We endorse the authors' claim that there is a gap between the Markov blanket qua statistical tool and biological and cognitive boundaries qua ontological structures in the world. We will offer an explanation for the gap's existence: It arises from the Markov blanket's inability to capture the dynamic process through which biological and cognitive agents create their own boundaries with the environment over time. Active inference presupposes the existence of a Markov blanket, but it is not envisioned as a process that actively generates the latter.

Biological systems actively produce their boundaries with the environment through autopoietic processes (Varela, Reference Varela1979). Autopoiesis refers to a network of processes that continually regenerates its components and constructs their own physical boundaries, which Varela and Maturana proposed as the essence of life and its autonomy (Varela, Reference Varela1979; Varela, Maturana, & Uribe, Reference Varela, Maturana and Uribe1974). For instance, a biological cell maintains its own identity distinct from the environmental medium with a membrane system constructed by a network of metabolic processes. Moreover, living beings maintain their bounded identity not only by exchanging energy and material through metabolism, but also by actively interacting with the environment over space and time. A good illustration is that of a single cell in a nutrient-poor environment that climbs up a glucose gradient to maintain its physical boundary with the environment, keeping its internal states within viable ranges (Egbert & Di Paolo, Reference Egbert and Di Paolo2009; Ikegami & Suzuki, Reference Ikegami and Suzuki2008; Suzuki & Ikegami, Reference Suzuki and Ikegami2009). A biological boundary, then, is not a given, but is actively defined by the system itself. The same principle applies across a wide variety of living systems, from single cell creatures to more complex, multicellular animals, which live sensorimotor lives (Thompson, Reference Thompson2007).

Cognitive systems likewise actively produce their boundaries through their interaction with the environment. We can see this in the case of extended cognition and mind (Clark, Reference Clark2008; Clark & Chalmers, Reference Clark and Chalmers1998), where cognitive boundaries extend beyond the biological body by incorporating environmental items as their constitutive parts. Cognitive extension is not a state upon which we stumble by chance; rather, it is a process we actively bring about, or “enact,” based on skills and habits cultivated over time (Miyahara & Robertson, Reference Miyahara and Robertson2021; Miyahara, Ransom, & Gallagher, Reference Miyahara, Ransom, Gallagher, Caruana and Testa2020). To illustrate, consider Otto from Clark and Chalmers’ (Reference Clark and Chalmers1998) famous thought experiment. Otto suffers a mild case of Alzheimer's disease and uses a notebook to compensate for his memory deficit. According to Clark (Reference Clark and Menary2010), Otto and his notebook exhibit a tight functional coupling with each other to constitute a unified cognitive system (Miyazono, Reference Miyazono2017) to the extent that they satisfy the following “trust and glue” conditions: (1) the resource (viz., the notebook) is reliably available and typically invoked; (2) any information thus retrieved is more or less automatically endorsed; and (3) information contained in the resource is easily accessible as and when required. Obviously, Otto will not meet these conditions merely by developing a memory problem. Rather, he would have to learn to use notebooks to complement his compromised cognitive capacities and continue to do so repeatedly until it became a habit for him to always carry around a notebook and use it for constant notetaking. The functional coupling is a product of Otto's active engagement with the notebook and his development of relevant skills and habits over time (which is why Otto's Markov blanket is malleable [Clark, Reference Clark, Metzinger and Wiese2017] or negotiable [Kirchhoff and Kiverstein, Reference Kirchhoff and Kiverstein2021]).

The main shortcoming of the Friston blanket approach concerns the relationship between action (i.e., active inference) and identity (i.e., the Markov blanket). In this approach, active inference depends upon the Markov blanket, but not the other way round. Biological and cognitive systems are defined by Markov blankets as boundaries with the external environment. These systems perform active inferences through looping interactions between sensory states, internal states, and active states defined by the Markov blanket to keep their internal parameters within viable bounds (Friston, Reference Friston2013). On the other hand, as we saw above, both biological and cognitive systems actively create and maintain their bounded identity by interacting with the environment. As Clark puts it: “Creatures like us […] are Nature's experts at knitting their own Markov blankets” (Clark, Reference Clark, Metzinger and Wiese2017, p. 14). To accommodate this within the free-energy principle (FEP) framework, we must conceive of active inference as playing an essential role in autopoiesis, that is, in creating and maintaining the system's bounded identity (cf. Kirchhoff, Reference Kirchhoff2018). In fact, Friston (Reference Friston2010) describes living systems as performing active inference to reduce sensory surprisal and consequently maintain its homeostasis. Nevertheless, on the FEP, active inference does not explicitly participate in the autopoietic formation of the boundary between the system and its environment, which defines the identity of living beings, that is, the Markov blanket. That is, the dynamic relationship between action and identity is missing in the Friston blanket approach that depicts Markov blankets not as a product, but only as a precondition of active inference (Friston, Reference Friston2013).

In short, the Friston blanket approach fails to identify the tailor who creates the boundaries. At most, Markov blankets coincide with the outcome of the boundary-making processes carried out by biological and cognitive agents. Markov blankets are tailored by statistical patterns but living agents do not outsource boundary-making: We actively weave our own boundaries with the world.