Mindreading through pure simulation.

In Simulating Minds Alvin Goldman develops his account of “mindreading”, i.e. the conditions under which we humans attribute to one another beliefs and desires (and feelings, high-level decisions, actions, perceivings, and so on). This isn’t an epistemological theory, of how our beliefs about other minds are justified, but an empirical theory of how we in fact go about the task. Accordingly, evidence from neuroscience, cognitive science and psychology looms large.

Goldman sees himself as defending a “simulation theory”, according to which the central mechanism for mindreading others is one where we use our own processes for forming beliefs and desires (on the basis of evidential inputs) or forming decisions (on the basis of beliefs of desires). The idea is that we have the capacity to deploy these processes “offline”, inputting pretend-evidence into our usual belief-and-desire forming processes, and extracting pretend-belief and pretend-desire. Or, indeed, inputting pretend-belief and pretend-desire into our normal decision-making procedures and extracting pretend-decisions. According to Goodman, a paradigmatic belief attribution might first, identify the presumed evidence of our target; second, use the simulation routine to generate pretend-belief that p, third, use an introspection/classification/attribution routine, with the overall upshot of that we form the belief that our target believes that p.

Goldman thinks of this only as a description of our “fundamental and default” mode of mindreading. Presumably the beliefs that are generated by the three-step procedure will, in the usual course of events, be brought into contact with other beliefs that we might have about the target. For example, we might have been told by a reliable source that the target definitely did not believe that p, or we might have general inductive evidence that people don’t believe that p. We might acquire information that the target is a spy who’s aiming to fool us into believing that p. Acknowledging that we are sensitive to these kind of rebutting or undercutting defeaters are quite consistent with Goldman’s theory. The point, I take it, of his describing the simulation mechanism as “default” is to allow its outputs to be weighed against all sorts of other evidence which might lead to us not ending up with the beliefs that simulation procedures direct us to. Goldman can also acknowledge that we sometimes form beliefs about others’ mental states through testimony or induction, without any role for simulation. The point, I take it, of his describing simulation-based mind-reading as “fundamental” is to acknowledge that we form beliefs about others mental states by induction and testimony, but to categorize these as subordinate to the mindreading method he is describing. All this seems pretty commonsensical–a basic but compelling foundationalist thought is that if we trace back the chains of testimony and induction, on pain of regress, we need some non-inductive or non-testimonial way of forming beliefs about others’ beliefs. Mindreading Goldman-style is a natural candidate.

A pure form of simulationism would claim that it can produce attributions of belief and desire to a target quite independently any prior views the attributer has about the target’s beliefs and desires. Pure simulationism is quite compatible with the kind of story just told about the need to integrate the outputs of simulation with prior views about the target’s psychology, but it would see that as strictly optional. But Goldman doesn’t endorse pure simulationism. Indeed, there are a couple of places where he implies that prior beliefs about the target’s psychology (either specific or in the form of generics that cover all targets) are required in order to implement simulation. The central example of this involves mindreading by retrodiction. Suppose the information we have about a target includes how they behave. We want to let this inform our opinion about what their beliefs and desires are. The interpreter has at their disposal their decision-making process, which takes in beliefs and desires and choice situations and spits out decisions. But that mechanism runs in the wrong direction, for the present interpretive problem–one can’t directly feed in pretended-decisions and get out pretended-beliefs and desires. Goldman conjectures that the simulationist instead proceed by a process of “generate and test”. We start from a hypothesis about what the target believes/desires, convert that into pretend-believes and desires which are fed, offline, into our own decision making procedure, resulting in pretend-decisions. If this matches the observed decisions, the hypothesis has passed the test and can be classified and attributed to the target. If not, we need to loop back to generate a new hypothesis. This generate-and-test loop for retrodictive mindreading involves simulation, but the “generation” (and regeneration upon failure) is not explained by simulative means. Some other explanation is owing about how we generate the hypothesis. And Goldman suggests that here we will appeal to our prior beliefs about what the target’s psychology is likely to be. If prior beliefs about target psychology are essential to retrodictive mindreading, as this suggests, then Goldman’s account is not pure simulationism. Goldman acknowledges and embraces this—he thinks that what we will end up is a compound story that involves elements of theorizing as well as simulating within the “fundamental and default” story about how we attribute higher mental states to others. (This concession generates some dialetical vulnerabilities for Goldman’s wider project, as Carruthers points out in his NDPR review of Goldman’s book.)

As I see it, the significant issue here is not that we *sometimes* use prior knowledge about others to generate hypotheses about their mental states which are fed into the generate-and-test simulation routine. That’s as unproblematic as the idea that the outputs of simulation routines need to be integrated into prior opinions about others, before they are ultimately endorsed. What’s significant is that it is looking like such prior knowledge is, for all Goldman says, always required for simulationist mindreading, to generate the hypotheses for retrodictive testing. And that means that some other method of arriving at beliefs about others’ psychology needs to be co-fundamental with the simulationist method, on pain of regress.

I want to describe a purer simulationism. Crucial to this is to think about the way the predictive and retrodictive mindreading combine. In essence, the idea will be that predictive mindreading can generate the hypotheses that are tested by retrodiction. Let’s see how that might work.

Rather than assume we *either* have information about a target’s evidential situation, or about their behaviour, let’s suppose (more realistically) that we have some information about both. Given this, here’s a doubly simulationist proposal. First, identify the target’s evidence, and run a simulation with those as pretend-outputs, to arrive at an initial set of “pretend-beliefs and desires”. This generates an initial “anchor” hypothesis about what the target’s psychology is like. But we haven’t yet brought to bear what is known about their behaviour. So we test the anchor hypothesis by feeding it into our decision-making process, arriving at pretend-decisions/behaviour. In the good case, the pretend-decisions and their behavioural signature match the behaviour that is observed, and we’ve just run a complete double cycle of simulationist mindreading, and are ready to classify and attribute the attitudes to the target.

In the bad case, though, there’s a mismatch between reality and the simulated decisions generated by the anchor hypothesis. In this case, we need to revise the hypothesis and try again. The pure simulationist at this point can conjecture that there is a fixed search space—an list of modifications to try, in order to generate a revised hypothesis. Picturesequely, we can imagine a space of possible interpretations, ordered by similarity to one another. The anchor interpretation generated by simulating the target’s reaction to evidence gives a starting point to be fed into the simulation-test-procedure to see if the decisions it predicts match those observed. If that fails to pass the test, then we move to try a sphere of closest interpretations to the anchor, and try those. If that fails, we move yet further out, and so on—until we run out of patience and give up. This is a hypothesis generating procedure that does not presuppose any prior information about the target’s psychology or the psychologies of agents in general, but only a measure of similarity between interpretations that defines a search method.

What would be the upshot of this combined way of mindreading? When starting with information about both evidential input and behavioural output of a target, we’d end up attributing that belief/desire psychology which is most similar (among psychologies that, under simulation, match the observed behaviour) to the belief/desire psychology produced by simulation of the evidence.

Just as before, this candidate psychological attribution will be tempered (rebutted/undercut/modified) by anything we happen to know about the target, and various stages in the pure form of the mindreading method can substituted by opinions we happen to hold. But even granted all this, what we have, I claim, is a possible pure simulationist routine for mindreading, one that could in principle be run entirely independently of prior views about what beliefs and desires the target has. It could, indeed, be the way we form the initial beliefs that are grist to the mill of induction and IBE by which we form psychological generalizations of that kind, within a foundationalist epistemology.

This last claim requires that prior opinion about beliefs/desires of others isn’t presupposed anywhere else in the method. Carruthers in his review of Goldman suggests that there’s another place where prior opinion matters—even predictive simulation requires we identify what the target’s evidence is, in order that this converted to pretend-evidence to be fed into our belief-forming mechanisms. That worry doesn’t seem so serious to me. As ever, the pure simulationist can concede it’s possible for prior beliefs to determine what we should take the pretend-evidence to be, to feed into the simulationist mindreading. But Goldman also defends “lower level”, automatic processes of mindreading, whereby visual information the attributer has about a target involuntarily triggers a “mirroring” within the attributer of the target’s perceptual processes. So this seems like a route entirely independent of beliefs about evidence, by which an Goldman style simulationist interpreter can identify the evidential situation of a subject, and then, by secondary “higher level” simulation, generate beliefs about what propositional attitudes they possess. The same goes for the identification of behaviour.

A second place where prior opinion might matter is the following. Goldman appeals to a distinction between those beliefs/desires of the attributer that are ‘quarantined’ (not used as auxiliary premises in simulating belief-formation/decisions) and those beliefs/desires that are not quarantined. What that distinction is, or should be, and whether *that* depends on prior opinion about targets, is I think a central question for the Goldman-style simulationist. And it’s another possible loci for simulation being infected by opinion. If the idea is that we quarantine those attitudes of ours we know to be idiosyncratic, then knowledge of how our psychological compares to those of others would be central to the operation of simulation itself. I think this, too, can be resisted, and a different story about quarantine given, but that is a matter for another time.

There are two things I take away from this discussion. The first is that it’s at least open to Goldman to develop a pure form of simulationism on which simulation is the *only* fundamental/default mindreading process, albeit one that only be operates in a “pure” way in the idealized limit. The second thing I take away from the discussion is the recipe that simulationism predicts, in that idealized limit. That is: in the ideal limit we attribute the behaviour-simulating interpretation which is most similar to the anchor interpretation simulated by the target’s evidence. That last bit of information is the sort of thing that is crucial to *my* larger project right now.

Leave a Reply

Please log in using one of these methods to post your comment:

WordPress.com Logo

You are commenting using your WordPress.com account. Log Out /  Change )

Google photo

You are commenting using your Google account. Log Out /  Change )

Twitter picture

You are commenting using your Twitter account. Log Out /  Change )

Facebook photo

You are commenting using your Facebook account. Log Out /  Change )

Connecting to %s