In his excellent target article, Seth reviews a wide range of evidence with respect to the possibility of phenomenal consciousness in artificial intelligence. The two main rivals in his discussion are computational functionalism – the view that implementing the right kinds of computations is sufficient for consciousness – and biological naturalism, according to which only living systems can be conscious. He concludes that there are “good reasons” to hold that biological naturalism is true and that (silicon-based) AI cannot be conscious. I will put forward three considerations that jointly favor the conclusion that AI consciousness is possible, despite Seth’s arguments.
First, AI systems (made from silicon) may be conscious even if computational functionalism is false. Seth seems to go back and forth on this, saying: “computational functionalism is sufficient for substrate flexibility. But it is not necessary.” But also (incorrectly): “Conscious AI requires […] computational functionalism to hold.” While it is important to keep in mind, Seth could reasonably reply that most of his arguments target the claim that silicon-based systems can be conscious directly, rather than engaging with computational functionalism specifically. However, it is also important to note that biological naturalism is a stronger claim than that AI consciousness is impossible: It also rules out consciousness in creatures, for example, hypothetical extraterrestrials, which are not made from silicon but which also do not qualify as living creatures. Therefore, even if we reject computational functionalism, we should be careful not to jump to the conclusion that biological naturalism is correct.
Second, Seth reviews and develops multiple lines of argument against the possibility of conscious AI. He points to processes in human brains which cannot be replicated in AI systems, such as mortal computation (Hinton, Reference Hinton2022), neural computation (Piccinini, Reference Piccinini2021), biological computation (Kirkpatrick, Reference Kirkpatrick2022), analog computation, and neuromorphic computing. Moreover, AI systems have informational closure, a distinction between software and hardware, and they can be characterized computationally independent of the actual dynamics of the implementing system, and so on – all properties potentially absent from living brains. In addition, Seth endorses a predictive processing (PP) account of conscious perception and characterizes PP mechanisms as driven by fundamentally biological imperatives related to allostasis and ultimately grounded in the free energy principle.
I do believe that these arguments establish biological naturalism as a coherent alternative to computational functionalism. However, in my view, they fail to provide a reason why we should believe biological naturalism over computational functionalism.
Seth’s different arguments seem to share the same form:
Biological naturalism argument
P1: There is some feature F (implementing mortal computations, neural computations, etc.) that only living systems can implement.
P2: F is necessary for consciousness.
C: Therefore, only living systems can be conscious.
Seth’s arguments support P1. However, it is not clear why one should believe P2. From my reading, Seth does not provide reasons for P2. Thus, computational functionalists can counter with the following argument:
AI consciousness argument
P1: There is some feature G (implementing certain standard digital computations, that is, non-mortal computations, non-neural computations, etc.) that AI systems can implement.
P2: G is sufficient for consciousness.
C: Therefore, AI systems can be conscious.
Until someone gives us a reason to accept P2 of one of these arguments, we are at a stalemate between both views. I also do not see that Seth’s recourse to PP provides such an argument. Setting aside that PP may not be best thought of as a theory of consciousness (Schlicht & Dolega, Reference Schlicht and Dolega2021), it can be construed algorithmically – so that AI systems can implement PP – or biologically, and the empirical evidence for PP does not distinguish between these options. If we do not know whether biological naturalism is true, we cannot know whether implementing the high-level algorithms of PP is sufficient for consciousness or whether biology is necessary. Thus, appeals to PP cannot help us assess whether biological naturalism is true. So far, biological naturalism and the view that AI can be conscious are on equal footing.
However, third, it seems to me that there are multiple reasons in favor of P2 of the AI consciousness argument that at least have some evidential weight: 1. Attributing consciousness to AI systems which are, on some abstract level, computationally similar to conscious humans and animals leads to more unifying explanations of behavior across creatures (see Dung Reference Dungn.d.). 2. Individuating the mechanism underlying consciousness computationally does justice to its evolutionary function (see Goldstein and Kirk-Giannini Reference Goldstein and Kirk-Gianninin.d.). 3. It is plausible that most ordinary people would, with respect to some possible AI systems, have the strong sense that they are conscious, while consciousness science often appeals to folk intuitions (see Shevlin Reference Shevlinn.d.). 4. In addition, if one believes that the question of AI consciousness is sensitive to how we conceptually engineer (Cappelen, Reference Cappelen2018) consciousness, then it seems most advisable to engineer consciousness to refer to high-level functional properties shared with AI rather than biological ones, since the former are the ones that matter most in ethics, in explanation of behavior, and in everyday life (for relevant illusionist perspectives, see Dung, Reference Dung2022, sect. 6; Frankish, Reference Frankish2024; see also Cappelen & Dever, Reference Cappelen and Dever2021). 5. Finally, biological naturalism specifically is subject to the charge that restricting consciousness to biological systems requires positing that it is an implausible coincidence that humans are conscious (see Schwitzgebel & Pober, Reference Schwitzgebel and Pober2024).
None of these arguments are, of course, conclusive, and space does not permit it to develop them in detail (see also the argument against AI consciousness in Schwitzgebel & Pober, Reference Schwitzgebel and Pober2024; as well as relevant discussions in Dung & Kersten, Reference Dung and Kersten2024; Shiller, Reference Shiller2024; Mogensen, Reference Mogensen2024). However, I wish to emphasize a more general point here: Various promising routes to support P2 of the AI consciousness argument are available which, combined, have substantial evidential weight in favor of the view that AI can be conscious. However, I am not aware of any strong arguments for P2 of the biological naturalism argument. Accordingly, until such arguments are provided, we should conclude that it is more likely that AI consciousness is possible than that it is not possible.
You have
Access
In his excellent target article, Seth reviews a wide range of evidence with respect to the possibility of phenomenal consciousness in artificial intelligence. The two main rivals in his discussion are computational functionalism – the view that implementing the right kinds of computations is sufficient for consciousness – and biological naturalism, according to which only living systems can be conscious. He concludes that there are “good reasons” to hold that biological naturalism is true and that (silicon-based) AI cannot be conscious. I will put forward three considerations that jointly favor the conclusion that AI consciousness is possible, despite Seth’s arguments.
First, AI systems (made from silicon) may be conscious even if computational functionalism is false. Seth seems to go back and forth on this, saying: “computational functionalism is sufficient for substrate flexibility. But it is not necessary.” But also (incorrectly): “Conscious AI requires […] computational functionalism to hold.” While it is important to keep in mind, Seth could reasonably reply that most of his arguments target the claim that silicon-based systems can be conscious directly, rather than engaging with computational functionalism specifically. However, it is also important to note that biological naturalism is a stronger claim than that AI consciousness is impossible: It also rules out consciousness in creatures, for example, hypothetical extraterrestrials, which are not made from silicon but which also do not qualify as living creatures. Therefore, even if we reject computational functionalism, we should be careful not to jump to the conclusion that biological naturalism is correct.
Second, Seth reviews and develops multiple lines of argument against the possibility of conscious AI. He points to processes in human brains which cannot be replicated in AI systems, such as mortal computation (Hinton, Reference Hinton2022), neural computation (Piccinini, Reference Piccinini2021), biological computation (Kirkpatrick, Reference Kirkpatrick2022), analog computation, and neuromorphic computing. Moreover, AI systems have informational closure, a distinction between software and hardware, and they can be characterized computationally independent of the actual dynamics of the implementing system, and so on – all properties potentially absent from living brains. In addition, Seth endorses a predictive processing (PP) account of conscious perception and characterizes PP mechanisms as driven by fundamentally biological imperatives related to allostasis and ultimately grounded in the free energy principle.
I do believe that these arguments establish biological naturalism as a coherent alternative to computational functionalism. However, in my view, they fail to provide a reason why we should believe biological naturalism over computational functionalism.
Seth’s different arguments seem to share the same form:
Biological naturalism argument
P1: There is some feature F (implementing mortal computations, neural computations, etc.) that only living systems can implement.
P2: F is necessary for consciousness.
C: Therefore, only living systems can be conscious.
Seth’s arguments support P1. However, it is not clear why one should believe P2. From my reading, Seth does not provide reasons for P2. Thus, computational functionalists can counter with the following argument:
AI consciousness argument
P1: There is some feature G (implementing certain standard digital computations, that is, non-mortal computations, non-neural computations, etc.) that AI systems can implement.
P2: G is sufficient for consciousness.
C: Therefore, AI systems can be conscious.
Until someone gives us a reason to accept P2 of one of these arguments, we are at a stalemate between both views. I also do not see that Seth’s recourse to PP provides such an argument. Setting aside that PP may not be best thought of as a theory of consciousness (Schlicht & Dolega, Reference Schlicht and Dolega2021), it can be construed algorithmically – so that AI systems can implement PP – or biologically, and the empirical evidence for PP does not distinguish between these options. If we do not know whether biological naturalism is true, we cannot know whether implementing the high-level algorithms of PP is sufficient for consciousness or whether biology is necessary. Thus, appeals to PP cannot help us assess whether biological naturalism is true. So far, biological naturalism and the view that AI can be conscious are on equal footing.
However, third, it seems to me that there are multiple reasons in favor of P2 of the AI consciousness argument that at least have some evidential weight: 1. Attributing consciousness to AI systems which are, on some abstract level, computationally similar to conscious humans and animals leads to more unifying explanations of behavior across creatures (see Dung Reference Dungn.d.). 2. Individuating the mechanism underlying consciousness computationally does justice to its evolutionary function (see Goldstein and Kirk-Giannini Reference Goldstein and Kirk-Gianninin.d.). 3. It is plausible that most ordinary people would, with respect to some possible AI systems, have the strong sense that they are conscious, while consciousness science often appeals to folk intuitions (see Shevlin Reference Shevlinn.d.). 4. In addition, if one believes that the question of AI consciousness is sensitive to how we conceptually engineer (Cappelen, Reference Cappelen2018) consciousness, then it seems most advisable to engineer consciousness to refer to high-level functional properties shared with AI rather than biological ones, since the former are the ones that matter most in ethics, in explanation of behavior, and in everyday life (for relevant illusionist perspectives, see Dung, Reference Dung2022, sect. 6; Frankish, Reference Frankish2024; see also Cappelen & Dever, Reference Cappelen and Dever2021). 5. Finally, biological naturalism specifically is subject to the charge that restricting consciousness to biological systems requires positing that it is an implausible coincidence that humans are conscious (see Schwitzgebel & Pober, Reference Schwitzgebel and Pober2024).
None of these arguments are, of course, conclusive, and space does not permit it to develop them in detail (see also the argument against AI consciousness in Schwitzgebel & Pober, Reference Schwitzgebel and Pober2024; as well as relevant discussions in Dung & Kersten, Reference Dung and Kersten2024; Shiller, Reference Shiller2024; Mogensen, Reference Mogensen2024). However, I wish to emphasize a more general point here: Various promising routes to support P2 of the AI consciousness argument are available which, combined, have substantial evidential weight in favor of the view that AI can be conscious. However, I am not aware of any strong arguments for P2 of the biological naturalism argument. Accordingly, until such arguments are provided, we should conclude that it is more likely that AI consciousness is possible than that it is not possible.
Acknowledgements
I thank Christian de Weerd and Kayla Carnation for very valuable comments.
Financial support
No funding to declare.
Competing interests
No competing interests to declare.