There is much that Seth and I agree on (as might be expected, since we have been colleagues at Sussex since the time of his PhD work). For example, I too have advocated for a predictive (“expectation-based”) processing approach to understanding consciousness (Chrisley & Parthemore, Reference Chrisley and Parthemore2007a; Chrisley & Parthemore, Reference Chrisley and Parthemore2007b). However, I suspect that (even his own, much more fully worked out version of) this approach does not have the implications he thinks it does. More confidently, I think his arguments in this area fail to establish their conclusions, even if some of those conclusions may turn out to be correct. So I am focusing here only on these argumentative differences.
First, Seth’s negative thesis, that conscious AI is impossible. He argues against conscious AI in two steps: The first step is to argue that computational functionalism, defined as the claim that “computation is sufficient for consciousness” (p. 1), is false. His second step is to claim (five times, and without argument) that if computational functionalism is false, conscious AI is impossible (target article, sect. 3.2, para. 4; sect. 3.6, para. 2; sect. 3.9, para. 1; sect. 5.0, para. 1; sect 5.3, para. 3, footnote 38).
But (as argued at length in (Chrisley, Reference Chrisley2008)) each of those five claims is incorrect, and so Seth’s master argument fails. An analogy might help show why. Consider the thesis of “McLarenism,” the thesis that owning a McLaren car is sufficient for winning a Formula One Grand Prix. Clearly, McLarenism is false. But its falsity does not imply that it is impossible for a McLaren car owner to win a Grand Prix! It also does not imply that acquiring a McLaren can’t be a key step to take in winning a Grand Prix, nor does it imply that an explanation of why the winner of the 1968 Belgian Grand Prix did in fact win does not involve their ownership of a McLaren, etc.
Just as a McLaren winning the Grand Prix is possible despite the falsity of McLarenism, so also is conscious AI (involving computation construed as narrowly as you like: digital, Turing-like, etc.) possible despite the falsity of computational functionalism. Further, computation might be a necessary part of any scientific explanation of consciousness, even if computational functionalism (as defined by Seth) is false. Even further, the analogy shows that computation might be a key component in an artificial conscious system even if computation is not necessary for consciousness.
These are not just philosophical niceties – they are crucial for understanding the true potential for conscious AI. A popular approach to artificial consciousness takes the view that computations must be grounded in (supplemented with) real-world, robotic (non-biological) sensorimotor capacities to yield consciousness, thus rejecting computational functionalism (at least as defined by Seth). Right or wrong (and we don’t know which it is!), and despite that rejection, no one sees such an approach as anything other than the pursuit of conscious AI.
What about Seth’s positive thesis, that consciousness is a biological phenomenon?
Despite detailing the intimate way in which human consciousness is interleaved with, and constrained by, our biology, Seth fails to show that this is necessarily essential to all consciousness. Consider flight: For living fliers (e.g., birds), “the connections between properties of the biological substrate and [particular instances of flying] have rich potential explanatory value. In the limit, one can think of each and every [particular instance of bird flight] as integrating multimodal bodily and environmental signals in the service of coordinating allostatic regulation and survival-relevant action. [Birds are flying creatures] because they are living, flesh and blood ‘beast machines’.” And: “Flight in biological systems is a dynamical and (plausibly) substrate-dependent process.” Do these facts imply that flight in general is best thought of as a biological phenomenon? Do they imply that artificial flight is impossible? Do they imply that artificial flight can only be achieved by building systems that closely emulate this biological complexity? Obviously not. Multiply-realizable aerodynamic principles can be abstracted from the biological context in which they first evolved and were exploited, and can instead be used to achieve artificial flight in non-living substrates. Pointing out that bird flight can only be truly and fully understood in the context of an evolved, self-maintaining system that metabolizes does not change that fact. Bird flight is a biological phenomenon, but flight in general is not. In fact, we only achieved artificial flight when we stopped trying too closely to emulate biological flight (compare Armer, Reference Armer and Chrisley1962). If consciousness is necessarily biological, it’s going to take a lot more work to demonstrate that than merely showing that it is a biological phenomenon when present in biological systems.
A final point: We’ll do a better job at understanding what computers (yes, even of the conventional – digital, von Neumann, Turing machine – sort) can and cannot do once we have a proper understanding of what computation itself is. The very fact that we recognize other, non-conventional forms of computation (analog computation, neuromorphic computation, quantum computation, etc.) shows that computation in general should not be defined as “what Turing machines do”; the fact that we do so anyway shows that we are not in the possession of a unifying account (Smith, Reference Smith1985; Smith, Reference Smith1996; Smith Reference Smith and Scheutz2002). The Turing machine theory reveals some properties of some kinds of computational systems, but it doesn’t capture the phenomenon of computation as a whole (Chrisley, Reference Chrisley1998). So, ultimately, we aren’t (or shouldn’t) be interested in questions like “can Turing machines be conscious?” What we’re really interested in is whether the concepts and methods that allow us best to explain, design, predict, and control the real-world phenomenon of computation (which already includes a lot more than what impoverished Turing machine theory captures) can also help us understand (natural) and build (artificial) minds – even conscious ones (Chrisley, Reference Chrisley and Scheutz2000).
You have
Access
There is much that Seth and I agree on (as might be expected, since we have been colleagues at Sussex since the time of his PhD work). For example, I too have advocated for a predictive (“expectation-based”) processing approach to understanding consciousness (Chrisley & Parthemore, Reference Chrisley and Parthemore2007a; Chrisley & Parthemore, Reference Chrisley and Parthemore2007b). However, I suspect that (even his own, much more fully worked out version of) this approach does not have the implications he thinks it does. More confidently, I think his arguments in this area fail to establish their conclusions, even if some of those conclusions may turn out to be correct. So I am focusing here only on these argumentative differences.
First, Seth’s negative thesis, that conscious AI is impossible. He argues against conscious AI in two steps: The first step is to argue that computational functionalism, defined as the claim that “computation is sufficient for consciousness” (p. 1), is false. His second step is to claim (five times, and without argument) that if computational functionalism is false, conscious AI is impossible (target article, sect. 3.2, para. 4; sect. 3.6, para. 2; sect. 3.9, para. 1; sect. 5.0, para. 1; sect 5.3, para. 3, footnote 38).
But (as argued at length in (Chrisley, Reference Chrisley2008)) each of those five claims is incorrect, and so Seth’s master argument fails. An analogy might help show why. Consider the thesis of “McLarenism,” the thesis that owning a McLaren car is sufficient for winning a Formula One Grand Prix. Clearly, McLarenism is false. But its falsity does not imply that it is impossible for a McLaren car owner to win a Grand Prix! It also does not imply that acquiring a McLaren can’t be a key step to take in winning a Grand Prix, nor does it imply that an explanation of why the winner of the 1968 Belgian Grand Prix did in fact win does not involve their ownership of a McLaren, etc.
Just as a McLaren winning the Grand Prix is possible despite the falsity of McLarenism, so also is conscious AI (involving computation construed as narrowly as you like: digital, Turing-like, etc.) possible despite the falsity of computational functionalism. Further, computation might be a necessary part of any scientific explanation of consciousness, even if computational functionalism (as defined by Seth) is false. Even further, the analogy shows that computation might be a key component in an artificial conscious system even if computation is not necessary for consciousness.
These are not just philosophical niceties – they are crucial for understanding the true potential for conscious AI. A popular approach to artificial consciousness takes the view that computations must be grounded in (supplemented with) real-world, robotic (non-biological) sensorimotor capacities to yield consciousness, thus rejecting computational functionalism (at least as defined by Seth). Right or wrong (and we don’t know which it is!), and despite that rejection, no one sees such an approach as anything other than the pursuit of conscious AI.
What about Seth’s positive thesis, that consciousness is a biological phenomenon?
Despite detailing the intimate way in which human consciousness is interleaved with, and constrained by, our biology, Seth fails to show that this is necessarily essential to all consciousness. Consider flight: For living fliers (e.g., birds), “the connections between properties of the biological substrate and [particular instances of flying] have rich potential explanatory value. In the limit, one can think of each and every [particular instance of bird flight] as integrating multimodal bodily and environmental signals in the service of coordinating allostatic regulation and survival-relevant action. [Birds are flying creatures] because they are living, flesh and blood ‘beast machines’.” And: “Flight in biological systems is a dynamical and (plausibly) substrate-dependent process.” Do these facts imply that flight in general is best thought of as a biological phenomenon? Do they imply that artificial flight is impossible? Do they imply that artificial flight can only be achieved by building systems that closely emulate this biological complexity? Obviously not. Multiply-realizable aerodynamic principles can be abstracted from the biological context in which they first evolved and were exploited, and can instead be used to achieve artificial flight in non-living substrates. Pointing out that bird flight can only be truly and fully understood in the context of an evolved, self-maintaining system that metabolizes does not change that fact. Bird flight is a biological phenomenon, but flight in general is not. In fact, we only achieved artificial flight when we stopped trying too closely to emulate biological flight (compare Armer, Reference Armer and Chrisley1962). If consciousness is necessarily biological, it’s going to take a lot more work to demonstrate that than merely showing that it is a biological phenomenon when present in biological systems.
A final point: We’ll do a better job at understanding what computers (yes, even of the conventional – digital, von Neumann, Turing machine – sort) can and cannot do once we have a proper understanding of what computation itself is. The very fact that we recognize other, non-conventional forms of computation (analog computation, neuromorphic computation, quantum computation, etc.) shows that computation in general should not be defined as “what Turing machines do”; the fact that we do so anyway shows that we are not in the possession of a unifying account (Smith, Reference Smith1985; Smith, Reference Smith1996; Smith Reference Smith and Scheutz2002). The Turing machine theory reveals some properties of some kinds of computational systems, but it doesn’t capture the phenomenon of computation as a whole (Chrisley, Reference Chrisley1998). So, ultimately, we aren’t (or shouldn’t) be interested in questions like “can Turing machines be conscious?” What we’re really interested in is whether the concepts and methods that allow us best to explain, design, predict, and control the real-world phenomenon of computation (which already includes a lot more than what impoverished Turing machine theory captures) can also help us understand (natural) and build (artificial) minds – even conscious ones (Chrisley, Reference Chrisley and Scheutz2000).
Acknowledgments
I thank Oliver Sharpe, Thomas Clark, Nigel Shardlow, and Uziel Awret for their helpful comments on a draft.
Financial support
This research received no specific grant from any funding agency, commercial, or not-for-profit sectors.
Competing interests
None.