In this paper we study an example in which agents misspecify the game that they are playing. We show that there is a class of self-sustaining equilibria which are perfectly consistent with the agents’ beliefs and which are achieved by O.L.S. learning in a trivial way. We show that only the equilibria found in an earlier paper by one of us can be sustained by O.L.S. learning. Lastly, we show that if we make a continuous time approximation, then all these equilibria are unstable.