Link Search Menu Expand Document

Conclusion: Six Questions



We started by asking six questions. In conclusion, we review discoveries about each.

In Which Domains Is There Substantial Evidence for a Two Systems Theory?

Substantial evidence is evidence from multiple studies from different labs using different approaches.

We have seen that there is substantial evidence for two systems theories of mindreading (in Mindreading: Signature Limits, and Development) and ethics (in Ethical Cognition). We have also seen some evidence for two systems theories of physical cognition (in Speed-Accuracy Trade-Offs (in Physical Cognition)). In no case is the evidence sufficient to entirely rule out alternative, one system theories.

There are also many other cases we did not consider (as listed in The Core Idea). In some of these cases a two systems theory is well established (e.g. memory, number and instrumental behaviour).

How Are the Two Systems Distinguished?

Our approach was to consider a stripped-down, core claim about processes which differ in how fast they are. Further respects in which systems can be distinguished—by appeal to automaticity, say—are captured by auxiliary hypotheses (see The Core Idea).

We saw that two systems for mindreading can be distinguished by (i) the different degrees to which they are automatic (see Mindreading: Automaticity) and (ii) the different models of minds and actions they employ (see Mindreading: Signature Limits, and Development).

Two systems for physical cognition can be distinguished by the range of models of the physical which can characterise their operations. One system appears limited to impetus mechanics while the other is more flexible. The more limited system also appears to be partly responsible for representational momentum and perhaps other broadly perceptual effects, and thereby to influence the overall phenomenal character of experiences (see Speed-Accuracy Trade-Offs (in Physical Cognition)).

In the ethical domain, two systems theories are wildly accepted but there is much uncertainty about what distinguishes them. In our view this is a significant open challenge (see Ethical Cognition).

What, If Any, Kind of Unity Is There Across Domains?

We did not identify substantial evidence for a hypothesis which generates readily testable predictions and could be used to characterise features of two systems in different domains.

Features commonly conjectured as common themes across domains include automaticity, informational encapsulation and domain specificity. We observed that there is not a lot of evidence to support these conjectures. For instance, evidence for or against automaticity is hard to identify in the domains of ethical and physical cognition.

Why Are There Two Systems?

  • speed–accuracy trade-offs: different challenges call for different trade-offs between speed and accuracy. Having more than one system allows for radically different trade-offs between speed and accuracy. (This was illustrated in The Core Idea.)

  • learning and development: having more than one system where the fast system is relatively unchanging over development can provide an optimal balance between reliably meeting everyday practical needs and making it possible to pursue learning where there is a high risk of error but also a large potential reward. (This was illustrated in Mindreading: Signature Limits, and Development.)

  • phylogeny and culture: the historical emergence of writing was a consequence of a slow system building on abilities made possible by some fast systems. As this suggests, there are some things best provided phylogenetically (or at least through learning processes that do not depend on large-scale cooperative cultural projects) and others that can be provided through large-scale cooperative cultural projects.

When, If Ever, Are Two Systems Better Than One?

If you are building a survival system you want quick and dirty heuristics that are good enough to keep it alive: you don’t necessarily care about the truth. If, by contrast, you are building a thinker, you want her to be able to think things that are true irrespective of their survival value. This cuts two ways. On the one hand, you want the thinker’s thoughts not to be constrained by heuristics that ensure her survival. On the other hand, in allowing the thinker freedom to pursue the truth there is an excellent chance she will end up profoundly mistaken or deeply confused about the nature of physical objects. If she turns to philosophy, she may even end up convincing herself that nothing exists apart from her. So you don’t want thought contaminated by survival heuristics and you don’t want survival heuristics contaminated by thought. Or if some contamination is inevitable, you at least want to limit it. This is beautifully achieved by giving your thinker two (or more) systems, one fast and the other slow. Providing, of course, that the two are not directly connected but rather linked only very loosely, via intentional isolators like metacognitive feelings.

How, If At All, Do the Two Systems Interact? What Are the Barriers to Interaction Between Them?

Because of how we characterised what it is for systems to be distinct, there is a tension between postulating two (or more) systems and postulating interactions between them. We suggested that the distinctness of systems consists in there being processes which differ in conditions which influence whether they occur, and which outputs they generate (in The Core Idea). As the scope for interaction increases, the grounds for distinguishing systems weaken.

In both mindreading and physical cognition, we saw that it is possible for distinct processes to yield incompatible outputs in response to a single stimulus. Importantly, in the case of mindreading we also saw that this can work both ways: there are situations in which fast processes support correct responses while slow processes support incorrect responses; and conversely (see Mindreading: Signature Limits, and Development). This suggests that there are barriers to interaction between systems. And perhaps that the representations they operate over are not inferentially integrated.

One conjecture, which we did not explore in depth, is that fast and slow processes differ in operating over representations which differ in format. This barrier to interaction may explain the lack of inferential integration.

We saw that it is possible for a fast process to influence a slow one indirectly and asynchronously if the fast system can modify the overall phenomenal character of experiences (see Speed-Accuracy Trade-Offs (in Physical Cognition)). This provides one model for understanding interactions between fast and slow systems.

It is also possible that metacognitive feelings provide a way for fast processes to influence slow processes synchronously (see Metacognitive Feelings: How Do Fast and Slow Processes Interact?).


automatic : On this course, a process is _automatic_ just if whether or not it occurs is to a significant extent independent of your current task, motivations and intentions. To say that _mindreading is automatic_ is to say that it involves only automatic processes. The term `automatic' has been used in a variety of ways by other authors: see Moors (2014, p. 22) for a one-page overview, Moors & De Houwer (2006) for a detailed theoretical review, or Bargh (1992) for a classic and very readable introduction
cognitively efficient : A process is cognitively efficient to the degree that it does not consume working memory and other scarce cognitive resources.
domain specific : A process is domain specific to the extent that there are limits on the range of functions its outputs typically serve. Domain-specific processes are commonly contrasted with general-purpose processes.
fast : A fast process is one that is to to some interesting degree cognitively efficient (and therefore likely also some interesting degree automatic). These processes are also sometimes characterised as able to yield rapid responses.
Since automaticity and cognitive efficiency are matters of degree, it is only strictly correct to identify some processes as faster than others.
The fast-slow distinction has been variously characterised in ways that do not entirely overlap (even individual author have offered differing characterisations at different times; e.g. Kahneman, 2013; Morewedge & Kahneman, 2010; Kahneman & Klein, 2009; Kahneman, 2002): as its advocates stress, it is a rough-and-ready tool rather than an element in a rigorous theory.
inferential integration : For states to be inferentially integrated means that: (a) they can come to be nonaccidentally related in ways that are approximately rational thanks to processes of inference and practical reasoning; and (b) in the absence of obstacles such as time pressure, distraction, motivations to be irrational, self-deception or exhaustion, approximately rational harmony will characteristically be maintained among those states that are currently active.
informational encapsulation : One process is informationally encapsulated from some other processes to the extent that there are limits on the one process’ ability to consume information available to the other processes. (See Fodor, 1983; Clarke, 2020, p. 5ff.)
intentional isolator : An event or state which links representations but either lacks intentional features entirely or else has intentional features that are only very distantly related to those of the two representations it links. Metacognitive Feelings and behaviours are paradigm intentional isolators.
metacognitive feeling : A metacognitive feeling is a feeling which is caused by a metacognitive process. Paradigm examples of metacognitive feelings include the feeling of familiarity, the feeling that something is on the tip of your tongue, the feeling of confidence and the feeling that someone’s eyes are boring into your back. On this course, we assume that one characteristic of metacogntive feelings is that either they lack intentional objects altogether, or else what their subjects take them to be about is typically only very distantly related to their intentional objects. (This is controversial---see Dokic, 2012 for a variety of conflicting theories.)
metacognitive process : A process which monitors another cognitive process. For instance, a process which monitors the fluency of recall, or of action selection, is a metacognitive process.
representational format : Format is an aspect of representation distinct from content (and from vehicle). Consider that a line on a map and a list of verbal instructions can both represent the same route through a city. They differ in format: one is cartographic, the other linguistic.
slow : converse of fast.


Bargh, J. A. (1992). The Ecology of Automaticity: Toward Establishing the Conditions Needed to Produce Automatic Processing Effects. The American Journal of Psychology, 105(2), 181–199.
Clarke, S. (2020). Cognitive penetration and informational encapsulation: Have we been failing the module? Philosophical Studies.
Dokic, J. (2012). Seeds of self-knowledge: Noetic feelings and metacognition. In M. J. Beran, J. L. Brandl, J. Perner, & J. Proust (Eds.), Foundations of metacognition (pp. 302–321). Oxford University Press Oxford, England.
Fodor, J. (1983). The modularity of mind: An essay on faculty psychology. Cambridge, Mass ; London: MIT Press.
Kahneman, D. (2002). Maps of bounded rationality: A perspective on intuitive judgment and choice. In T. Frangsmyr (Ed.), Le prix nobel, ed. T. Frangsmyr, 416–499. (Vol. 8, pp. 351–401). Stockholm, Sweden: Nobel Foundation.
Kahneman, D. (2013). Thinking, fast and slow. New York: Farrar, Straus; Giroux.
Kahneman, D., & Klein, G. (2009). Conditions for intuitive expertise: A failure to disagree. American Psychologist, 64(6), 515–526.
Moors, A. (2014). Examining the mapping problem in dual process models. In Dual process theories of the social mind (pp. 20–34). Guilford.
Moors, A., & De Houwer, J. (2006). Automaticity: A Theoretical and Conceptual Analysis. Psychological Bulletin, 132(2), 297–326.
Morewedge, C. K., & Kahneman, D. (2010). Associative processes in intuitive judgment. Trends in Cognitive Sciences, 14(10), 435–440.