I just submitted an article to eLearn Magazine that starts to get at one of the reasons why I am skeptical about emergent learning as a panacea. I’m not going to give away too many of the surprises here (unless eLearn decides not to publish the article, in which case I will post it here), but let’s just say that you can find the roots of the argument in the article Conversation, Learning, and Informational Cascades [PDF]. From the abstract:
We offer a model to explain why groups of people sometimes converge upon poor decisions and are prone to fads, even though they can discuss the outcomes of their choices. Models of informational herding or cascades have examined how rational individuals learn by observing predecessors’ actions, and show that when individuals stop using their own private signals, improvements in decision quality cease. A literature on word-of-mouth learning shows how observation of outcomes as well as actions can cause convergence upon correct decisions. However, the assumptions of these models differ considerably from those of the cascades/herding literature. In a setting which adds ‘conversational’ learning about both the payoff outcomes of predecessors to a basic cascades model, we describe conditions under which (1) cascades/herding occurs with probability one; (2) once started there is a positive probability (generally less than one) that a cascade lasts forever; (3) cascades aggregate information inefficiently and are fragile; (4) the ability to observe past payoffs can reduce average decision accuracy and welfare; and (5) delay in observation of payoffs can improve average accuracy and welfare.
