The fanciest sort of intentionality: Active inference, mindshaping and linguistic content

ABSTRACT

In this paper, I develop an account of linguistic content based on the active inference framework. While ecological and enactive theorists have rightly rejected the notion of content as a basis for cognitive processes, they must recognize the important role that it plays in the social regulation of linguistic interaction. According to an influential theory in philosophy of language, normative inferentialism, an utterance has the content that it has in virtue of its normative status, that is, in virtue of the set of commitments and entitlements that the speaker undertakes by producing this utterance. This normative status is determined by the normative attitudes shared by members of the utterer’s linguistic community. I propose here an account of such normative attitudes based on the ecological interpretation of the active inference framework. I explain how social normativity can be understood in that framework as the way in which members of a group shape their social niche to make it more predictable. Finally, I apply this account of social normativity to basic communicative practices, thereby explaining how social normative expectations can emerge to regulate these communicative practices, eventually leading to the institution of the sort of normative statuses constitutive of linguistic content.

KEYWORDS:

• Linguistic content
• meaning
• active inference
• social normativity
• mindshaping
• inferentialism

Acknowledgments

This work was supported by the SSHRC (Social Science and Humanities Research Council) under grant 767-2019-1371.

Disclosure statement

No potential conflict of interest was reported by the author(s).

Notes

1. The notion of content is often introduced by saying that the states or performances that have a content are those that have truth conditions (or more generally satisfaction conditions; see eg. Searle, 1983). While this way of introducing the notion of content is useful to provide an intuitive grasp on the notion, it can be misleading because it already presupposes, or at least suggests, a particular type of theory of what it is for a state or performance to have content (namely, a representationalist approach to content and a truth-conditional account of linguistic meaning). Another way to introduce the notion of content is to say that a state or performance has a content just in case it is inferentially articulated, that is, can stand in inferential relations with other states and performances. This alternative entry point into the notion of content is just as much suggestive of a particular theory of content (namely, an inferentialist account of content). As we will see, my own proposal has much more in common with this second proposition.

2. There is a distinction to be drawn between content and (literal) linguistic meaning (“I am here” uttered by different speakers has the same linguistic meaning but expresses a different content; Allan, 1986; Carston, 2008; Grice, 1957; Recanati, 2002), but I will not have the space to address this question in this paper. See Drobák (2020) for an account of this distinction in an inferentialist framework similar to the one I adopt below.

3. In his “private language argument”, Wittgenstein famously argued that the norms determining the correctness of a move are necessarily public, and therefore social, because if such norms could be private, there would be no way of distinguishing between a move that seems correct and a move that is correct according to these private norms (Wittgenstein, 1953, §244–271). If no such distinction can be made, we then lose an essential part of what it is for something to be a norm: that there can be correct and incorrect applications of it.

4. Gibbard (2012) affirms that the concept of meaning can be fully explained in normative terms, but leaves open the question of whether the property of meaning is also normative. I propose here an account according to which the property of meaning is constituted by normative statuses. However, this way of talking is misleading because, in the view expressed here, content attributions do not consist in the attribution of properties to an agent, but rather in the assignment of proprieties to this agent; see Brandom (1994, pp. 9–12).

5. If normative statuses depend on normative attitudes, and content is a normative status (as suggested by the strong normativity thesis), then content depends on normative attitudes; that is, a state or performance of an individual has a content and the content that it has in virtue of the fact that some members of its community adopt the appropriate normative attitudes toward them. The idea that content essentially depends on content attribution is characteristic of what has been called interpretationism, which is an approach to intentionality principally championed by Davidson (1973, 1974) and Dennett (1978); Dennett (1987) that states, very roughly, that something has a content insofar as it is correctly interpreted as having this content. Brandom’s approach therefore counts as a form of interpretationism thus broadly understood (Brandom, 1994, pp. 55–62).

6. Gallagher nevertheless takes issue with the fact that Brandom does not seem to recognize the intentional nature of lower-level behavior and cognitive phenomena exhibited by infants and animals (Gallagher, 2017, pp. 73–74). This criticism is justified, and our answer to it is similar to that of Gallagher, which is of recognizing a more basic form of intentionality, described in the skilled intentionality framework (Rietveld et al., 2018) as the skillful navigation of a field of affordance.

7. Kiverstein and Rietveld’s proposal seems to be an ecological-enactive reformulation of the classic strategy of explaining mental content on the basis of linguistic content. The opposite strategy, explaining linguistic content on the basis of mental content, is well represented by broadly Gricean approaches accounting for linguistic meaning in terms of the communicative intentions of the speaker (Grice, 1957; Neale, 1992; Scott-Phillips, 2014). In my view, both these strategies will fail if they assume that we have an independent grasp on either linguistic or mental content which do not recognize their social normative nature.

8. Brandom himself holds that normative attitudes cannot be reduced to non normative attitudes and dispositions, and that it is therefore “norms all the way down” (Brandom, 1994, p. 625). I do not think that this position necessarily follows from a view of content as determined by normative attitudes, and I instead believe that a coherent and plausible naturalist account of these normative attitudes can be derived from the active inference framework (see section Social normativity in active inference).

9. While some believe that the thesis of the normativity of content must be incompatible with naturalism to be interesting (Boghossian, 2005, p. 207–217; Whiting, 2007, p. 135), I believe that it is an interesting thesis in its own right, independently of whether normative statuses can be reduced to “natural” facts or not. Moreover, a naturalistic approach to the normative attitudes that are in my view constitutive of content could give us important insights into the empirical workings of linguistic interaction.

10. The approach to content adopted here is similar to the “two-stages” program proposed by Hutto, Myin and Satne to naturalize intentionality (Hutto & Myin, 2013, 2017; Hutto & Satne, 2015). At the first stage is a basic form of contentless intentionality, what they call Ur-intentionality, which consists in a form of directionality of organisms toward elements of their environment, but without specifying these elements under a mode of presentation (which is often taken to be an essential trait of content). The second stage is contentful intentionality, which is scaffolded on sociocultural practices and essentially corresponds to propositional attitudes and linguistic meaning. Although my approach is importantly similar to theirs, there are some differences concerning the details of the two stages. At the first stage, I prefer, to the notion of Ur-intentionality, the notion of an optimal grip on a field of affordances, as proposed in the ecological interpretation of the active inference framework (Bruineberg & Rietveld, 2014; Bruineberg, Kiverstein et al., 2018). I developed an account of communicative behavior based on this basic form of intentionality in previous work. At the second stage, I propose to give pride of place to the normative inferentialism developed by Brandom (as explicitly suggested by Varga, 2015) to account for the social practices that confer content on our communicative behavior. In this paper, I sketch an account of these social practices within the active inference framework.

11. Zawidzki additionally suggests that mindshaping should be seen as the key to the “human sociocognitive syndrome”, which is the set of sociocognitive features that differentiate human cognition and culture from animal cognition and culture (Zawidzki, 2013, ch. 1 and 2018, p. 735). I accept the view that human social cognition fundamentally has a mindshaping function, but I am not committed to the additional idea that mindshaping is limited to human social cognition. Indeed, there is some evidence that chimpanzee communities can exhibit some forms of rudimentary social normativity (Andrews, 2020; Fitzpatrick, 2020). The difference between eventual animal mindshaping and human mindshaping should accordingly be seen as a difference of degree.

12. It is clear that content attribution often also has an epistemic function. For instance, in communities where shared norms are well enforced, discovering what commitments are undertaken by an agent is a good way of knowing what the agent will do. Some have thus argued that content attribution has equally mindshaping (social-normative) and mindreading (epistemic) functions (Peters, 2019). Although I recognize that content attribution sometimes has this epistemic function, I maintain that the social-normative function is primary. Thanks to an anonymous reviewer for bringing this question to my attention.

13. It must be kept in mind that the notion of inference involved in the active inference framework (bayesian optimization in a generative model) has nothing to do with the inferences constitutive of linguistic meaning in normative inferentialism (socially regulated transitions between behaviors).

14. It must be noted that while agents attempting to minimize prediction error will have a preference for regular and predictable sensory input, the active inference framework must make room for some degree of interest in novelty and surprising stimuli (Kiverstein, Miller et al., 2019) if it wants to avoid the notorious “dark room problem” (Friston et al., 2012). This is compatible with the general preference for regular social stimuli that I believe to be generating social normativity. For instance, a ROE could be in force in a community even if there is sometimes some degree of tolerance, and even interest, for minor deviations to this ROE.

15. This close relationship between active inference and mindshaping is highlighted by Dewhurst in a paper on folk psychology and predictive processing: “this reversal of the usual way in which folk psychology is thought to operate, from mindreading to mindshaping, resembles the relationship between passive and active inference under predictive processing accounts. An explicit mental state attribution that leads to conformative behaviour could be seen as a kind of active inference, whilst folk psychology understood as mindreading is more akin to a form of passive inference” (Dewhurst, 2017, p. 8).

16. As remarked by an anonymous reviewer, non-human animals are also free energy minimizing systems that interact with other organisms, and yet we do not find in their behavior the sort of social normativity that I have argued is implied by free energy minimization in social contexts. What could explain such a difference? My suggestion is that, before normative practices were stabilized, human communities were already characterized by basic cooperative activities (see note 18), which implied frequent social interactions and selection for mechanisms of social tolerance and compliance (see, Zawidzki, 2013, ch. 4). The importance of social interaction (both in terms of frequency and impact on fitness) in the human ecological niche eventually made it crucial for them to track and regulate the regularities of their social environment. In sum, to the difference of humans, social interactions are not frequent and important enough in the ecological niche of (most) non-human animals (and the behavior of the other organisms that inhabit it is not malleable enough) for it to be worth it for them to integrate the regularities involved in these social interactions in their generative models and to regulate their social niche with regard to these regularities. As I point out in note 11, some argue that chimpanzee communities exhibit some rudimentary forms of social normativity. They might thus be somewhere on the path to the sort of social normativity that we find in human communities.

17. The ecological account of communicative behavior presented below is not strictly speaking necessary for the account of content introduced here. As long as we have a robust account of cooperative communication which does not rely on the notion of content, we can derive linguistic content by applying to it the social normative practices entailed by the active inference framework. Our ecological account of communication is just a prime candidate for such a role.

18. As such, my account of linguistic content presupposes joint action. This could spell trouble, because traditional analyses of joint action often rely on the capacity to entertain shared intentions, which are contentful mental states (Bratman, 1999, 2014; Searle, 1990, 2010; Tomasello, 2008, 2014). However, these analyses are often recognized as extremely cognitively demanding (Butterfill, 2012; Pacherie, 2013; Tollefsen, 2005). I therefore prefer more minimalist accounts (Vesper et al., 2010; Tollefsen & Dale, 2012; Butterfill, 2017; see also Cuffari et al., 2015), which do not presuppose contentful mental states and on the basis of which normative practices can emerge, in turn giving rise to contentful intentionality.

19. What follows is obviously not a story of how communicative practices in fact evolved to eventually contain normative linguistic practices. Such an evolution depends on a wide variety of contingent events and contextual environmental parameters that I cannot specify here. My model is rather a proof of concept showing that we can get basic forms of linguistic normativity if we make the sort of active social niche construction and mindshaping entailed by the active inference framework interact with our ecological account of cooperative communication.

20. It must be kept in mind that these linguistic descriptions of these various acts of communication come from our own language, that we use to characterize these basic communicative practices. The acts of communication themselves do not have any linguistic structure or content. They directly trigger changes in the shared field to alter the dynamics of the hunters' joint action in various ways, but do not aim to describe anything or say “that the mammoth is beside the river”, for instance.

21. This account of content, as a particular kind of inferentialism, entails that propositional content is in an important way fundamental, and that the content of subsentential expressions (i. e. concepts) is only derivative from the content of the propositions in which these concepts take place. This is because content is defined in terms of its inferential articulation, and it is primarily propositions, and not concepts, that take place in inferences and “make a move in the language game” (Brandom, 1994, pp. 79–84). This primacy of the propositional is not new (Frege, 1879, 1884; Wittgenstein, 1953), but it is opposed to an influential traditional approach in semantics which takes the meaning of concepts to be given independently of their role in propositions, and builds the meaning of propositions by combining these independently meaningful concepts (see, eg. Fodor, 1975, 1987). The opposite “substitutional” strategy starts with the inferential role of propositions, and derives the content of their constituting concepts by identifying the difference that their substitution makes to the inferential role of the propositions in which they appear (Brandom, 1994, ch. 6; see Tanter, 2022 for recent work on this topic). This substitutional method is akin to the “segmentation” process described by Mithen (2005, 2009) and Wray (1998, 2002, 2005), according to which the proto-languages used by early hominins, originally composed of “holistic” acts of communication (i. e. proto-sentences), were in time segmented in individually contentful parts (i. e. proto-words) contributing their individual meaning to the use of the sentences in which they appear.

22. In previous work, we have described how proto-acts of language can emerge from our ecological account of cooperative communication (Tison & Poirier, 2021b). However, strictly speaking, these acts do not constitute proper acts of language because they do not yet involve the sort of norms that can confer force to an act of communication: they are “utterances without force” (Moore, 2019). Only with the right kind of normative status can an act of communication be said to have the pragmatic significance of an act of language.

23. It is certainly possible that some crucial normative elements of our contemporary practices of assertion are still missing from this quite barebones account of assertion, and that, contra Brandom, the nature of assertion is not exhausted by the dual responsibility and authority that it involves. In that case, I would not be opposed to rather call the practices described above as practices of proto-assertion. However, I believe that extending my account to cover these missing normative elements would probably be feasible with the resources of the account of social normativity provided in this paper. The same can be said of the extension of the basic practices of challenge and justification sketched above to proper practices of giving and asking for reasons. Thanks to an anonymous reviewer for asking me to clarify this question.

24. There is another important difference between Brandom’s framework and the reconstruction that I propose here. In Brandom’s work, the practices of attribution of normative status seem to appear out of nowhere and there is no satisfying description of the cooperative communicative practices on which linguistic normativity is built (this point echoes Satne’s (2021) criticism of Brandom’s normative inferentialism). In our account, linguistic normativity is built on the basis of preexisting cooperative communicative interactions, as described in our ecological account of communication.

Additional information

Funding

This work was supported by the Social Sciences and Humanities Research Council of Canada [767-2019-1371].