Bruineberg and colleagues contend that so-called “Friston blankets” introduce a number of “non-arbitrary assumptions” in applying Markov blankets to the boundaries of living systems (sect. 4, paras. 1 and 2). The application of Markov blankets to living systems requires prior observations providing “a principled justification for why to start from one particular model rather than a different one” (sect. 6, para. 1). In this sense, they conclude, Markov blankets owe part of their explanatory power to these prior assumptions to a point where “it is not clear that the Markov blanket formalism is doing much additional work” (sect. 6.1.2, para. 7).

If the application of Markov blankets to living systems is indeed determined by such underlying assumptions, this would seem to imply that at least some of the confusions that Bruineberg and colleagues have set out to untangle run deeper than our attitudes toward Markov blankets.

If so, then a strong instrumentalism about Markov blankets may itself be insufficient as a measure to untangle the root causes of the confusions between realist and instrumentalist readings of Markov blankets (see Andrews [Reference Andrews2021] and Kirchhoff, Kiverstein, & Robertson [Reference Kirchhoff, Kiverstein and Robertson2022] for recent discussion of realism and instrumentalism qua free-energy principle [FEP] models). Besides the eternal vigilance demanded by our models and metaphors, we may need to reevaluate some of the starting observations informing their application.

The assumption that the organism and the environment constitute two conditionally independent interactants defines many Bayesian approaches to living systems, including the Fristonian one targeted by Bruineberg and colleagues. This guiding assumption behind designating living systems in terms of an inner organism contraposed by an outer environment may be interpreted as a variant of preformationism: The notion that organisms and environments constitute and should be evaluated in our theorizing as separate entities with inherent properties, and whose interaction is essentially secondary to their independent existence (see Anderson, Reference Anderson, Metzinger and Wiese2017; see also Oyama, Reference Oyama2000).

This assumption, of a pre-established conditional independence between organisms and their respective environments, presents a potential point of theoretical (Colombo & Wright, Reference Colombo and Wright2021) and empirical (Aguilera, Millidge, Tschantz, & Buckley, Reference Aguilera, Millidge, Tschantz and Buckley2021) incongruity between Markov blankets and the essentially coupled character of sensorimotor interfaces. It has moreover been brought into question by more recent accounts emphasizing the constitutive role that interaction plays in producing and sustaining the separate forms of organism and environment (see, e.g., Bruineberg & Rietveld, Reference Bruineberg and Rietveld2019; Gallagher & Allen, Reference Gallagher and Allen2018; Kirchhoff & Kiverstein, Reference Kirchhoff and Kiverstein2019, Reference Kirchhoff and Kiverstein2021).

We believe that the risk of preformationism echoes earlier debates within the literature in that it “force[s] us to recognize that the picture of biological agents as free-energy-minimizing systems requires something closer to a process-based (rather than a static or state-based) ontology” (Clark, Reference Clark, Metzinger and Wiese2017, p. 17). In this regard, existing accounts have already shown how temporally deep hierarchical models provide for adaptive models with less sharp distinctions between organisms and their environments (see Kirchhoff, Parr, Palacios, Friston, and Kiverstein, Reference Kirchhoff, Parr, Palacios, Friston and Kiverstein2018).

What we here want to briefly suggest is that a process-based perspective may furthermore avoid preformationism not only in the application of Markov blankets, but also at the level of the underlying assumptions that inform this application.

The sort of process-based perspective that we have in mind serves to preclude preformationism specifically by reconceptualizing stabilized forms on either side of the (Markov) boundary as products of ongoing exchanges that serve to perpetuate the living system. That is, under a process-based perspective on living systems, we may understand the organism and its respective environment not as a preformed substance but as an ensemble of processes (e.g., metabolism). The process view we refer to echoes the view of process ontology that takes processes – instead of substantive forms – as the fundamental unit of analysis in biology. Process ontology seeks to reverse the explanatory relation between entities and processes: Rather than explaining processes in terms of interactions between distinct entities, process ontology explains entities as relatively stable phases of continuous processes (Nicholson & Dupré, Reference Nicholson, Dupré, Nicholson and Dupré2018; see also Griffiths & Stotz, Reference Griffiths, Stotz, Nicholson and Dupré2018).

Narratively, as applied to active inference, a process-based perspective conceptualizes organismic boundaries as “hard-won achievements” of living systems (Kirchhoff & Kiverstein, Reference Kirchhoff and Kiverstein2019; see also Kirchhoff, Reference Kirchhoff2015; Sutton, Reference Sutton and Menary2010). This reversal is decisive for at least one of the underlying assumptions that Bruineberg and colleagues ascribe to Friston blankets: It eliminates the need for the assumption of a preformed organism qua model and environment qua modeled distal world, which arguably commits Friston blankets (and other Bayesian accounts) to a particular variant of substantialist realism. In its stead, processes are what is taken to be the fundamental unit of biological analysis. Under a process-based view, then, one need not assume the organism and environment since these may be derived from the continuous exchanges.

While Bruineberg and colleagues' prescribed strong instrumentalism might still furnish us with helpful resources for clearing up confusions surrounding the application of Markov blankets to living systems, we find that some such confusions may still be traced to the prior observations that inform this application. We believe that a process-based perspective may aid us in upending a central assumption that prefigures some of the forms of confusion targeted by Bruineberg and colleagues. While a far cry from absolving us of the duty to attend to other crucially important issues pointed out by Bruineberg and colleagues in their insightful target article, we nonetheless believe that critically assessing the starting assumptions underlying these issues may ultimately prove to be indispensable in their resolution.