In this commentary, I discuss what the use of computational models to simulate the dynamics or time evolution of some phenomena – here, consciousness – entails for our understanding of those phenomena. I argue that being able to simulate the dynamics of a phenomenon using computational models does not entail that that phenomenon is literally performing the computations that are prescribed by the model.

I applaud Seth for his compelling discussion of artificial consciousness, computational functionalism, and biological naturalism. In his target article, Seth argues that for artificial intelligence systems to be conscious, two conditions must hold. First, computational functionalism must be true of consciousness. That is, it must be possible to define the states and processes associated with consciousness in terms of their functional relations to other states, and it must be possible to formalise these functional relations and their dynamics using the theory of computation (Putnam, Reference Putnam, Capitan and Merrill1967, Reference Putnam1975). Here, consciousness is not merely modelled as a computation but is equated to computation, metaphysically speaking. That is, in this view, consciousness is literally a form of computation. Second, the computations associated with consciousness must be characterised by multiple realisability or substrate flexibility (Piccinini & Maley, Reference Piccinini, Maley, Sprevak and Kallestrup2014; Polger & Shapiro, Reference Polger and Shapiro2016), which denotes different versions of the idea that computations must not depend, or not depend so tightly, on the physical properties of the medium that implements them, such that they can be implemented in different media (i.e., in biological neuronal networks as well as in silicon-based computing systems). Hence, for a conscious computing artefact to exist, the computations associated with consciousness must at minimum be implementable in silico.

In the article, Seth reviews “negative” arguments against the possibility of conscious artificial intelligence, arguing from the implausibility of computational functionalism to that of conscious computing artefacts. Here, I make a complementary point that is independent of the truth of computational functionalism. I point out that one can use computational tools to study and simulate what are – at least on the face of it – non-computational phenomena. For example, when one simulates the orbit of Jupiter around the sun, one does not usually assume that either Jupiter itself or its orbit is literally computing an integral. The general point is that the fact that some phenomenon can be simulated or modelled using computational tools says precisely nothing about the metaphysical status of the phenomena being simulated. In particular, it provides no evidence at all for the idea that these phenomena are themselves literally performing computations.

Analogously, in the computational neuroscience of consciousness, we routinely leverage computational tools to model conscious processes. Here, it is crucial to distinguish computational neuroscience, defined as the study of the structure, function, and dynamics of the brain using computational modelling and simulation methods, from the Computational Theory of Mind, which is a metaphysical position that equates cognition and computation (Fodor, Reference Fodor1975). Their difference matters for the possibility of conscious artefacts.

To say that a physical system, such as a digital computer or a brain, implements a given computation is to say that there exists a systematic mapping between two descriptions of the system: On the one hand, a dynamical description of some phenomenon and the observables or data that it generates, and on the other hand, the set of algorithmic steps required to perform that computation (Horsman et al., Reference Horsman, Stepney, Wagner and Kendon2014). In computational neuroscience, this relationship enables a form of analogical explanation (Chirimuuta, Reference Chirimuuta2024), in which we explain a phenomenon that we are trying to understand – the structure, function, and dynamics of the brain – in terms of an engineered artefact that we already understand – a designed artefact, the computing device. Mathematical frameworks like nonequilibrium statistical mechanics (González & Davis, Reference González and Davis2017; Niven, Reference Niven2010) and the free energy principle (Friston, Reference Friston2019; Ramstead et al., Reference Ramstead, Sakthivadivel, Heins, Koudahl, Millidge, Da Costa, Klein and Friston2023) can be used to determine the mathematical and physical conditions under which these descriptions are in fact equivalent or interchangeable. But this is an epistemic equivalence, defined and operative at the level of our descriptions of phenomena. It is not a metaphysical statement about whether the system or phenomenon being described is computational, metaphysically speaking.

I believe that the plausibility of the idea that it is possible to design conscious computing artefacts derives in part from not having properly made the distinction between computational modelling and computational metaphysics. The implication for the existence of conscious artificial intelligence is to disqualify a popular line of evidence for its existence: Even if we did in fact have access to a complete and final computational model of an episode of conscious experience, or even consciousness generally, we nonetheless would have no guarantee that the dynamics or processes being modelled or simulated are themselves computational in nature. In other words, the availability of computational models of consciousness provides no evidence for computational functionalism about consciousness, and hence no evidence in favour of conscious artificial intelligence.

In summary, the fact that we have access to computational models that seem to simulate aspects of consciousness provides no evidence that these models are themselves consciously experiencing anything. This point dovetails with Seth’s “negative argument” against the plausibility of computational functionalism.