Stich, S. P., & Nichols, S. (1992). Folk Psychology: Simulation or Tacit Theory? Mind And Language, 7(1-2), 35–71. doi:10.1111/j.1468-0017.1992.tb00196.x

Summary

In this paper, Stich & Nichols argue against simulation theory in favor of theory theory. They first give an overview of their interpretation of what simulation theory is positing, which seems consistent with what I’ve read so far. They make a particular point about calling the simulation theory off-line simulation of the decision-making system. That is, they argue that simulation theory says that to run a simulation, the decision-making system is taken off-line. It is fed pretend beliefs and desires, and it outputs a decision, which is then fed to another system which explains and interprets decisions in terms of their inputs. They also clarify what they mean by “theory”, which is (as they put it) in the “wide” sense: that a theory need not necessarily be symbolic/logical, but could take the form of other representations like a neural network as well.

Stich & Nichols go through a number of arguments for the simulation theory, and give counter evidence to each one.

What are the rules of the folk psychological theory? This is interpreting a “theory” in the narrow sense. And besides, not being able to specify the rules doesn’t make the theory invalid (example: language, folk physics).
Mental simulation models have been successful. “Simulation” used in the context of “mental simulation” is actually closer to a theory that is ust not rule/sentence like.
Children can’t possibly have as sophisticated a theory as theory theory suggests. Theory theory doesn’t necessarily posit a theory that is the same type of stuff that you learn in the classroom. Besides, children learn other very complicated things (language, folk physics)
Simulation theory is simpler than theory theory. Specifying the “control mechanism” for coming up with pretend beliefs and desires is non-trivial. Essentially, in theory-theory, you get the control for free, while in simulation theory, you get the “database” of knowledge for free. They are equally complex theories.
Sometimes we imagine how others behave by imagining ourselves in their situation. Imagery is not the same thing as simulation. For example, to imagine yourself walking through your house and counting the number of windows, it’s not just running an off-line simulation—the knowledge about the number of windows has to come from somewhere.
Simulation theory is supported by developmental evidence. It is consistent with the evidence; that doesn’t it predicts it. Theory theory is also consistent with the evidence.
Autistic children are poor at engaging in pretend play and at reasoning about theory of mind. As in the previous point, simulation might be able to account for this, but so can theory theory.

After refuting the arguments of off-line simulation theory, Stich & Nichols present a few arguments in favor of theory theory:

There is developmental data that simulation theory can’t explain. In particular, there is evidence that children have accurate knowledge about what other children have seen and yet still make incorrect judgments about their beliefs. That is, it seems like children have all the correct inputs that would be used to run an off-line simulation, yet they still come up with the wrong answer, but only for other people—they give a different answer themselves.
Theory theory is cognitively penetrable. Simulation theory is not cognitive penetrable. In particular, because it’s a simulation of one’s own behavior, if there are any quirks in our own behavior that we don’t consciously know about, they should still show up in our predictions of other people’s behavior. However, if the theory posited by theory theory doesn’t include these things explicitly, then it will make the wrong predictions because it is cognitively penetrable. There are many examples of cases where people predict that they (or others) would behave in a certain way, but then actually behave differently themselves. Thus the can’t have just been running a simulation.

Takeaways

I agree with pretty much all the points made by Stich & Nichols, though I do think they are a bit vague on the notion of what a theory is. Ideally, even if you can’t specify a set of rules, you should still be able to roughly sketch the mechanism by which it works, or by which it is acquired.

I do like how Stich & Nichols explicitly contrast offline simulation with other forms of simulation. They state that they don’t think offline simulation is involved at all, which I would probably agree with—though I do think some form of “simulation” probably is used. Perhaps the better term to use here is “emulation” (ala Grush), to imply that it is a simulation from a model, not a simulation from oneself. Such a model could easily be a component in a theory; thus, this notion isn’t at odds with theory theory.