Bruineberg et al. argue for a crucial distinction between inference with and within a model, with Pearl blankets pertaining to the former and Friston blankets the latter. However, any set of variables in a graphical model possesses a Pearl blanket (which therefore says nothing about system boundaries), while Friston blankets are taken to pick out living subsystems of a larger ecosystem. Unfortunately, Friston blankets have been applied almost as liberally as their statistical counterparts, including to individual neurons (Palacios, Isomura, Parr, & Friston, Reference Palacios, Isomura, Parr and Friston2019), body substructures such as the brain (Seth & Friston, Reference Seth and Friston2016), and eyes (Parr & Friston, Reference Parr and Friston2018) as well as larger organisms (Buckley, Kim, McGregor, & Seth, Reference Buckley, Kim, McGregor and Seth2017; Veissière, Constant, Ramstead, Friston, & Kirmayer, Reference Veissière, Constant, Ramstead, Friston and Kirmayer2019). This plurality of blankets is acknowledged by Parr (Reference Parr2020) and celebrated by Kirchhoff, Parr, Palacios, Friston, and Kiverstein (Reference Kirchhoff, Parr, Palacios, Friston and Kiverstein2018) as evidence for the ubiquity of the free-energy principle (FEP). We contend that this flexibility in what is cast as internal, external, sensory, or active states, is dangerously confused; it gives the false impression that the theory can recruit causal concepts, for example, Markov blankets, without committing to the full implications of a causal model-based understanding of perception and action.

The causal nature of the world is implicit in active inference, where sensory states are depicted as caused by external states that are, in turn, causally influenced by active states (Friston, Daunizeau, & Kiebel, Reference Friston, Daunizeau and Kiebel2009; Reference Friston, Mattout and Kilner2011). However, Friston et al. (Reference Friston, Daunizeau and Kiebel2009) propose that agents do not represent the world as such, but simply as a statistical coupling between the distribution of internal and external states through the blanket states. Worryingly, FEP theorists assume this is sufficient for agents to evaluate the consequences of their actions (Ramstead, Kirchhoff, & Friston, Reference Ramstead, Kirchhoff and Friston2020), and do everything else associated with cognition such as thinking, planning, imagining, and explaining (Sloman & Lagnado, Reference Sloman and Lagnado2015). While Ramstead et al. (Reference Ramstead, Kirchhoff, Constant and Friston2021) claim that the recognition density (the agents' approximate distribution over external states conditional on sensory states) represents the world, nothing is said about how this density encodes causal relations that are separable from actions and sensations. If the self-evidencing agent only represents relationships between their active and sensory states, and not the external world of causes that give rise to these, how can they arbitrate between inputs caused by their own actions and those that “would have happened anyway,” for example, those caused by ongoing dynamics out in the world? How too are they to do the myriad other things we associate with cognition?

In other words, active inference seems to conflate two different forms of inference. One is simply conditioning one's internal model on observations to update probabilities and make predictions. This includes both inferring likely consequences of observations – if the light turns on, we predict that the room is illuminated – but also their likely causes – that someone else must be home and have turned on the switch. A much-discussed limitation of such “passive” learning is that it struggles to answer questions about causal directionality (Bramley, Dayan, Griffiths, & Lagnado, Reference Bramley, Dayan, Griffiths and Lagnado2017; Lagnado & Sloman, Reference Lagnado and Sloman2004, Reference Lagnado and Sloman2006; Steyvers, Tenenbaum, Wagenmakers, & Blum, Reference Steyvers, Tenenbaum, Wagenmakers and Blum2003). Thus, a second form of inference is through active interventions, local alterations to the world that allow the learner to identify causal effects – for example, that the switch controls the light rather than the reverse. Clearly, if they then conclude that the light coming on means someone else is home, or that turning on the light would make someone else appear, they would have made a foundational mistake. Learning from intervention, or imagining actions, requires updating one's model in a more sophisticated way than simply conditioning on observations (Pearl, Reference Pearl2009). One must represent one's own action as coming from outside the system being modelled. This is a subtlety that active inference overlooks but one that humans are highly sensitive to (Bramley, Lagnado, & Speekenbrink, Reference Bramley, Lagnado and Speekenbrink2015; Bramley et al., Reference Bramley, Dayan, Griffiths and Lagnado2017; Bramley, Gerstenberg, Mayrhofer, & Lagnado, Reference Bramley, Gerstenberg, Mayrhofer and Lagnado2018, Reference Bramley, Gerstenberg, Mayrhofer, Lagnado and Kleinberg2019; Hagmayer, Sloman, Lagnado, & Waldmann, Reference Hagmayer, Sloman, Lagnado and Waldmann2007; Lagnado & Sloman, Reference Lagnado and Sloman2004; Rothe, Deverett, Mayrhofer, & Kemp, Reference Rothe, Deverett, Mayrhofer and Kemp2018; Sloman & Lagnado, Reference Sloman and Lagnado2005). Even rats are sensitive to the distinction between light or noise as signals (for food) or as consequences of their own action, that is, pressing a button (Blaisdell, Sawa, Leising, & Waldmann, Reference Blaisdell, Sawa, Leising and Waldmann2006; Clayton & Dickinson, Reference Clayton and Dickinson2006). To avoid interpreting the consequences of their own actions as signals for food, rats must treat themselves as independent from the light–food system. Critically, whether a sensory input is perceived as observational or interventional is agent-relative. One agent's intervention is, from the perspective of another agent, a worldly cause. This highlights that deciding what falls inside or outside a system's boundaries is a modelling choice that depends on the goal of the modeller and so does not resolve questions about actual physical boundaries.

To exhibit adaptive behaviour in a causal world, cognizers should not only approximate the expected observational distribution of external states but also the expected distribution under potential actions. This latter task requires that cognizers treat themselves as separate from the system they are learning about. To choose and evaluate the effect of its actions, an agent must perform inference with a model encoding asymmetric causal relations – in the sense that only actions on causes influence effects but not the reverse (Griffiths & Tenenbaum, Reference Griffiths and Tenenbaum2005, Reference Griffiths and Tenenbaum2009; Lagnado, Waldmann, Hagmayer, & Sloman, Reference Lagnado, Waldmann, Hagmayer, Sloman, Gopnik and Schulz2007; Tenenbaum, Griffiths, & Kemp, Reference Tenenbaum, Griffiths and Kemp2006) and should exhibit behaviour aimed at disambiguating these asymmetries. As such, we suggest that the notion of Markov blankets is critical to the agent's model of its own interactions with the world. In this sense, both the agent and the theorist describing it are performing inference with a model, and the cognition-relevant blankets are those that are properties of self-world representations rather than ontological features of living systems.

To sum up, we agree that casting behaviour as action–perception loops has yielded theoretical insights into self-regulatory (Barrett, Reference Barrett2017; Pezzulo, Rigoli, & Friston, Reference Pezzulo, Rigoli and Friston2015; Seth & Friston, Reference Seth and Friston2016) and habitual behaviour (Friston et al., Reference Friston, Rigoli, Ognibene, Mathys, Fitzgerald and Pezzulo2015, Reference Friston, FitzGerald, Rigoli, Schwartenbeck, O'Doherty and Pezzulo2016). However, we fear that inattention to causal representational structure means active inference suffers the same pitfalls as predictive processing (Sloman, Reference Sloman2013), and behaviourism before it, consigned to explain only simple autonomic or reflex behaviours and not those that make intelligent systems such fascinating and unique parts of the natural world.