We will examine the three main worries that have been raised so far that conveniently fall into the main categories of questions that guide a mechanist, what Bechtel has called looking down, around and up. We first discuss the decomposability worry that concerns Bechtel’s looking down and around, namely the possibility of decomposing higher-level phenomena into lower-level components organized in a certain way, which enactivists consider reductionist. We reply that such a view relies on a misconception of the mechanistic approach.
The second worry we turn to is related to the effect of higher-level emergent levels onto lower level constituents (looking up in the sense of situating the components inside a larger mechanism). Enactivists fear that a mechanistic approach precludes such top-down causality which they think essential to the autonomy (i.e. the non-reducibility) of the social level and the possibility of so-called ‘circular causation’. Most mechanists do indeed reject inter-level causation. We argue, however, that a recent proposal to explain how mechanism can allow for inter-level causation is convincing and fully compatible with enactivism.
Finally, the extended cognition worry has to do with looking up in situating the individual in a social and physical context. We contend that the core mechanistic literature does not seem to cover extended and supra-personal mechanisms, as required by enactivism. However, recent developments in the field show that there are ways to create such a possibility while staying true to both mechanistic framework and enactivism.
The decomposability worry
The main worry enactivists and other REC-ers seem to have with the mechanistic approach is that it allegedly views cognitive systems as decomposable or near-decomposable while in reality they are non-decomposable. For example, Lamb and Chemero (2014) argue that according to the mechanists, producing an explanation requires (1) “decomposition [that] involves developing a model of a system’s behavior by identifying discrete component parts and their linear, or weakly non-linear, interactions” and (2) “localization [that] involves mapping those discrete components and interactions onto features of a physical system” (pp. 809-810). What is often added to this charge is that such an explanatory strategy views cognitive systems as component-dominant, i.e. the behavior of the whole is a simple additive result of the behavior of its components, whose properties and functions are rigid and pre-determined (Favela 2015). Therefore, a single component can be analyzed in isolation as responsible for some particular capacity of the system. Applied to the brain, for example, it would mean that we can identify and localize particular brain modules responsible for particular cognitive tasks like vision, processing information about other agents, reading written text and so on. Taking out that part of the brain or disrupting it would mean that the whole system loses that particular capacity.
In an opposition to this view on the brain and cognition, REC-ers argue that in fact living cognitive systems are non-decomposable into components and interaction-dominant. That is, the behavior of the whole is more than a simple sum of the parts because interactions between parts are mostly non-linear, the behavior of each part dynamically depends on all other parts of the system and it is not possible to assign any specific task to any component. Therefore, interactions between components are more important than the components themselves (Richardson and Chemero 2014). Viewed through that lens, it is not possible to analyze the brain and cognition into separable modules. Removing a part of the brain will obviously lead to some loss of function. However, this is not because the brain lost a particular component that realizes a particular property but because the brain is then a different whole operating differently (Maturana 1980a). One should note that we have not claimed that enactivists or interaction-dominant explanations deny the existence of components altogether. What they do is deny identification of components and their contribution to the realization of the phenomenon as an important part of the explanation of this phenomenon. Differently put, the contribution of a component to the overall behavior of a system is not tractable or identifiable in terms of the taxonomy that is appropriate to describe the overall behavior of the system (much like this is the case with, say, the contribution of genes to the complex behavior of animals).
As can be seen from the comparison between a component-dominant and interaction-dominant view on the cognitive system, they are two extremes of a continuum of positions that might be held by supporters of REC. Rejecting a component-based explanation could mean rejecting explanations that (a) take into account parts only but not their configuration, (b) take into account only parts that interact linearly or (c) statically but not dynamically, (d) take into account all sorts of interactions in addition to parts but not the modulating effect of the environment.Footnote 5 It is unclear at present which of these options enactivists subscribe to. However, we can examine where the mechanistic approach places itself on this continuum.
If neural and cognitive systems are indeed non-decomposable and the mechanistic framework can only be applied to decomposable systems, then obviously enactivists cannot make use of it. However, these arguments betray a misunderstanding of the mechanistic framework and explicit dismissal of the new developments in this field.
First of all, mechanists explicitly argue against mere aggregation of components and place heavy emphasis on their organization (Wimsatt 1997). It is because the way parts are organized in space and time that they together can exhibit behavior that they cannot exhibit on their own. It is because the parts are on a lower level than the whole they comprise that they cannot have the same properties (the properties of the hydrogen and the oxygen atoms are clearly not the same as the properties of water molecules).
Second, there is no reason to suppose that only linear and sequential modes of organization are allowed in mechanisms. Especially when dealing with biological mechanisms, non-linear and cyclic modes are ever-present. Such a focus on biology has led mechanists to stress the necessity for dynamic mechanistic explanation because in a system organized non-linearly “the operations performed by parts of the mechanism vary dynamically, depending on activity elsewhere in the mechanism” (Bechtel 2011, p. 551). Therefore, an explanation has to include not just a static diagram of components and their organization but also a description of how the functioning of these parts is orchestrated in time.
Finally, Bechtel (1997), in response to REC challenges has argued that cognitive systems are likely to lie on a continuum between the extremes of non-decomposable and fully-decomposable. They are, instead, integrated systems. In such systems, it is still possible to identify components. However, their functions are not necessarily predetermined and fixed. Rather, their contribution to the operation of the whole might dynamically depend on other parts of the system, the larger context and be variable in time. It does not mean that when studying a mechanism for a particular phenomenon it is impossible to identify these contributions.
In reply to such arguments, Lamb and Chemero (2014) state that
If aneo-mechanist wishes to discard the condition of decomposability, then she does so at the cost of discarding the feature of neo-mechanistic explanations that makes them distinct from more general accounts of naturalistic explanation (p. 813).
We believe this is incorrect. What is distinctive about neo-mechanistic explanations is not decomposability understood by REC-ers to mean “decomposability into linearly interacting static components”. What is distinctive about neo-mechanistic explanation is (1) a concern for capturing different levels of the system and understanding the relations within and between levels and (2) a concern for a particular target phenomenon and the concrete working parts and operations that underlie it. The latter property makes mechanistic explanation different from different types of explanation, such as, for instance, dynamic explanation (which seems to be what Lamb and Chemero mean with ‘naturalistic explanation’) in which generality and an ability to subsume a variety of phenomena under a particular regularity is seen as a virtue. The former property distinguishes mechanistic explanation from a general REC view on inter-level relations that take us into the discussion of inter-level causality, to which we now turn.
The causality worry
Connected with the decomposability worry is another misgiving of enactivists about mechanisms, which we will call the ‘causality worry’. An important part of the enactivist framework is the so-called ‘circular causality’ that is allegedly operative in many enactive systems. The idea here is that the elements or components that make up a system ‘cause’ the emergence of properties at a higher level of aggregation that cannot be reduced to the component parts and their interactions. These emergent properties, in turn ‘cause’ specific effects at the component level, by ‘enslaving’ components and their properties, as it is called. The PC experiment is a case in point: the overall dynamics of the experimental set-up, including participants, involves the more frequent occurrence of avatar-avatar interactions. This is caused by the actions of individuals, but the overall dynamics of the whole system causes individuals to move their mouses such as to contribute to this effect. Although mechanistic explanations are keen on levels as well as on causation, causation between levels has not traditionally been part of the mechanistic picture (Bechtel 2008; Craver 2007a, 2007b).
Mechanists typically think of causation as an intra-level phenomenon. Inter-level relations are relations of constitution, according to them, and it would be wrong to put these on a par with causal relations. The problem is that on a mechanistic account higher-level phenomena – system S’s \({\Psi }\)-ing, say – are constituted by the causal interactions of components of a given mechanism – such as component C’s \({\Phi }\)-ing. This means that C is a part of S and that C’s \({\Phi }\)-ing is part of S’s \({\Psi }\)-ing. Thus, top-down causation in a mechanistic framework would seem to involve causal interactions between a whole and its parts. This is problematic because according to many, if C and S are related as part and whole, they cannot be related as cause and effect. Causes and effects are thought to be (i) wholly distinct, (ii) temporally asymmetric (causes precede effects) and (iii) unidirectionally dependent (effects depend on causes, but not vice versa). However, wholes and parts are (i) not wholly distinct, (ii) temporally coincidental, and (iii) dependent in a direction (wholes are constituted by parts, not vice versa), that is incompatible with causal dependency in top-down causation (where parts should depend on wholes). For reasons such as these, mechanists reject the idea that the constituent relations between levels leave room for causal relations.
The notion of constitution at play here is synchronic. Or better: it is a notion in which the diachronic nature of processes – whether at the component level or the system level – does not play an explicit role. It is for this reason that Kirchhoff (2015) has argued that the notion of constitution as employed by REC is radically different from the way analytic philosophers, including mechanists, use that notion. Constitution on REC accounts is essentially and fundamentally diachronic; it is the dynamic unfolding of interconnected lower-level processes that constitutes events at a higher level. The notion of constitution that Kirchhoff uses as a contrast class for this dynamic, diachronic constitution, though, is taken from the kind of analytical metaphysics that is not concerned with cognition or processes in the first place: Gibbard’s (1975) example of a piece of marble that constituted Michelangelo’s David is used as the main model. Though this model has been used to argue that persons are constituted by bodies (Rudder-Baker 2000), it should be clear that a constitution relation that is as static as the relation between a piece of marble and a statue cannot used as be a model for the way in which lower-level processes constitute higher-level cognition. Kirchhoff is right when he claims that diachronicity has been disregarded by mechanists. This is not because they think interconnected components of mechanisms are as static as pieces of marble. It is because they have failed to be explicit about the fact that these components are processes too.
This is exactly what Krickel (2017) has done in a recent paper. By doing so she has killed two birds with one stone: not only is her diachronic notion of constitution a plausible diachronic extension of the standard mechanistic picture drawn by e.g. Bechtel and Craver. More importantly for our discussion, she shows how a diachronic notion of mechanistic constitution, in which interconnected lower-level processes together constitute a higher-level process, makes room for inter-level causation. In order to see how, we first need to distinguish between two ways in which a system-level process can be subdivided in parts. Temporal parts of such a process are parts of the overall system-level process – they are time-slices of such a process. If the process is the process of a person dying – to use a sinister but simple example – a temporal part of it may be the moment in which a person looks shocked and brings his hands to his chest. Spatial parts are at the component-level; but like temporal parts they can occur during only a part of the overall system-level process. In the example of a dying person, the event of a heart that stops beating would be a case in point. Inter-level causation becomes possible, according to Krickel, because spatial parts of an overall process and temporal parts of such a process are not related as parts and wholes. Suppose that the event of diving in ice cold water and the event of ceasing to move are temporal parts of the process of some person’s dying, and that the event of a heart that stops beating is a spatial part of that overall process. The heart that stops beating is not related as a part to either the process of diving into the water or the process of stopping to move. If spatial and temporal parts of a single overall process are not related as parts and wholes, then they can be related as causes and effects: they are distinct and temporally related and they can have asymmetrical dependence relations. In our example it would mean that we can say that the diving in ice-cold water caused the heart to stop beating, which in turn caused the person to stop moving. And these causal relations are all constitutive of the overall process of dying.
Krickel’s mechanistic notion of inter-level causation fits our model of mechanistic enactivism (where the tooth-belt of global processes causes movements in the component cogs and vice versa). It also fits the enactivist diachronic/process view of constitution. In fact, prominent enactivists cite Krickel’s position with approval (Gallagher in press; Gallagher stops short of explicitly accepting diachronic mechanism but he certainly does not reject it).
The extended cognition worry
The third major worry enactivists could have about mechanisms is that they prevent cognition from being understood as extended, i.e. done not by the brain alone but rather by a brain-body-environment system. In the case of social cognition, it is rather an extended brain-body-environment-body-brain system (Froese et al. 2013).
That the worry is justified is illustrated by the following critique by Herschbach (2012). In his article on social cognition sub-titled “A mechanistic alternative to enactivism” (emphasis added), he very acutely points out that enactivists have not been very clear on what they mean by constitution in their claim that social interaction constitutes social cognition. Constitution is a part-whole relationship and if the claim is that supra-personal interaction constitutes individual cognition, then it is somehow a category mistake and a confusion of levels of organization.Footnote 6 On the other hand, if constitution is aimed at emphasizing the causal links between agents engaged in the interaction, then enactivists are committing a well-known coupling-constitution fallacy (Adams and Aizawa 2010). In this fallacy, frequently ascribed to proponents of extended cognition in general, one points out extensive causal coupling between a cognitive agent and some external factors and then concludes that therefore these factors are part of cognition. Such a conclusion is thought to be unwarranted because coupling and constitutive relations are in general not equivalent.
Herschbach proposes that adopting amechanistic framework can capture everything that enactivists want to say about social interaction without committing the fallacy. He states that the perceptual crossing example would be described by mechanists as
an autonomous social network composed of two interacting agents [with some emergent properties to be explained by] (a) decomposing the system into its parts – the agents and potentially other environmental objects – and determining how each part behaves, (b) examining how those parts are organized spatially and temporally to constitute the entire social network, and (c) determining how that network interacts with anything external to it (p. 482).
Enactivists probably would not find this description troubling. However, Herschbach moves on to claim that a mechanist would focus on the lower level of the individual agents and how their internal mechanisms are responsible for the scanning behavior observed in the experiment. This behavior is responsive to the kind of sensory input received by the agent (from another avatar vs their shadow). The main point of difference between enactivists and mechanists, according to Herschbach, is that while the former would like to say that such environmental input constitutes social cognition, the latter would say that only the agent-internal mechanism constitutes the phenomenon of interest (the scanning behavior) while the environmental input is merely an external influence on that mechanism. The mechanism succeeds only when situated in the appropriate social context of having contact with another agent. In conclusion, rather than talking of the constitutive role of the interaction, Herschbach suggests that the emerging pattern is to be explained by internal capacities and dynamics of the agents (this is the constitutive part) that are situated in the appropriate social environment that causally interacts with it.
Why would Herschbach say this? Are mechanists necessarily internalists with respect to cognition? They are not. While most mechanists say little about the issue of extended cognition, Zednik (2011), for one, has argued for a possibility of truly extended mechanisms. He argues that dynamical explanations “are well suited for describing extended mechanisms whose components are distributed across brain, body, and the environment” (p. 239). That is, body and its brain on the one hand and the environment on the other can be said to be the two working parts of the mechanism (see also Beer 2003). Following this idea, Rucińska (2016) adds that the said parts can be conceived in a non-representational manner to fit wider enactivist commitments, by focusing on the ‘know-how’ in the animal’s body and affordances as constituents of the environmental side of the equation.
The link between mechanism and internalism that Herschbach assumes is not inherent in mechanism but rests on a further assumption, not about the nature of explanation, but about the nature of cognition. Herschbach thinks that only parts that participate in a self-organized autonomous individual can be truly said to constitute cognition. He follows Bechtel (2009a) in this claim, who in turn argued that it is the autonomous living system that is the proper “locus of control”, differentiated from the environment, because it is the living system that needs to maintain itself as a unity in constantly changing external conditions. Thus, even if we were to regard the whole PC set-up as a large mechanistic system, it would simply not be a cognitive system, according to Herschbach.
The obvious thing to note here is that Herschbach replaces the enactive explanandum – the enactivist what-level description – with his own by switching from the phenomenon of interest being social interaction as a whole to the scanning behavior of the individual – which is the standard TCS what-level description. Like Bechtel, enactivists think that organisms are loci of autonomous control. However, they are autonomous in being operationally closed. This, however, applies not just to the bio-chemical processes of self-maintenance, but also to the closure of the sensorimotor loop of the organism. This loop is closed not to the environment but through the environment, which is merely an additional step in the loop, not an input or output external to the system (see Villalobos and Ward 2015, for a more detailed argument). The point here is that if the enactivist notion of autonomy did not allow for the role of the environment in the cognitive process, they could not coherently advance extended cognition type of claims.
Herschbach may reject this enactivist notion of environment-involving autonomy. But he cannot do this on the grounds that it precludes a mechanistic explanation. The fight between Herschbach and the enactivists is not at the how-level where mechanism is at home, but at the what-level.