I argue that the inference from predictive processing and the free energy principle to a form of substrate-dependence is not as straightforward as Seth suggests. Instead of justifying a form of substrate-dependence, it must presuppose that consciousness is substrate-dependent. As a remedy, I suggest further clarifying the difference between simulating and realising consciousness.
If the free energy principle and predictive processing support a form of substrate-dependence, they must specify what non-computational aspects are required to realise consciousness, and why. Seth formulates two core ideas that speak to this. One is that the mechanisms underlying conscious experience are, according to predictive processing, entangled with homeostatic processes that are characteristic of living organisms (sect. 4.1). The other is that, according to the free energy principle, the process of minimising variational free energy is identical to homeostatic processes (sect. 4.2).
One could object that this only contingently holds for conscious living organisms. Simply put, it is an evolutionary advantage if an organism can instantiate consciousness and cognition (in part) by redeploying mechanisms that it also uses to stay alive. But this leaves it open whether artificial systems, which are not bound by the constraints of natural selection, can realise consciousness in different ways – in particular, in ways that are less energy-efficient, but are devoid of the homeostatic processes found in living organisms.
Does the free energy principle suggest a stronger link between conscious experience and homeostatic processes? After all, if they are “in some physical sense […] the same thing” (sect. 4.2), then it might seem that you cannot implement one without the other. But Seth is careful to avoid biopsychism, i.e., the view that all living organisms are conscious (sect. 4.0). That is, some homeostatic processes are not identical to conscious experience.
Furthermore, it seems that the identity between homeostatic processes and the information processing associated with consciousness must be qualified: the free energy principle provides two equivalent descriptions of certain self-organising systems – one in terms of physical dynamics, the other in terms of a form of information-processing. With respect to systems for which this equivalence holds, the information processing (minimisation of free energy) underlying consciousness may be numerically the same processing that undergirds homeostatic processes. But we cannot infer that this must also be true for conscious artificial systems. This is because the computations that are implicated in minimising variational free energy can be performed using digital computers. Hence, assuming that these computational processes must be implemented in the same way as they are implemented in conscious living organisms, in order to realise consciousness, would be unwarranted. Consequently, we are still left with the question of whether performing the right computations is sufficient for replicating consciousness or not.
To make progress, we need to address the following questions: why might the way in which conscious living organisms implement the minimisation of free energy be relevant to realising consciousness? What are the relevant properties of conscious living organisms that distinguish simulating consciousness from realising consciousness?
As I argue in Wiese (Reference Wiese2024), conscious living organisms instantiate causal relations that are not instantiated by computer simulations of conscious agents (at least not when implemented in conventional hardware). The assumption that causal relations are relevant to realising (instead of merely simulating) consciousness is not specifically an assumption about realising consciousness. Rather, it is in line with more general considerations (which I will not recapitulate here) to the effect that a key difference between merely simulating a phenomenon and realising it lies in the respective forms of causation that are instantiated (see sect. 3.7). That is, a mere simulation will not replicate all relevant causal relations that are required for realising the phenomenon. In particular, one can argue that the (virtual) causation found in computer simulations is illusory, not real (Kohár, Reference Kohár2024). To be sure, a computer that implements a computational process will thereby also implement a causal process. But the computer will, in most cases, not thereby replicate the same causal relations that are instantiated in the simulated target.
Conway’s game of life (Gardner, Reference Gardner1970) provides a useful analogy. It can be regarded as a coarse-grained simulation of an evolutionary process, in which over- or underpopulation can cause the death of individuals. For instance, if a living cell in the game of life is surrounded by less than two or more than three living neighbours, it dies. Notably, these causal relations don’t exist in the computational implementation, because there are no direct causal interactions between the parts of the computer that represent the states of individual cells (for a similar point, see Kohár, Reference Kohár2024). This can be seen even more clearly when one imagines implementing the game by hand, using a game board and tokens. The “birth” and “death” of individual tokens are not caused by other tokens, but by the human being who is implementing the rules of the game.
Consciousness, by contrast, is real in the sense that it is causally efficacious. In a (conventional) computer simulation of conscious agents, causal interactions between the agents’ internal states and the environment are represented, but not replicated. In a robot (with conventional hardware), the information processing associated with consciousness is instantiated, and it is also causally relevant (especially for outward behaviour), but internal causal relations that are associated with consciousness (e.g., interactions between mental processes) will be much sparser or absent. To the extent that conscious experience depends on causal interactions between unconscious mental processes (e.g., processes that integrate sensory information from multiple modalities; see also sect. 4.5), even a robot (with conventional hardware) will at most simulate conscious experience.
In sum, the free energy principle can help to discern forms of causal relations that are characteristic of conscious living organisms, but are absent in mere simulations. That a difference in causation is relevant to the distinction between simulating and realising a phenomenon is an additional assumption, but one that can be justified independently of considerations about consciousness. Hence, the free energy principle does support the view that consciousness is substrate-dependent, without begging the question.
You have
Access
I argue that the inference from predictive processing and the free energy principle to a form of substrate-dependence is not as straightforward as Seth suggests. Instead of justifying a form of substrate-dependence, it must presuppose that consciousness is substrate-dependent. As a remedy, I suggest further clarifying the difference between simulating and realising consciousness.
If the free energy principle and predictive processing support a form of substrate-dependence, they must specify what non-computational aspects are required to realise consciousness, and why. Seth formulates two core ideas that speak to this. One is that the mechanisms underlying conscious experience are, according to predictive processing, entangled with homeostatic processes that are characteristic of living organisms (sect. 4.1). The other is that, according to the free energy principle, the process of minimising variational free energy is identical to homeostatic processes (sect. 4.2).
One could object that this only contingently holds for conscious living organisms. Simply put, it is an evolutionary advantage if an organism can instantiate consciousness and cognition (in part) by redeploying mechanisms that it also uses to stay alive. But this leaves it open whether artificial systems, which are not bound by the constraints of natural selection, can realise consciousness in different ways – in particular, in ways that are less energy-efficient, but are devoid of the homeostatic processes found in living organisms.
Does the free energy principle suggest a stronger link between conscious experience and homeostatic processes? After all, if they are “in some physical sense […] the same thing” (sect. 4.2), then it might seem that you cannot implement one without the other. But Seth is careful to avoid biopsychism, i.e., the view that all living organisms are conscious (sect. 4.0). That is, some homeostatic processes are not identical to conscious experience.
Furthermore, it seems that the identity between homeostatic processes and the information processing associated with consciousness must be qualified: the free energy principle provides two equivalent descriptions of certain self-organising systems – one in terms of physical dynamics, the other in terms of a form of information-processing. With respect to systems for which this equivalence holds, the information processing (minimisation of free energy) underlying consciousness may be numerically the same processing that undergirds homeostatic processes. But we cannot infer that this must also be true for conscious artificial systems. This is because the computations that are implicated in minimising variational free energy can be performed using digital computers. Hence, assuming that these computational processes must be implemented in the same way as they are implemented in conscious living organisms, in order to realise consciousness, would be unwarranted. Consequently, we are still left with the question of whether performing the right computations is sufficient for replicating consciousness or not.
To make progress, we need to address the following questions: why might the way in which conscious living organisms implement the minimisation of free energy be relevant to realising consciousness? What are the relevant properties of conscious living organisms that distinguish simulating consciousness from realising consciousness?
As I argue in Wiese (Reference Wiese2024), conscious living organisms instantiate causal relations that are not instantiated by computer simulations of conscious agents (at least not when implemented in conventional hardware). The assumption that causal relations are relevant to realising (instead of merely simulating) consciousness is not specifically an assumption about realising consciousness. Rather, it is in line with more general considerations (which I will not recapitulate here) to the effect that a key difference between merely simulating a phenomenon and realising it lies in the respective forms of causation that are instantiated (see sect. 3.7). That is, a mere simulation will not replicate all relevant causal relations that are required for realising the phenomenon. In particular, one can argue that the (virtual) causation found in computer simulations is illusory, not real (Kohár, Reference Kohár2024). To be sure, a computer that implements a computational process will thereby also implement a causal process. But the computer will, in most cases, not thereby replicate the same causal relations that are instantiated in the simulated target.
Conway’s game of life (Gardner, Reference Gardner1970) provides a useful analogy. It can be regarded as a coarse-grained simulation of an evolutionary process, in which over- or underpopulation can cause the death of individuals. For instance, if a living cell in the game of life is surrounded by less than two or more than three living neighbours, it dies. Notably, these causal relations don’t exist in the computational implementation, because there are no direct causal interactions between the parts of the computer that represent the states of individual cells (for a similar point, see Kohár, Reference Kohár2024). This can be seen even more clearly when one imagines implementing the game by hand, using a game board and tokens. The “birth” and “death” of individual tokens are not caused by other tokens, but by the human being who is implementing the rules of the game.
Consciousness, by contrast, is real in the sense that it is causally efficacious. In a (conventional) computer simulation of conscious agents, causal interactions between the agents’ internal states and the environment are represented, but not replicated. In a robot (with conventional hardware), the information processing associated with consciousness is instantiated, and it is also causally relevant (especially for outward behaviour), but internal causal relations that are associated with consciousness (e.g., interactions between mental processes) will be much sparser or absent. To the extent that conscious experience depends on causal interactions between unconscious mental processes (e.g., processes that integrate sensory information from multiple modalities; see also sect. 4.5), even a robot (with conventional hardware) will at most simulate conscious experience.
In sum, the free energy principle can help to discern forms of causal relations that are characteristic of conscious living organisms, but are absent in mere simulations. That a difference in causation is relevant to the distinction between simulating and realising a phenomenon is an additional assumption, but one that can be justified independently of considerations about consciousness. Hence, the free energy principle does support the view that consciousness is substrate-dependent, without begging the question.
Acknowledgements
I am grateful to Leonard Dung for feedback on a draft of this comment.
Financial support
This research received no specific grant from any funding agency, commercial, or not-for-profit sectors.
Competing interests
None.