for R’s and R’s choosing the action that is best for L’s)

(A reminder of how that works.  Say that an L moves first.  He chooses the action that looks the best to him say A.  Now suppose the next guy is an R and by chance action B looks best to him.  The third guy is going to look at the previous two and infer from their choices that there is strong information that the true state is such that A is good for L’s and B is good for R’s. This information can be so strong that it swamps his one little private signal and he follows the herd:  choosing A if he is L or B if he is R.  This perpetuates itself with all subsequent decision makers.)

In effect the L’s choose A just because the R’s are choosing B and vice versa.