Keyboard Shortcuts?

×
  • Next step
  • Previous step
  • Skip this slide
  • Previous slide
  • mShow slide thumbnails
  • nShow notes
  • hShow handout latex source
  • NShow talk notes latex source

Click here and press the right key for the next slide (or swipe left)

also ...

Press the left key to go backwards (or swipe right)

Press n to toggle whether notes are shown (or add '?notes' to the url before the #)

Press m or double tap to slide thumbnails (menu)

Press ? at any time to show the keyboard shortcuts

\title {The Developing Mind \\ The Mindreading Puzzle: a Solution?}
 
\maketitle
 

The Mindreading Puzzle: a Solution?

\def \ititle {The Mindreading Puzzle: a Solution?}
\begin{center}
{\Large
\textbf{\ititle}
}
 
\iemail %
\end{center}
The Mindreading Puzzle

There is an age at which:

1. in performing an A-task, the child relies on a model of minds and actions not incorporating beliefs;

2. in performing false belief tasks which are not A-tasks, the child relies on a model of minds and actions incorporating beliefs; and

3. the child has a single model of minds and actions.

An A-task is any false belief task that children tend to fail until around three to five years of age.
So far I argued that we can’t plausibly reject either 1 or 2 outright. In this talk I want to suggest an alterative.
But before I do that, I want to introduce a related puzzle.

another puzzle: automaticity

Schneider et al (2014, figure 1)

One way to show that mindreading is automatic is to give subjects a task which does not require tracking beliefs and then to compare their performance in two scenarios: a scenario where someone else has a false belief, and a scenario in which someone else has a true belief. If mindreading occurs automatically, performance should not vary between the two scenarios because others’ beliefs are always irrelevant to the subjects’ task and motivations.

Schneider et al (2014, figure 3)

\citet{Schneider:2011fk} did just this. They showed their participants a series of videos and instructed them to detect when a figure waved or, in a second experiment, to discriminate between high and low tones as quickly as possible. Performing these tasks did not require tracking anyone’s beliefs, and the participants did not report mindreading when asked afterwards.
on experiment 1: ‘Participants never reported belief tracking when questioned in an open format after the experiment (“What do you think this experiment was about?”). Furthermore, this verbal debriefing about the experiment’s purpose never triggered participants to indicate that they followed the actor’s belief state’ \citep[p.~2]{Schneider:2011fk}
Nevertheless, participants’ eye movements indicated that they were tracking the beliefs of a person who happened to be in the videos.
In a further study, \citet{schneider:2014_task} raised the stakes by giving participants a task that would be harder to perform if they were tracking another’s beliefs. So now tracking another’s beliefs is not only irrelevant to performing the tasks: it may actually hinder performance. Despite this, they found evidence in adults’ looking times that they were tracking another’s false beliefs. This indicates that ‘subjects … track the mental states of others even when they have instructions to complete a task that is incongruent with this operation’ \citep[p.~46]{schneider:2014_task} and so provides evidence for automaticity.% \footnote{% % quote is necessary to qualify in the light of their interpretation; difference between looking at end (task-dependent) and at an earlier phase (task-independent)? %\citet[p.~46]{schneider:2014_task}: ‘we have demonstrated here that subjects implicitly track the mental states of others even when they have instructions to complete a task that is incongruent with this operation. These results provide support for the hypothesis that there exists a ToM mechanism that can operate implicitly to extract belief like states of others (Apperly & Butterfill, 2009) that is immune to top-down task settings.’ It is hard to completely rule out the possibility that belief tracking is merely spontaneous rather than automatic. I take the fact that belief tracking occurs despite plausibly making subjects’ tasks harder to perform to indicate automaticity over spontaneity. If non-automatic belief tracking typically involves awareness of belief tracking, then the fact that subjects did not mention belief tracking when asked after the experiment about its purpose and what they were doing in it further supports the claim that belief tracking was automatic. }
Further evidence that mindreading can occur in adults even when counterproductive has been provided by \citet{kovacs_social_2010}, who showed that another’s irrelevant beliefs about the location of an object can affect how quickly people can detect the object’s presence, and by \citet{Wel:2013uq}, who showed that the same can influence the paths people take to reach an object. Taken together, this is compelling evidence that mindreading in adult humans sometimes involves automatic processes only.

Evidence for non-automaticity

Back & Apperly (2010, fig 1, part)

This is the data for answers that required a ‘yes’ response.
So does all mindreading in adult humans involve only processes which are automatic? No: it turns out that verbal responses in false belief tasks that are A-tasks are not typically a consequence of automatic belief tracking. To show this, \citet{back:2010_apperly} instructed people to watch videos in which someone acquires a belief, either true or false, and then, after the video, asked them an unexpected question about the protagonist’s belief \citep[see also][]{apperly:2006_belief}. They measured how long people took to answer this question. Starting with the hypothesis that answering a question about belief involves automatic mindreading only, they reasoned that the mindreading necessary to answer a question about belief will have occurred before the question is even asked. Accordingly there should be no delay in answering an unexpected question about belief—or, at least, no more delay than in answering unexpected questions about any other facts that are automatically tracked. But they found that people were slower to answer unexpected questions about belief than predicted. Importantly this was not due to any difficulty with questions about belief as such: when such questions were expected, they were answered just as quickly as other, non-belief questions. It seems that, when asked an unexpected question about another’s belief, people typically need time to work out what the other believes. We must therefore reject the hypothesis that answering a question about belief involves automatic mindreading only.% \footnote{% \citet[ms~p.~9]{carruthers:2015_mindreading} objects (following \citealp{cohen:2009_encoding}) that these experiments are ‘not really about encoding belief but recalling it.’ Note that this objection is already answered by \citet[p.~56]{back:2010_apperly}. }
\citep[p.\ 132]{Wel:2013uq}: ‘In support of a more rule-based and controlled system, we found that response initiation times changed as a function of the congruency of the participant’s and the agent’s belief in the explicit group only. Thus, when participants had to track both beliefs, they slowed down their responses when there was a belief conflict versus when there was not. The observation that this result only occurred for the explicit group provides evidence for a controlled system.’

van der Wel et al (2014, figure 3)

Two puzzles, not just one. What’s nice here is that the distinction between A-tasks and non-A-tasks appears to be associated with a further performance factor: automaticity vs non-automaticity of reasponses.

Belief tracking

not A-tasks A-tasks
automatic

Schneider et al (2012,2014a,2014b); Kovács et al (2010); van der Wel et al (2014)

How can we resolve the two puzzles? I think the puzzles are fundamentally about how mindreaders *model* the mind. But before getting to that, let me introduce a hypothesis about *systems*.

hypothesis:

There are two (or more) systems for tracking others’ beliefs, one more automatic than the other.

[Aside on systems: It is difficult to say what a system is without courting controversy. Fortunately, we don't need a detailed theoretical account of systems. It should be almost uncontroversial that humans have multiple systems for tracking numbers (counting vs subetizing), causes, colours and actions. I use ‘system’ as a placeholder for whatever turns out to be the right theoretical notion for understanding these cases. The conjecture that there are multiple systems is not tied to a particular notion of system. Rather, the conjecture is that belief is like number, colour, causation and actions in that there are multiple systems for tracking all of these things. (Caffeine free diet coke is coke and alcohol free, gluten free beer is beer.)]
The hypothesis that humans have two or more systems for tracking others beliefs makes a prediction which distinguishes it from competitors.

prediction:

automatic and non-automatic tracking can yield inconsistent responses to a single scenario.

The prediction is that, where different measures reflect different belief-tracking processes, it should be possible for these measures to yield evidence of inconsistent responses to the same stimuli. This is best illustrated with a diagram ...
One inconsistent response. Might just be that automatic responses is a more sensitve measure ...
Here the idea is that, on the first task, one measure indicates that the subjects predict an action that would be rational given a false belief whereas the other measure indicates that the same subejcts predict a different action, one that would not be rational given the false belief.
Switch from two tasks to 2 ages and location vs identity.
So I think there is some evidence, not decisive but still compelling, for he hypothesis that humans have two or more systems for tracking others beliefs.

hypothesis:

There are two (or more) mindreading systems for tracking others’ beliefs, one more automatic than the other.

prediction:

automatic and non-automatic tracking can yield inconsistent responses.

So far I have only been talking about tracking beliefs. By saying that a system \emph{tracks} beliefs I mean that, within limits, it enables a subject to respond differently depending on what others believe.
Now tracking beliefs does not necessarily involve \emph{mindreading}, the process of identifying mental states as the mental states of a particular subject. Next I want to move towards a bolder conjecture:
I want to say not just that there are multiple belief-tracking systems but, further, that there are multiple mindreading systems. that is multiple systems that track beliefs by means of identifying mental states. To explain this idea, I need to return to the question in my title, How do mindreaders model minds?
Recall the Mindreading Puzzle

There is an age at which:

1. in performing an A-task, the child relies on a model of minds and actions not incorporating beliefs;

2. in performing false belief tasks which are not A-tasks, the child relies on a model of minds and actions incorporating beliefs; and

3. the child has a single model of minds and actions.

An A-task is any false belief task that children tend to fail until around three to five years of age.
So far I argued that we can’t plausibly reject either 1 or 2 outright. Now, finally, I want to suggest an alterative.
To even make sense of the question, How do mindreaders model minds?, we need to reject a dogma. The dogma is that there is one model of the mental and mindreading involves the use of that model. Or, more carefully (to accommodate Wellman et al), the dogma is that there is either just one model or else a family of models where one of the models, the best and most sophisticated model, contains all of the states that are contained in any of the models.
[Is this way of putting is clearer? : the mental states included in each model are a subset of the mental states included in the best, most sophisticated model. (The idea is that there is a model containing all the states in the union of the sets of states contained in each model.) ]
You can see that the dogma is not something we should take for granted by drawing a parallel between mindreading and physical cognition.

How do physical thinkers model the physical?

How do mindreaders’ belief-tracking systems model minds?

The notion of model complements that of system. The idea is going to be that different belief-tracking systems rely on different models of the mental. But I'm getting ahead. Let me start with a simple question.
How do mindreaders model minds?
Or, if you accept the conjecture, we should really be asking about how different belief-tracking systems model minds.
This question needs explanation. Let me explain the question by comparing an anlogous question about the physical ...
How do physical thinkers model the physical?
Focus just on the physical case first.
The question is, How do physical thinkers model the physical?
To say that a certain group of subjects can represent physical properties like weight and momentum leaves open the question of how they represent those things.
In asking how the subjects, infants say, weight or momentum, we are aiming to understand these things as infants understand them; we are aiming to see them as infants see them. (NB: I'm going to focus on human adults!)
How can we do this? We need a couple of things [STEPS: (1) theories; (2) signature limits; ]

1. theories

Impetus vs Newtonian

The first thing we need is theories of the physical or mental.
It is a familiar idea, from the history of science, that there are multiple coherent \textbf{theories} of the physical: impetus and Newtonian mechancs, for example. The impetus theory says that moving objects have something, impetus, that they gradually loose. When they loose their impetus they stop moving. If you push them you impart impetus to them, and that is why they move. With Newtonian mechanics this is not the case; there is no impetus.
Theories specify models. A theory isn't a model, and to say that someone relies on a model is not to say that they necessarily know any theory. But the theory describes a way the universe could be, and so specifies one way that a process could model it as being.
The impetus theory isn't right but it is (broadly) coherent; it describes a way the world could be. This is why it specifies a model.

minimal vs canonical

2. flexibility/efficiency trade-offs

The second thing we need (in order to answer the question about how humans model the physical) is to understand why different humans use different models of the physical.
What makes it useful to have different models of the physical for, say putting up a garden fence or landing a robot on a comet, is that some models allow you to get answers quickly without too much effort whereas other models, although harder to use, will provide accurate answers even in situations far from the mundane.
So because we're interested in actual thinkers, we are looking for pairs (or sets) of models that allow different trade-offs between efficiency and flexibiliy. It's for just this reason that the contrast between impetus and Newtonian models is interesting ...

impetus vs friction+air-resistance+...

Whereas a Newtonian theory requires computing several independent forces such as friction and air resistance, which are both distinct from the forces imparted in launching an object, impetus mechanics in effect rolls these all into a single thing, the object’s impetus.

relational vs propositional

codifiable vs uncodifiable

3. signature limits

A signature limit of a model is a set of predictions derivable from the model which are incorrect, and which are not predictions of other models under consideration.
In limited but common range of cases, impetus and Newtonian mechanics coincide. However, the two theories make different predictions about the acceleration of falling objects, and of ascending objects (those launched vertically, in the manner of a rocket).
Consider ascending objects. We're fixing density and shape and considering how the size of objects changes things.
According to Newtonian mechanics, if we ignore air resistance, then size makes no difference to accelleration. If we include air resistance, larger objects accellerate faster (because of the difference in ratio of mass to surface).
By contrast, according to an impetus principle: ‘More massive objects accelerate at a slower rate. An object’s initial impetus continually dissipates because it is overcome by the effect of gravity. The more massive the ascending object, the more gravity counteracts its impetus.’ \citep[p.\ 445]{kozhevnikov:2001_impetus}

vertical launch

So vertical launch is a signature limit of impetus models of the physical.
Interestingly, there is evidence that some automatic responses show this signature limit of impetus models of the physical.

numerical identity

Here is, in a much simplified form, what Kozhenikov and Hegarty showed. First you see that adults without knowledge of physics tend to make non-automatic---explicit verbal---predictions in line with an impetus model, whereas experts make predictions that conform to a Newtonian model. But here's the really cool thing ...
[The simplification consists in ignoring the fact that they actually measured the difference in predictions for small vs large objects; so strictly speaking what the bars in the figure here represent are the prediction about the large object’s position minus the prediction about the small object’s position. Note that (unlike the mind case), there is no claim about chance performance.]
Irrespective of whether you are a novice or an expert, your automatic responses (representational momentum) are subject to a signature limit of impetus mechanics. This is evidence that those automatic responses rely on such an impetus model of the physical, as Kozhevnikov and Hegarty argue.
So in the case of physical cognition, it makes sense to ask, Which model of the physical does a particular process rely on? And the answer seems to be that some automatic processes rely on an impetus model, whereas some non-autoamtic processes rely on a Newtonian model (at least in experts).
OK, that should be enough to show that the dogma should not be assumed without argument. Where someone is a mindreader, that is, is capable of identifying mental states, we need to understand what model of the mental underpins her abilities. So any time we have a mindreading process, it is sensible to ask, What model of the mental does that process rely on?
Is there evidence for the positive claim that the dogma is wrong and that different mindreading systems in humans rely on different models of the mental? I think there is. Let me explain ...
Start with the theories first. Where do we get theories that might allow us to identify how mindreaders model the mental? As in the case of the physical, we're interested in simple theories that only need to be approximately correct.
Theories are the things that philosophers create when they do things like trying to explain what an intention is, (Bratman says he is giving a theory of intention, for example). Or when they try to explain how belief differs from supposing, guessing and the rest. Most of these theories are highly sophisticated and concern propositional attitudes only. But what about the bad theories, the mental analogues of impetus theories?
Instead of going to the history of science for our bad theories, we turn to early philosophical attempts to characterise mental states. My favourite is Jonathan Bennett's. These theories are hopeless considered as accounts of adult's explicit thinking about mental states. But, like impetus theories of the physcial, they provide inspiration for very simple theories about the mental which make correct predictions of action in a limited but important range of circumstances.
One attempt to codify a the core part of a theory of the mental analogous to impetus mechanics is provided in Butterfill and Apperly's paper about how to construct a minimal theory of mind.
I can't explain it in detail here, but minimal theory of mind is like impetus mechanics. It's obviously flawed and gets things quite wildly wrong but still useful in a limited range of circumstances.
Butterfill and Apperly's minimal theory of mind identifies a model of the mental.
I'm not going to describe the construction of minimal theory of mind, but I've written about it with Ian Apperly and outlined the idea on your handout.
The construction of minimal theory of mind is an attempt to describe how mindreading processes could be cognitively efficient enough to be automatic. It is a demonstration that automatic belief-tracking processes could be mindreading processes.
For this talk, the details don't matter. What matters is just that it's possible to construct minimal models of the mental which are powerful enough that using them would enable you to solve some false belief tasks.
\section{Minimal theory of mind\citep{butterfill_minimal}} An agent’s \emph{field} is a set of objects related to the agent by proximity, orientation and other factors. First approximation: an agent \emph{encounters} an object just if it is in her field. A \emph{goal} is an outcome to which one or more actions are, or might be, directed. %(Not to be confused with a \emph{goal-state}, which is an intention or other state of an agent linking an action to a particular goal to which it is directed.) \textbf{Principle 1}: one can’t goal-directedly act on an object unless one has encountered it. Applications: subordinate chimps retrieve food when a dominant is not informed of its location;\citep{Hare:2001ph} when observed scrub-jays prefer to cache in shady, distant and occluded locations.\citep{Dally:2004xf,Clayton:2007fh} First approximation: an agent \emph{registers} an object at a location just if she most recently encountered the object at that location. A registration is \emph{correct} just if the object is at the location it is registered at. \textbf{Principle 2}: correct registration is a condition of successful action. Applications: 12-month-olds point to inform depending on their informants’ goals and ignorance;\citep{Liszkowski:2008al} chimps retrieve food when a dominant is misinformed about its location;\citep{Hare:2001ph} scrub-jays observed caching food by a competitor later re-cache in private.\citep{Clayton:2007fh} %,Emery:2007ze \textbf{Principle 3}: when an agent performs a goal-directed action and the goal specifies an object, the agent will act as if the object were actually in the location she registers it at. Applications: some false belief tasks \citep{Onishi:2005hm,Southgate:2007js,Buttelmann:2009gy}
Unlike the full-blown model, a minimal model distinguishes attitudes by relatively simple functional roles, and instead of using propositions or other complex abstract objects for distinguishing among the contents of mental states, it uses things like locations, shapes and colours which can be held in mind using some kind of quality space or feature map.
Let me put it another way. The canonical model of the mental is used for a wide range of things: its roles are not limited to predicton; instead it also supports explanation, regulation and story telling. In this way it's more like a myth-making framework than a scientific one, although this is rarely recognised.
A minimal model of the mental gets efficiency by being suitable only for prediction and retrodiction.
What about signature limits?
One signature limit on minimal models of the mental concerns false beliefs about identity. These are the kind of false belief Lois Lane has when she falsely believes that Superman and Clark Kent are different people: for the world to be as she believes it to be, there would have to be two objects rather than one; her beliefs expand the world.
(This is for illustrating mistakes about identity.) You might not realise that your bearded drinking pal ‘Ian’ and the author ‘Apperly’ are one and the same person.
[Explain why minimal models can't cope with false beliefs about identity.] Now on a cananical model of the mental, false beliefs involving identity create no special probelm. This is because the (Fregean) proposition that Superman is flying is distinct from the proposition that Clark Kent is flying. Different propositions, different beliefs. By contrast, a minimal model of the mental uses relational attitudes like registration; this means that someone using a minimal model is using the objects themselves, not representational proxies for them, to keep track of different beliefs. Consequently knowing that Superman is Clark Kent prevents a minimal mindreader from tracking Lois’ false beliefs about identity.
This is why false beliefs about identity are a signature limit of minimal models of the mental.

‘an impetus heuristic could yield an approximately correct (and adequate) solution ... but would require less effort or fewer resources than would prediction based on a correct understanding of physical principles.’

Hubbard (2014, p. 640)

\citet[p.\ 450]{kozhevnikov:2001_impetus}: ‘To extrapolate objects’ motion on the basis of physical principles, one should have assessed and evaluated the presence and magnitude of such imperceptible forces as friction and air resistance operating in the real world. This would require a time-consuming analysis that is not always possible. In order to have a survival advantage, the process of extrapolation should be fast and effortless, without much conscious deliberation. Impetus theory allows us to extrapolate objects’ motion quickly and without large demands on attentional resources.’
\citep[p.\ 640]{hubbard:2013_launching}: ‘prediction based on an impetus heuristic could yield an approximately correct (and adequate) solution [...] but would require less effort or fewer resources than would prediction based on a correct understanding of physical principles.’

‘To extrapolate objects’ motion on the basis of physical principles, one should have assessed and evaluated the presence and magnitude of such imperceptible forces as friction and air resistance operating in the real world. This would require a time-consuming analysis ... ... the process of extrapolation should be fast and effortless, without much conscious deliberation. Impetus theory allows us to extrapolate objects’ motion quickly and without large demands on attentional resources.’

Kozhevnikov and Hegarty (2001, p. 450)

Kozhevnikov & Hegarty (2001, figure 1)

simplified from Kozhevnikov & Hegarty (2001)

simplified from Kozhevnikov & Hegarty (2001)

Given that we can coherently make testable hypotheses about models, I want to finish by considering one for which a variety of evidence has recently been offered.

Hypothesis:

Some automatic belief-tracking systems rely on minimal models of the mental.

Prediction:

Automatic belief-tracking is subject to the signature limits of minimal models.

False-Belief:Identity task

adapted from Low & Watts (2013); Low et al (2014); Wang et al (2015)

There is some evidence that this prediction is correct. Jason Low and his collegaues set out to test it. They have now published three different papers showing such limits; and Hannes Rakoczy and others have more work in progress on this. Collapsing several experiements using different approaches, the basic pattern of their findings is this ...
Take non-automatic responses first; in this case, communicative responses. When you do a false-belief-identity task, you see the pattern you also find for false-belief-locations tasks. But things look different when you measure non-automatic responses ...

False-Belief:Identity task

adapted from Low & Watts (2013); Low et al (2014); Wang et al (2015)

The non-automatic responses all show the signature limit of minimal models of the mental. This is evidence for the hypothesis that Some automatic belief-tracking systems rely on minimal models of the mental.
I also hear that quite a few scientists have pilot data that speaks against this signature limit.
One particular task for future research will be to examine whether other automatic responses to scenarios involving false beliefs about identity, such as response times and movement trajectories, are also subject to this signature limit.

conclusion

In conclusion, I’ve argued that we should not assume without argument the dogma that there is just one model of the mental. Nor should we assume that there models of the mental form a family where one of the models, the best and most sophisticated model, contains all of the states that are contained in any of the models.
By analogy with physical cognition, it is coherent to hold that there are multiple, incommensurable models of the mental and that different mindreading processes rely on different models of the mental. [Incommensurable in the sense that no coherent model could contain all states that feature in any model.]
I’ve also introduced two hypotheses ...

1. Systems

Hypothesis:

There are ≥ two belief-tracking systems, one more automatic than another.

Prediction:

Automatic and non-automatic tracking can yield inconsistent responses to a single scenario.

2. Models

Hypothesis:

Some automatic belief-tracking systems rely on minimal models.

Prediction:

Automatic belief-tracking is subject to signature limits.

You might not yet accept the hypotheses I’ve offered; this is reasonable because there will soon be quite a bit more evidence that bears on them.
But what I hope you do accept is that, by analogy with physical cognition, everyone, regardless of their theoretical commitments, needs to face the question, How do mindreaders model minds? Without an answer to this question we have not understood the first thing about mindreading.
The Mindreading Puzzle

There is an age at which:

1. in performing an A-task, the child relies on a model of minds and actions not incorporating beliefs;

2. in performing false belief tasks which are not A-tasks, the child relies on a model of minds and actions incorporating beliefs; and

3. the child has a single model of minds and actions.

An A-task is any false belief task that children tend to fail until around three to five years of age.
Claim 2 is false: it’s not the child but a system in the child, and the model does not incorporate belief but a belief-like state (e.g. registration).
Claim 3 is false: the child’s responses to scenarios involving false beliefs do involve multiple models

‘representational changes in mental state understanding develop as a function of intersections between language, cognitive control, and early core knowledge.’

\citet[p.~613]{low:2010_preschoolers}

Low (2010, p. 613)

3-year-olds really don’t get false belief

3-year-olds’ performance is not simply the result of performance factors. After all, they succeed on structurally similar tasks about pretence, saying or desiring.
Further, changing the EF demands of a FB task doesn’t affect performance, and, conversely, nor do differences between cultures in EF affect FB performance. Consistently with this, success on FB predicts later social competence independently of EF.
But then what about the developmental relation between EF and FB performance? It's emergence not performance (?).

Low & Watt’s identity task

[This is here in case I have to explain Low et al’s identity task]

van der Wel et al non/-automatic

This is what you as subject see. Actually you can't see this so well, let me make it bigger.
This is what you as subject see. There is are two balls moving around, two barriers, and a protagonist who is looking on. Your task is very simple (this is the 'implicit condition'): you are told to track one of these objects at the start, and at the end you're going to have to use a mouse to move a pointer to its location.
This is how the experiment progresses.
You can see that the protagonist leaves in the third phase. This is the version of the sequence in which the protagonist has a true belief.
This is the version of the sequence in which the protagonist has a false belief. (Because the balls swap locations while she's not absent.') OK, so there's a simple manipulation: whether the protagonist has true or false beliefs, and this is task-irrelevant: all you have to do is move the mouse to where one of the balls is. Why is this interesting?

van der Wel et al (2014, figure 1)

Just look at the 'True Belief' lines (the effect can also be found when your belief turns out to be false, but I'm not worried about that here.) Do you see the area under the curve? When you are moving the mouse, the protagonist's false belief is pulling you away from the actual location and towards the location she believes this object to be in!

van der Wel et al (2014, figure 2)

Here's a zoomed in view. We're only interested in the top left box (implicit condition, participant has true belief). To repeat, When you are moving the mouse, the protagonist's false belief is pulling you away from the actual location and towards the location she believes this object to be in!

van der Wel et al (2014, figure 2)

Some processes involved in tracking others’ beliefs are automatic.

Using the same task, van der Wel et al also show that some processes are NOT automatic ...
\citep[p.\ 132]{Wel:2013uq}: ‘In support of a more rule-based and controlled system, we found that response initiation times changed as a function of the congruency of the participant’s and the agent’s belief in the explicit group only. Thus, when participants had to track both beliefs, they slowed down their responses when there was a belief conflict versus when there was not. The observation that this result only occurred for the explicit group provides evidence for a controlled system.’

van der Wel et al (2014, figure 3)

Let me emphasise this because we'll come back to it later:

‘they slowed down their responses when there was a belief conflict versus when there was not’

Knudsen & Liszkowski