Here, I use the following strategy to draw a distinction between simulating and being a conscious system. First, I highlight some characteristic features of conscious living organisms that conform to the FEP. Systems satisfying these features fulfill what I shall call the “FEP Consciousness Criterion” (FEP2C). Since FEP2C is based on consideration about conscious living organisms, we should not expect that artificial systems must satisfy FEP2C, in order to be conscious—just as we should not presuppose that having a neocortex or a biological nervous system is necessary for being conscious. A benefit of FEP2C is that it abstracts away from the underlying (biological) implementational details. Hence, FEP2C can be satisfied by non-biological artificial systems; but it is not satisfied by most current computers. This can be seen clearly by deconstructing FEP2C into a set of conditions entailed by this criterion. Furthermore, we can then determine to what extent it is plausible to regard these conditions as necessary for consciousness in artificial systems.
What does FEP2C consist in? Under the assumption that formulating a mechanical theory of consciousness is possible (Sect. 3) we can express the internal dynamics of conscious systems in two conjugate ways (Sect. 2). In other words, if we start with a description in terms of the probability of internal states (or paths), we can equivalently express the dynamics in terms of a probability distribution encoded by internal states (or paths). In doing so, we move from a description of a physical system to a description of a computational system that minimises variational free energy, with respect to an internally encoded probability density (generative model). For the sake of simplicity, call the former the physical dynamics, and the latter the computational dynamics.
If the FEP is correct, then a given physical dynamics uniquely specifies the corresponding computational dynamics. Crucially, the reverse does not hold. By map** internal states (or paths) to a probability density, information about some physical details is lost. This assumption is justified by theorems such as the slaving principle (Haken, 1977/2012) or the center manifold theorem (Carr, 1971/2012; Davis, 2006).
According to these theorems, trajectories of self-organising systems that are not in equilibrium with their environment unfold in a relatively low-dimensional manifold, compared to their high-dimensional state space. In the brain, this means that the activity of neural population can be described in terms of their ensemble properties (e.g., statistical averages, Friston et al., 2020). Random fluctuations at the level of individual neurons can be averaged out, because they do not influence the behaviour of the ensemble (Palacios et al., 2020).
In particular, this means that a relatively coarse-grained description of the computational dynamics does not uniquely specify the underlying physical dynamics. In principle, one could implement the computational dynamics of a conscious organism in a computer simulation. There would thus be a level at which both activity in a conscious organism and in a computer could be described as implementing variational free energy minimisation. The underlying physical dynamics, however, would in general differ dramatically.
This brings us to the “FEP Consciousness Criterion” (FEP2C). For all conscious living organisms that conform to the FEP, the following holds:
(FEP2C) The system‘s physical dynamics entail computational dynamics that include computational correlates of consciousness.
The FEP2C makes two claims about conscious organisms: (i) The physical dynamics of conscious organisms entail computational dynamics. (ii) These computational dynamics include computational correlates of consciousness. The first claim (i) directly follows from the FEP (see Sect. 2), together with the definitions of “physical dynamics” and “computational dynamics” provided at the beginning of this section: according to the FEP, we can interpret an organism’s physical dynamics as a process of variational free energy minimisation (which entails approximate Bayesian inference). This establishes the first claim (i).
To show that the second claim (ii) holds, I will make an additional assumption, viz. that consciousness contributes in some way to the sustained existence of conscious living organisms. In other words, I assume that consciousness has a function for conscious organisms, by regularly contributing to the goals of organisms (where I take it that staying alive is among the goals of organisms, see Piccinini, 2020, p. 68). For instance, consciousness might enable a cluster of learning abilities (Birch et al., 2020; Birch, 2022; Kanai et al., 2019) that make some contribution to staying alive. Of course, perhaps the same cognitive capacities can be realised without consciousness (“conscious inessentialism,” Flanagan, 1993). But I submit that it is nevertheless plausible to assume that consciousness has a function for conscious living organisms (even if we cannot rule out epiphenomenalism about consciousness).
In Sect. 2, I noted that the physical dynamics of a conscious organism that contribute to surprisal minimisation include the material realisers of consciousness, under the assumption that consciousness contributes to the organism’s survival. By reformulating these physical dynamics as a process of minimising variational free energy, we end up with a description of the organism’s computational dynamics. If the physical dynamics include the supervenience base of consciousness in that organism (which they will, if consciousness contributes to the organism’s survival), the computational dynamics will include computational correlates of consciousness. And that’s exactly what the second claim (ii) entailed by FEP2C says.
Recall that by “computational correlates of consciousness” I mean computational processes that correlate with consciousness in living organisms and can be formulated in terms of minimising variational free energy. These processes are sufficient for consciousness in living organisms (but not necessarily in artificial systems). Computational correlates are thus a particular form of computational dynamics. Let us call them “conscious* computational dynamics” (with an asterisk to indicate that the system instantiating these dynamics need not be conscious).
FEP2C is not fulfilled by current computers, even if they were to simulate a conscious system by instantiating conscious* computational dynamics. We can see this by deconstructing FEP2C into a set of conditions entailed by FEP2C. If a system, e.g., the computer in your office, fails to fulfill any of these conditions, it also fails to fulfill FEP2C (even if it instantiates conscious* computational dynamics). (In the following section, I discuss whether failure to fulfill a condition entailed by FEP2C gives us reason to infer the absence of consciousness.)
If a system S satisfies FEP2C, then:
-
[Implementation condition] S’s conscious* computational dynamics are strongly constrained by S’s hardware (or by the particular underlying mechanisms that implement these computations).
-
[Energy condition] The “thermodynamic cost of computation” paid by S is relatively low (compared to current computers).
-
[Causal-flow condition] The causal flow of S’s conscious* computational dynamics matches the causal flow of S’s physical dynamics.
-
[Existential condition] S sustains its existence (partly) by virtue of its conscious* computational dynamics.
I shall explain these conditions, and how they follow from FEP2C, in the remainder of this section.Footnote 6
4.1 The implementation condition
According to the implementation condition, an organism’s conscious* computational dynamics are, as it were, deeply tied to what it means to be that organism. What does this mean? In a nutshell: variational free energy is defined with respect to a generative model, the details of which depend on the particular organism. The organism’s computational dynamics are therefore shaped by the particularities of the organism.
As an analogy, consider how the same chord can be played using different analog instruments, but also using software synthesisers. When using an analog instrument, the sound of the chord will be shaped by the physical properties of the instrument. When the chord is played using a software synthesiser, different analog instruments can be emulated (to some extent), and so the sound of the chord is significantly less constrained by the physical properties of the machine on which the software is running.
Recall from Sect. 2 that the organism’s physical dynamics capture the characteristic features of the organism (e.g., human beings tend to be found in regions of their state space in which their body temperature is close to \(37\)° C). Put differently, to the extent that the organism displays these dynamics it “remain[s] the kind of thing that it is” (Ramstead et al., 2023, p. 3). According to the FEP, the organism’s physical dynamics can be redescribed as approximate Bayesian inference (by variational free-energy minimisation). That is, the FEP entails that the physical dynamics of living organisms can be interpreted as a form of computational dynamics. In doing this, we are just taking another perspective on the same dynamics. And since the physical dynamics capture what it means to be the organism, the same goes for the computational dynamics: the computations performed by the organism (i.e., those entailed by its physical dynamics) are deeply tied to what it means to be that organism.
In other words, the system’s hardware (the material basis of the computational correlates found in that system) puts strong constraints on its conscious* computational dynamics. This formulation is relatively vague. What does “strong” mean in this context? It may not be possible to quantify the strength of the constraints, but there is a clear qualitative difference between the way in which computational correlates are implemented by living organisms, and how they are implemented in current computers with a classical architecture. In current computers, there is a separation of software and hardware: the same software can be run on different tokens of the same type of hardware. This is extremely useful, because apps can be copied and installed on different computers, without having to modify the apps for each particular computer (as long as they are of the same type or use the same operating system). Once a large language model has been trained, its weights can be copied, and multiple instances of the same model can be run. The involved computational processes are “immortal”, because the same computational processes can be instantiated over and over again, in different pieces of hardware of the same type.
Hinton (2022) contrasts this form of computation with what he calls “mortal computation”. In mortal computation, the algorithms run by a given system are strongly constrained by the system’s particular hardware. This is the case for biological brains: even if you could record and copy the “connection weights” of my brain, trying to implement the same connection weights in your brain would be hopeless. Aside from practical complications, the individual differences between our brains (especially differences in connectivity) would make it impossible to instantiate the same computations, merely by “copying” connection weights.
Hinton (2022) suggests that allowing for differences between different tokens of the same type of hardware may reduce the cost of hardware production and save energy. In turn, every instance of a model would have to learn the model parameters that work for the particular piece of hardware on which it is running: “These parameter values are only useful for that specific hardware instance, so the computation they perform is mortal: it dies with the hardware.” (Hinton, 2022, p. 13).
Similarly, conscious organisms that satisfy FEP2C implement computational correlates of consciousness (conscious* computational dynamics) in a particular way. Recall that I am assuming that the computational correlates can be described in terms of minimising variational free energy. Variational free energy is defined with respect to a generative model, the details of which depend on the particular organism. There may be some general abstract properties shared by all conscious organisms, but the way in which the computational processes in a particular organism instantiate these properties differs from the way in which computational processes in other organisms instantiate them.
A stronger version of the implementation condition (which I am not presupposing here) would entail that the only way to implement computational correlates of consciousness is by using biological hard/wetware.Footnote 7 For instance, it can be questioned whether a digital computer could ever simulate the functional organisation of a human brain in real time; this puts pressure on the claim that the functional roles played by the neural realisers of consciousness are multiply realisable (Cao, 2022).
Here, I am not presupposing that consciousness depends on a functional organisation that is so tightly integrated with the properties of biological neurons, as to make it substrate-dependent. That is, I am remaining open to the possibility that consciousness is multiply realisable. The implementation condition only suggests that, within a given system, there is a tight integration between the physical properties of the material realisers of consciousness and the (computational) functional roles they realise.
That is, the implementation condition suggests that, although consciousness may be multiply realisable, there is a special way in which it is realised in conscious living organisms (just as a chord is realised in a special way by an analog musical instrument). Whether this is required for consciousness, or just a contingent fact about conscious living organisms, is a further question (to which I return below).
4.2 The energy condition
The energy condition is a corollary of the implementation condition. Above, I said that the computations performed by the organism (i.e., those entailed by its physical dynamics) are deeply tied to what it means to be that organism. This is because, according to the FEP, particular self-organising systems follow the “path of least surprisal” (Miller et al., 2022, p. 4) as long as they continue to exist, and pursuing the path of least surprisal can alternatively be described as variational free energy minimisation (and thereby as a process of approximate Bayesian inference). In other words, there are computations an organism “automatically” performs, simply by virtue of its continued existence. But this means that the energy an organism needs to sustain its existence includes the energy it needs to minimise variational free energy. Put differently, these computations “come for free”.Footnote 8 Hardware that uses mortal computation may have a similar benefit, which is why Hinton (2022) suggests that mortal computation might be the future of computing: “If you want your trillion parameter neural net to only consume a few watts, mortal computation may be the only option.” (Hinton, 2022, p. 13).
4.3 The causal-flow condition
Conscious living organisms that conform to the FEP instantiate conscious* computational dynamics not just in an efficient way. There is also, by assumption, a separation between internal and external states, and a circular causal flow between internal and external states, mediated by blanket states (i.e., perceptual and active states). Crucially, the internal states (or paths) that figure in the description of the physical dynamics are numerically identical with the internal states that figure in the description of the conjugate computational dynamics.
In general, such a match between the realisers of physical and computational dynamics cannot be taken for granted.Footnote 9 For the sake of illustration, assume that a computer with a von Neumann architecture can be described as a self-organising system that conforms to the FEP. Furthermore, assume that the computer simulates a system that satisfies FEP2C. This means the computer instantiates computational correlates of consciousness (which can be described in terms of minimising variational free energy). In particular, the computer must encode a probability density over some external states, given blanket states. Denote the states that encode the probability density with \({\mu }_{c}\). Here, the subscript “c” emphasises that these states are presupposed by the description of the computational dynamics that is simulated by the computer.
The computer’s physical states that represent \({\mu }_{c}\) are part of the computer’s memory. The computer simulates the computational dynamics by implementing a gradient descent on variational free energy. Hence, we can assume that the computations performed by the computer include those that are performed by the simulated conscious organism. But the way in which these computations are implemented differ in the following respect. Note that in computers with a von Neumann architecture, the central processing unit (CPU) is separated from the memory unit, and the memory unit stores both programme instructions and data. Since the states that encode \({\mu }_{c}\) are part of the computer’s memory, and since the computations that update the values of \({\mu }_{c}\) are performed in the CPU, these states never directly causally interact with states that represent the organism’s external, sensory, and active states.
To make it even more explicit, denote the simulated external states with \({\eta }_{c}\), and sensory and active states with \({s}_{c}\) and \({a}_{c}\), respectively. Because of the separation between CPU and memory unit, any causal influence of one data element (stored in the memory unit) on another data element must always be mediated by the CPU. Even if there are further memory units within the CPU, causal relations between elements of those memory units will always be mediated by other parts of the CPU, as well. That is, since a computer simulation must store the values of \({\mu }_{c}\), \({b}_{c}\) (comprising \({s}_{c}\) and \({a}_{c}\)), and \({\eta }_{c}\) in the memory unit, any causal relations between these representations is indirect, because it is mediated by the CPU. This differs from the basic causal flow between system states in the simulated conscious organism: in the simulated system, (some) external and internal states directly causally interact with (some) blanket states.Footnote 10
The difference in the basic causal flows is illustrated in Fig. 2.
4.4 The existential condition
Since the FEP analyses the concept of the existence of particular self-organising systems (Hohwy, 2021), it follows that being such a system entails minimising variational free energy. The computations such a system performs, which contribute to minimising variational free energy, therefore contribute to the sustained existence of the system. Put differently, the system exists (in part) by virtue of performing those computations. Notably, this does not mean that minimising variational free energy is sufficient for one’s continued existence. On the contrary, free energy minimisation is only necessary for the sustained existence of particular self-organising systems (including living organisms, Constant, 2021). But this means that, if a living organism exists for a certain period of time, we can (partly) explain this fact in terms of the computations it performed by virtue of existing (i.e., in terms of the computational dynamics entailed by its physical dynamics).
Contrast this with a simulation in a von Neumann computer, which may perform the same computations, by representing the organism’s states in its memory and by updating these representations in accordance with rules that specify how to minimise variational free energy. The relevant parts of the memory unit (or the whole computer) do not exist by virtue of their role in these computations.
Another way of expressing this is that living organisms give a damn (Haugeland, 2000), because they exist (in part) by virtue of performing certain computations. Since minimising variational free energy is necessary for survival, failing to do so (over a certain period of time) will lead to death. Hence, it matters what kinds of computations living organisms perform, their continued existence depends on it (again, this does not mean that performing the right computations is sufficient for survival, it is only necessary).
