Bruineberg and collaborators distinguish three philosophical positions about the status of Markov blankets in active inference modelling, namely: literalism, realism and instrumentalism (sect. 6.1). Literalism is the view that the world is fundamentally a Markov blanket; so, everything in it – including your brain, my tortoises, and the symbols you are currently reading on this page – is a Markov blanket or grounded in a Markov blanket. Realism concerns the relation of active inference models to the world. It says that, for any active inference model, there is some mapping between some of its theoretical posits and some worldly features. Particularly, it says that the Markov blankets posited by any of these models can be mapped onto some boundaries of some objects. Instrumentalism concerns the relation of Markov blankets in active inference models both to the world and to their users. It presumes that at least some active-inference models are useful for achieving some epistemic or pragmatic goal.

Though these -isms might come across as futility, they foreground very general questions about the point of the free-energy principle (FEP) and its relevance for understanding life and mind, the utility and correct way of interpreting active inference modelling, and the rational epistemic attitude towards the content of the theories and models grounded in the FEP.

Bruineberg et al. criticize literalism because it reifies abstract mathematical structures and is “removed from the empirical and naturalistic research programme that FEP purports to be” (sect. 6.1.2, para. 6). Yet literalists may reply they are engaged in revisionary metaphysics that makes no appeal to the supernatural. Noticing the FEP is a research programme in mathematical physics quite removed from empirical data (Colombo & Palacios, Reference Colombo and Palacios2021), literalists may point out reification (or Platonism) is widespread among mathematicians to make sense of their achievements, and may also draw an analogy with computation, showing how cellular automata have been stripped of their “metaphysical modesty” for arguing that the universe is fundamentally a cellular automaton (e.g., Wolfram, Reference Wolfram2002; Zuse, Reference Zuse1970).

Bruineberg et al. criticize realism because the Markov blanket formalism doesn't tell modellers how to find “a non-arbitrary mapping that is privileged for principled reasons” (sect. 6.1.2, para. 7). As this criticism accepts that idealized active inference models can be evaluated for their accuracy, realists may insist that somebody's finding a boundary (counter)intuitive has no bearing on when an active inference model is accurate. Insofar as a given active inference model is accurate and the Markov blankets it posits enjoy referential success, then the system being modelled does possess the properties to which the formal structure of the model successfully refers – whether counterintuitively or not.

Perhaps, Bruineberg et al.'s criticism is not that realism and literalism are counterintuitive or incoherent, but that these positions in the FEP literature are often based on bad (or no) arguments uninformed by relevant results in metaphysics and the philosophy of modelling about, say, the boundaries of objects (e.g., Varzi, Reference Varzi, Campbell, O'Rourke and Slater2011), the notions of structure, thing, and fundamentality (e.g., Sider, Reference Sider2020) or the bearing of imagination and fiction on scientific modelling (e.g., Levy & Godfrey-Smith, Reference Levy and Godfrey-Smith2019). Or perhaps, Bruineberg et al.'s criticism is that, to account for the achievements of active inference modelling, we should focus attention on the utility of such models rather than their accuracy; and to explain their utility, it's irrelevant whether Markov blankets are “fundamental,” “real,” or “fictitious.”

This last idea seemingly coheres with Bruineberg et al.'s recommended instrumentalism. But any plausible, instrumentalist position towards a scientific model is premised on the success of that model in furthering some scientific aim. The problem, in the context of the FEP, is that it's unclear how active inference modelling is successful.

It's uncontroversial that Markov blankets, and other statistical and algorithmic tools for causal search, discovery, and inference often play an incredibly useful role in helping scientists to represent and study systems of interest at a suitable scale, make reliable predictions, discover causal mechanisms, and facilitate interventions in the world (e.g., Marinescu, Lawlor, & Kording, Reference Marinescu, Lawlor and Kording2018; Spirtes, Glymour, Scheines, & Heckerman, Reference Spirtes, Glymour, Scheines and Heckerman2000). It's also uncontroversial that some tools such as the digital computer (and methods for inferential statistics) have historically inspired new theories like the computational theory of mind (Gigerenzer, Reference Gigerenzer1991). What is contentious is whether successful active inference models are nothing but instruments for prediction, control, or achieving some other aim, and whether the theoretical claims of these models constitute knowledge of the world (including of its unobservable aspects). Traditionally, instrumentalists don't limit themselves to make the banal claim that maths and stats are useful tools to do science. They want to make sense of successful scientific practices, accounting for what warrants scientists' reliance on empirically successful models in inquiry (cf. Psillos, Reference Psillos1999; Stanford, Reference Stanford, Sarkar and Pfeifer2006). So, any plausible instrumentalist position towards an active inference model, or FEP more generally, should presume the model under consideration is useful, successful, or furthers some scientific aim.

While computational neuroscientists, as well as modellers in other sciences have a diversity of aims (Kording, Blohm, Schrater, & Kay, Reference Kording, Blohm, Schrater and Kay2018; Potochnik, Reference Potochnik2017), widely shared modelling aims include empirical adequacy (Van Fraassen, Reference Van Fraassen1980, pp. 11–13), novel predictions (Lakatos, Reference Lakatos, Worrall and Currie1978, pp. 31–34), and guidance on how to intervene in the world (Cartwright, Reference Cartwright2007, Ch. 3; Woodward, Reference Woodward2003, pp. 7–9). If it's unclear that active inference models are empirically adequate, make novel predictions, and guide cognitive and life scientists to successfully intervene in the world, then there is no obvious question for instrumentalists (and scientific realists) to address about the epistemic status of these models.

In fact, when they describe Baltieri, Buckley, and Bruineberg's (Reference Baltieri, Buckley and Bruineberg2020) active inference model of the Watt governor, Bruineberg et al. give us reason to believe that active inference modelling is not a useful approach for studying and understanding any self-organizing system. As Bruineberg et al. themselves recognize, it's unclear “what can possibly be gained by thinking of the behaviour of a coupled engine-mechanical governor system in terms of perception-action loops under the banner of free energy minimization” (sect. 7, para. 2). If the utility and empirical successes of active-inference modelling are contentious, then an instrumentalist position towards the FEP and the modelling practices it grounds cannot be as unproblematic and uninteresting as Bruineberg et al. suggest.