Glancing at the Gul & Pesendorfer paper, “The Case for Mindless Economics,” my first thought was that these guys good really use a good philosophy of science seminar. (Tyler discusses it, as does Bryan Caplan.) Now, I don't really understand the drive to secure the autonomy of formal economics. Either economics seeks to decribe actual human behavior or it doesn't. It is fine to create any kind of formal model you like, and to call predicates in your model anything you like. You can, say, create a model that speaks of 'kittens', 'meowing', and 'estrus,' and as long as your model hangs together in the way models should, then that's nice. Maybe it will even be interesting. But if it has nothing much to do with actual purring furry pets, don't get upset when a zoologist comes along says that you don't really have a theory of cats at all, but a theory of 'cats.'
My sense is that the formal neoclassical theory of economics is a theory of 'behavior' not behavior. Models don't need to be models of the actual world in order to be worthwhile. But you just can't get away with too much model/world three card monte, maintaining that your model both does apply to the behavior of real people, but cannot be touched by investigations into the sources of the behavior of real people.
Right on page 1, Gul and Pesendorfer say “Neuroscientific evidence cannot refute economic models because the latter make no assumptions and draw no conclusions about the physiology of the brain,” but this just has to be false if economic models are actually supposed to be models of human behavior. Because, see, human beings have brains. And brains are actual phsyical systems that have physical constraints that limit, say, the amount of information that can be processed within a period of time. If an economic model of the agent assumes instantaneous, zero-error updating, or the all-at-once representation of an ordering including the uncountably large number of options in the feasible set, or knowledge of the entire set of all other agents uncountably large preference orderings, then the model is in fact making a large number of assumptions about the information-processing capacities of agents. And if it turns out that human information processing capacities fall far short of these assumptions (and, oh, do they ever) then you have a choice when it comes to the economic model. You can either admit that it is not a model of actual human behavior, but just of some imaginary agent not subject to the laws of physics, such that they can complete all this computation before the heat death of the sun, or the implosion of the Universe. Or you can admit that it is bad model—false, due to non-correspondence—but, perhaps, of some heuristic value in hypothesis formation, or some such thing.
What you can't do is say you've got a model of something that is in fact a biological system, and then argue that real information about the nature of the biological system is irrelevant to the truth of your model. I may be crazy, but that seems to me exactly what Gul and Pesendorfer are trying to argue. If it is, I am completely sure that it won't work.