When functionalism entered analytic philosophy of mind, it was seen as a way to avoid substance dualism without endorsing behaviorism – which attributed mentality too liberally – or identity theory – which chauvinistically denied it too narrowly. The notion of computation, formalized in the 1930s, gave early functionalist philosophers (e.g., Armstrong, Reference Armstrong1968; Fodor, Reference Fodor1968; Putnam, Reference Putnam and Block1967) a vocabulary to identify the mental with a property that could fit a monist worldview. Soon after, Block (Reference Block and Savage1978) argued that “machine state functionalists,” for whom mental states were to be identified with states of a Turing-machine table, risked over-attributing mentality to systems like the Bolivian economy or a population slice of China equipped with radios.
Block’s challenge can be read in two ways. On one hand, if neurons are the relevant computational units and their computations can be described in a Turing machine table, then one must attribute mentality to systems that intuitively lack it. On the other hand, if neurons are the relevant units, their computations likely cannot be captured by such a Turing-machine table implementable by people and radios. This latter view inspired Andy Clark’s (Reference Clark1989) “microfunctionalism,” a biologically plausible version of functionalism meant to address doubts about machine-table models of the brain.
Microfunctionalism holds that computations are in principle multiply realizable, but in practice very few things – neurons, glia, etc. – can instantiate those constituting our mental states. Computation remains substrate independent, yet computational complexity rules out most materials. This isn’t new: computer engineers abandoned vacuum tubes for transistors, then integrated circuits, then silicon microprocessors, as earlier materials proved too slow, fragile, or costly to run ever more complex computations. But from the fact that certain computations cannot be carried out by certain materials, it does not follow that computational processes aren’t substrate independent. As Clark noted, we must attend to the modal scope of the multiple realizability claim.
The same ambiguity in modal scope is present in the characterization of “substrate independence” Seth uses to argue against the possibility of consciousness in AI systems. According to Seth, substrate independence is “the claim that the same kind of mental states (including conscious states) can occur in systems with different physical substrates – i.e., systems made out of different kinds of stuff” (my italics). But to say that a mental state can occur in a different physical substrate does not mean that any physical substrate would do. Again, the more complex the computation, the fewer actual physical substrates can likely instantiate it. Thus, the fact that current conventional digital AI systems implemented in existing computational architectures do not instantiate conscious states does not speak against the in-principle claim that, given the right substrate, there could one day be a physical AI system that could successfully implement them.
We do not know how consciousness arises in biology, but if it depends on complex autopoietic processes absent from current digital AI systems, then I agree that current networks cannot be conscious. From the start, neural networks failed to capture brain processes likely crucial to mental life. Most are based on oversimplified neuronal models – known to be incomplete or false, at least since Bullock (Reference Bullock1959) challenged the neuron doctrine behind McCulloch and Pitts’s (Reference McCulloch and Pitts1943) architecture. Today, many of these very same processes do not feature in any way in the computational structure of our best AI systems. Yet, that does not mean that they cannot be characterized computationally, nor that they are not computational. And if one day we can characterize them computationally – or, at least, replicate them with artificial materials even if we don’t know how to characterize them computationally – then the possibility of AI systems that can instantiate consciousness may not sound so far-fetched.
The brain is profoundly complex, and Seth is exactly right in claiming that its generative entrenchment “imposes many constraints on possible alternative implementations, which likely exclude alternate substitutes such as silicon.” But this claim is entirely compatible with a microfunctionalist view, according to which as the complexity of the computation increases, the number of possible substrata that can instantiate it shrinks. Indeed, it may be possible to accept Seth’s claim that, at the end, it may be really hard – if not impossible – to separate “what brains do from what they are,” and yet still accept that what they do is, in fact, compute. For here the difficulty, even the impossibility, of separating them has simply to do with the extraordinarily reduced number of things in the universe that can possibly instantiate said computations. As Dennett (Reference Dennett2005, 21) once put it, as we learn more about brains, “minds will turn out not to be simple computers, and their computational resources will be seen to reach down into the subcellular molecular resources available only to organic brains, but the theories that emerge will still be functionalist in the broad sense” (p. 25).
Whatever computations our brains may be doing, they are the result of years and years of cellular evolution that has benefited from an ever-growing sequence of improbable tricks, all of which require physical substrata with very precise timings, durability, trophic cycles, energy consumption, etc. Yet, nothing in this improbability implies that what brains do is not computing – broadly understood, of course, so as to include non-digital computations. Separating what brains do from what they are may be hard, if not impossible, given the contingent universe we live in, but if what ultimately explains why brains can instantiate conscious states is something to do with what they are and not with what they do – as Seth’s biological naturalism implies – we risk falling into a kind of vitalism (not far from Bergson’s!) whereby all we could say is that brains are conscious because of some sort of élan consciente. And as long as we have reason to believe that there is a way to characterize what brains do in terms of computations, we won’t need to reject the possibility that, maybe one day, complex AI systems will be able to instantiate conscious states.
You have
Access
When functionalism entered analytic philosophy of mind, it was seen as a way to avoid substance dualism without endorsing behaviorism – which attributed mentality too liberally – or identity theory – which chauvinistically denied it too narrowly. The notion of computation, formalized in the 1930s, gave early functionalist philosophers (e.g., Armstrong, Reference Armstrong1968; Fodor, Reference Fodor1968; Putnam, Reference Putnam and Block1967) a vocabulary to identify the mental with a property that could fit a monist worldview. Soon after, Block (Reference Block and Savage1978) argued that “machine state functionalists,” for whom mental states were to be identified with states of a Turing-machine table, risked over-attributing mentality to systems like the Bolivian economy or a population slice of China equipped with radios.
Block’s challenge can be read in two ways. On one hand, if neurons are the relevant computational units and their computations can be described in a Turing machine table, then one must attribute mentality to systems that intuitively lack it. On the other hand, if neurons are the relevant units, their computations likely cannot be captured by such a Turing-machine table implementable by people and radios. This latter view inspired Andy Clark’s (Reference Clark1989) “microfunctionalism,” a biologically plausible version of functionalism meant to address doubts about machine-table models of the brain.
Microfunctionalism holds that computations are in principle multiply realizable, but in practice very few things – neurons, glia, etc. – can instantiate those constituting our mental states. Computation remains substrate independent, yet computational complexity rules out most materials. This isn’t new: computer engineers abandoned vacuum tubes for transistors, then integrated circuits, then silicon microprocessors, as earlier materials proved too slow, fragile, or costly to run ever more complex computations. But from the fact that certain computations cannot be carried out by certain materials, it does not follow that computational processes aren’t substrate independent. As Clark noted, we must attend to the modal scope of the multiple realizability claim.
The same ambiguity in modal scope is present in the characterization of “substrate independence” Seth uses to argue against the possibility of consciousness in AI systems. According to Seth, substrate independence is “the claim that the same kind of mental states (including conscious states) can occur in systems with different physical substrates – i.e., systems made out of different kinds of stuff” (my italics). But to say that a mental state can occur in a different physical substrate does not mean that any physical substrate would do. Again, the more complex the computation, the fewer actual physical substrates can likely instantiate it. Thus, the fact that current conventional digital AI systems implemented in existing computational architectures do not instantiate conscious states does not speak against the in-principle claim that, given the right substrate, there could one day be a physical AI system that could successfully implement them.
We do not know how consciousness arises in biology, but if it depends on complex autopoietic processes absent from current digital AI systems, then I agree that current networks cannot be conscious. From the start, neural networks failed to capture brain processes likely crucial to mental life. Most are based on oversimplified neuronal models – known to be incomplete or false, at least since Bullock (Reference Bullock1959) challenged the neuron doctrine behind McCulloch and Pitts’s (Reference McCulloch and Pitts1943) architecture. Today, many of these very same processes do not feature in any way in the computational structure of our best AI systems. Yet, that does not mean that they cannot be characterized computationally, nor that they are not computational. And if one day we can characterize them computationally – or, at least, replicate them with artificial materials even if we don’t know how to characterize them computationally – then the possibility of AI systems that can instantiate consciousness may not sound so far-fetched.
The brain is profoundly complex, and Seth is exactly right in claiming that its generative entrenchment “imposes many constraints on possible alternative implementations, which likely exclude alternate substitutes such as silicon.” But this claim is entirely compatible with a microfunctionalist view, according to which as the complexity of the computation increases, the number of possible substrata that can instantiate it shrinks. Indeed, it may be possible to accept Seth’s claim that, at the end, it may be really hard – if not impossible – to separate “what brains do from what they are,” and yet still accept that what they do is, in fact, compute. For here the difficulty, even the impossibility, of separating them has simply to do with the extraordinarily reduced number of things in the universe that can possibly instantiate said computations. As Dennett (Reference Dennett2005, 21) once put it, as we learn more about brains, “minds will turn out not to be simple computers, and their computational resources will be seen to reach down into the subcellular molecular resources available only to organic brains, but the theories that emerge will still be functionalist in the broad sense” (p. 25).
Whatever computations our brains may be doing, they are the result of years and years of cellular evolution that has benefited from an ever-growing sequence of improbable tricks, all of which require physical substrata with very precise timings, durability, trophic cycles, energy consumption, etc. Yet, nothing in this improbability implies that what brains do is not computing – broadly understood, of course, so as to include non-digital computations. Separating what brains do from what they are may be hard, if not impossible, given the contingent universe we live in, but if what ultimately explains why brains can instantiate conscious states is something to do with what they are and not with what they do – as Seth’s biological naturalism implies – we risk falling into a kind of vitalism (not far from Bergson’s!) whereby all we could say is that brains are conscious because of some sort of élan consciente. And as long as we have reason to believe that there is a way to characterize what brains do in terms of computations, we won’t need to reject the possibility that, maybe one day, complex AI systems will be able to instantiate conscious states.
Acknowledgments
N/A.
Financial support
N/A.
Competing interests
No competing interests to declare.