A central premise of Seth’s argument is that, because brains are complex living systems that do not seem to resemble today’s (non-living) silicon-based computers, there is no reason to believe that computers are capable of consciousness. In this account, consciousness may not be computational and/or may be substrate-dependent, i.e. may only be realizable by living systems.
We believe that this account takes too narrow a view of computation, misconstrues the nature of platform independence or multiple realizability, and doesn’t acknowledge the fundamentally computational nature not only of the brain, but of life itself.
Multiple mathematical definitions of computation have been advanced, beginning with Church (Reference Church1936) and Turing (Reference Turing1936), later shown to be equivalent (Turing, Reference Turing1937). Another definition is von Neumann’s Universal Constructor (von Neumann & Burks, Reference Neumann and Burks1966). Von Neumann’s version differs from Turing’s in defining computation as embodied: while the Turing Machine consists of a head that can read and write symbols on a tape, the Universal Constructor is a cellular automaton, a machine made out of the very symbols it can read and write. Von Neumann’s construction thus has closure with its physical universe, whereas Turing’s construction does not.
Despite this important conceptual difference, though, von Neumann’s theory of computation is mathematically equivalent to that of Turing and Church, in that one can emulate a Universal Turing Machine with a Universal Constructor, or vice versa. We can therefore characterize physical processes as “computation” as long as the functions that describe those processes are computable by either of these mathematically equivalent definitions.
This is highly relevant to autopoietic systems. Per Kleene’s Second Recursion Theorem (Kleene Reference Kleene1938), in any Turing-complete system, there exist programs that can compute a function while also having access to their own source code. This allows for the creation of self-replicating programs. Building on this concept, Von Neumann realized that heritable, embodied self-replication was required for Darwinian evolution; thus, living systems must include a Universal Constructor to grow, maintain themselves, or replicate. Since a Universal Constructor is a computer, it follows that any system capable of autopoiesis must be computational. This is why DNA, RNA, ribosomes, etc., can be used to carry out arbitrary computations (Adleman Reference Adleman1994): Life is inherently computational.
This brings us to what we believe to be Seth’s fundamental error: when we examine the processes that he correctly links to consciousness (prediction, homeostasis, and autopoiesis), we see that these processes can only be implemented by physical systems that can be characterized as computers engaged in computation.
Of course, each of these processes has different relationships to the question of closure with their physical environment. Prediction can be disembodied: “simulating” predicting is predicting, just as a “simulation” of a calculator that tells us 2 + 2 = 4 is a calculator, having computed the same function. Homeostasis is more embodied, as it demands the ability to modify some physical variable, e.g., temperature or blood-glucose level. Autopoiesis is at the extreme end of embodiedness. Hardware cannot fix itself or replicate without a Universal Constructor. But autopoiesis is still computation.
Given these considerations, Seth’s theory of consciousness is in fact a computational theory of consciousness, even if he doesn’t wish to apply this word.
Some confusion arises here due to the translation of “substrate independence” from the realm of pure mathematics to engineering practice and linguistic usage. In engineering, computational substrates matter. For example, without GPUs and other massively parallel hardware, modern AI would be impossible. This is why neural network-based AI didn’t take off until the 2000s despite having first been attempted in the late 1950s.
The linguistic issue is more insidious. One could assert that even if an artificial kidney successfully filters urea, it isn’t really a kidney. Alternatively (and, perhaps, more usefully), one could define a kidney functionally. However, we should not confuse this debate over terminology with anything more profound. Philosophers Susan Schneider and David Chalmers run afoul of this error when they speculate that computation on a silicon platform might be functionally identical to neural computation in the brain, but the silicon version might be a kind of “philosophical zombie” lacking any subjective consciousness (Schneider, Reference Schneider2021; Chalmers, Reference Chalmers2022).
This is dubious. The very nature of life is functional – from molecules, to cells, to organs (including the brain), to whole organisms, and beyond. Every living system, or part thereof, has inputs and outputs, and evolution constantly “invents” alternative implementations: consider aerobic vs. anaerobic respiration, or the way sophisticated camera eyes arose in cephalopods and in vertebrates, or the “multiple discovery” of wings in insects, birds, and bats. These all have technological analogs too: fuel combustion, digital cameras, and airplanes. Both biological and technological systems can be characterized as evolving ecologies of functional relationships, where the substitution of one implementation or substrate for another is commonplace.
We hypothesize that consciousness, like flight, is functional, not only in the mathematical or computational sense described above, but also in the sense that it serves important functions for intelligent social agents. We do not see any evidence to suggest that consciousness must involve an inherently embodied form of computation, like the filtering of urea, though we remain open to the possibility. As far as we understand, though, consciousness supports the ability to look after oneself and plan for the future, as Seth implies with the nod to homeostasis, as well as supporting introspection, theory of mind, empathy, and social cooperation (Graziano Reference Graziano2013).
Like prediction, these are all computations that can be carried out purely in the information domain, implying that there is no impediment to the realization of consciousness in silico; a perfect “simulation” of consciousness is consciousness, just as a perfect “simulation” of a calculator is a calculator. We would even speculate that some current AI systems, given their abilities to predict, model themselves and others, and self-correct, may already be partly conscious. We do not claim certainty either way. But we certainly do not share Seth’s confidence that “real” artificial consciousness is unlikely along current trajectories. On the contrary, we think it is likely.
You have
Access
A central premise of Seth’s argument is that, because brains are complex living systems that do not seem to resemble today’s (non-living) silicon-based computers, there is no reason to believe that computers are capable of consciousness. In this account, consciousness may not be computational and/or may be substrate-dependent, i.e. may only be realizable by living systems.
We believe that this account takes too narrow a view of computation, misconstrues the nature of platform independence or multiple realizability, and doesn’t acknowledge the fundamentally computational nature not only of the brain, but of life itself.
Multiple mathematical definitions of computation have been advanced, beginning with Church (Reference Church1936) and Turing (Reference Turing1936), later shown to be equivalent (Turing, Reference Turing1937). Another definition is von Neumann’s Universal Constructor (von Neumann & Burks, Reference Neumann and Burks1966). Von Neumann’s version differs from Turing’s in defining computation as embodied: while the Turing Machine consists of a head that can read and write symbols on a tape, the Universal Constructor is a cellular automaton, a machine made out of the very symbols it can read and write. Von Neumann’s construction thus has closure with its physical universe, whereas Turing’s construction does not.
Despite this important conceptual difference, though, von Neumann’s theory of computation is mathematically equivalent to that of Turing and Church, in that one can emulate a Universal Turing Machine with a Universal Constructor, or vice versa. We can therefore characterize physical processes as “computation” as long as the functions that describe those processes are computable by either of these mathematically equivalent definitions.
This is highly relevant to autopoietic systems. Per Kleene’s Second Recursion Theorem (Kleene Reference Kleene1938), in any Turing-complete system, there exist programs that can compute a function while also having access to their own source code. This allows for the creation of self-replicating programs. Building on this concept, Von Neumann realized that heritable, embodied self-replication was required for Darwinian evolution; thus, living systems must include a Universal Constructor to grow, maintain themselves, or replicate. Since a Universal Constructor is a computer, it follows that any system capable of autopoiesis must be computational. This is why DNA, RNA, ribosomes, etc., can be used to carry out arbitrary computations (Adleman Reference Adleman1994): Life is inherently computational.
This brings us to what we believe to be Seth’s fundamental error: when we examine the processes that he correctly links to consciousness (prediction, homeostasis, and autopoiesis), we see that these processes can only be implemented by physical systems that can be characterized as computers engaged in computation.
Of course, each of these processes has different relationships to the question of closure with their physical environment. Prediction can be disembodied: “simulating” predicting is predicting, just as a “simulation” of a calculator that tells us 2 + 2 = 4 is a calculator, having computed the same function. Homeostasis is more embodied, as it demands the ability to modify some physical variable, e.g., temperature or blood-glucose level. Autopoiesis is at the extreme end of embodiedness. Hardware cannot fix itself or replicate without a Universal Constructor. But autopoiesis is still computation.
Given these considerations, Seth’s theory of consciousness is in fact a computational theory of consciousness, even if he doesn’t wish to apply this word.
Some confusion arises here due to the translation of “substrate independence” from the realm of pure mathematics to engineering practice and linguistic usage. In engineering, computational substrates matter. For example, without GPUs and other massively parallel hardware, modern AI would be impossible. This is why neural network-based AI didn’t take off until the 2000s despite having first been attempted in the late 1950s.
The linguistic issue is more insidious. One could assert that even if an artificial kidney successfully filters urea, it isn’t really a kidney. Alternatively (and, perhaps, more usefully), one could define a kidney functionally. However, we should not confuse this debate over terminology with anything more profound. Philosophers Susan Schneider and David Chalmers run afoul of this error when they speculate that computation on a silicon platform might be functionally identical to neural computation in the brain, but the silicon version might be a kind of “philosophical zombie” lacking any subjective consciousness (Schneider, Reference Schneider2021; Chalmers, Reference Chalmers2022).
This is dubious. The very nature of life is functional – from molecules, to cells, to organs (including the brain), to whole organisms, and beyond. Every living system, or part thereof, has inputs and outputs, and evolution constantly “invents” alternative implementations: consider aerobic vs. anaerobic respiration, or the way sophisticated camera eyes arose in cephalopods and in vertebrates, or the “multiple discovery” of wings in insects, birds, and bats. These all have technological analogs too: fuel combustion, digital cameras, and airplanes. Both biological and technological systems can be characterized as evolving ecologies of functional relationships, where the substitution of one implementation or substrate for another is commonplace.
We hypothesize that consciousness, like flight, is functional, not only in the mathematical or computational sense described above, but also in the sense that it serves important functions for intelligent social agents. We do not see any evidence to suggest that consciousness must involve an inherently embodied form of computation, like the filtering of urea, though we remain open to the possibility. As far as we understand, though, consciousness supports the ability to look after oneself and plan for the future, as Seth implies with the nod to homeostasis, as well as supporting introspection, theory of mind, empathy, and social cooperation (Graziano Reference Graziano2013).
Like prediction, these are all computations that can be carried out purely in the information domain, implying that there is no impediment to the realization of consciousness in silico; a perfect “simulation” of consciousness is consciousness, just as a perfect “simulation” of a calculator is a calculator. We would even speculate that some current AI systems, given their abilities to predict, model themselves and others, and self-correct, may already be partly conscious. We do not claim certainty either way. But we certainly do not share Seth’s confidence that “real” artificial consciousness is unlikely along current trajectories. On the contrary, we think it is likely.
Acknowledgments
We would like to thank our colleagues from the Paradigms of Intelligence team at Google for discussions that helped us to prepare this commentary.
Financial support
Blake Richards was supported by NSERC (Discovery Grant: RGPIN-2020-05105; Discovery Accelerator Supplement: RGPAS-2020-00031) and CIFAR (Canada AI Chair; Learning in Machine and Brains Fellowship).
Competing interests
Both authors of this commentary are employees at Google. There are no other competing interests to report.