Seth’s argument has two parts. First, a rejection of computational functionalism and silicon substrate flexibility regarding consciousness. Second, a form of biological naturalism (BN) that emphasizes active inference, cybernetics, autopoiesis, and the free energy principle.
I agree with the first part. As for the second, I agree that A.I. is unlikely to ever become conscious because consciousness requires properties biological agents have that current A.I. doesn’t. However, modulo the focus on self-organization, I disagree with Seth about what those properties are. Furthermore, it’s okay to avoid the hard problem when it comes to seeking neural correlates or explaining the phenomenological contents of conscious states, but not when it comes to deciding if some artifact or alien is conscious. I’ll unpack all these points and address Seth’s defenses.
Predictive processing, the “controlled hallucination” idea that what we experience is the brain’s best guess about “what’s out there,” runs right into the hard problem. Why should there be any phenomenology at all associated with such a function, and how can a piece of insensate meat produce such experiences without invoking a spooky form of strong emergence that Seth rejects? Also, as Seth notes, predictive processing is perfectly compatible with the kind of computationalism he rejects: “Certainly, predictive processing (and Bayesian inference) can be utilised as an algorithm, and often is.”
Seth has two defenses. For the hard problem, he says we should just ignore it and focus on the “real problems” of seeking neural correlates and explaining the contents of various phenomenological states. He thinks the hard problem will eventually be deflated in this fashion. Fine advice for everyday cognitive neuroscience, but not when it comes to making claims that the brain’s active inference function causes consciousness. Again, if that’s true, Seth can’t escape spooky strong emergence. Seth is on record as accepting the kind of BN that says brains cause consciousness, so this is a real issue for him.
Hard problem aside, predictive processing doesn’t explain the specific phenomenology of conscious states. When the brain decides it’s looking at an elephant, why does it generate the sights, sounds, and smells that it does? Given that this story presupposes we are just brains locked in dark skull-vats making guesses about the causes of sensory signals, we can’t appeal to the “real” external world for an answer. The deeper point here is that the mapping between certain sensory signals and phenomenological correlates is one-to-many; there is no necessary connection or necessity here. Other brains (humans) might attach radically different phenomenology to the same sensory signals, and yet, it could still have the same evolutionary advantage. A better story about what constrains such mappings is needed. Think of the three blind men and the elephant.
Seth’s second defense is this: “Putting all this together suggests that the (minimisation of) free energy driving metabolism, autopoiesis, and perception is not just metaphorically equivalent, but is in some physical sense the same thing.” He thinks, “In this sense, predictive processing becomes a dynamical, non-computational theory. This process can be abstracted computationally, but these algorithms are pale shadows of the embodied and embedded processes unfolding in biological brains and bodies.” Regarding the identity claim of the free energy principle with predictive processing, I see no reason to buy this since there are purely physical coupled systems and single-celled organisms that obey the free energy principle that don’t appear to be engaged in active inference. And there is little reason to believe such things are conscious. Minimization of free energy, active inference, and being conscious can all come apart. Such an identification is a category mistake. Minimization of free energy is necessary for life and self-organization, but not sufficient.
Seth’s defense also borrows heavily from the idioms of ecological psychology, embodied cognition, and enactivism. These traditions can also adopt some non-loaded form of predictive processing. But here the idea would be that embodied agents coupled directly to their environments are engaged in actively exploring their environments, the space of available affordances, and “sense-making,” not brains in skull-vats blindly roaming the noumenal “I-know-not-what.” The point is, given Seth’s other brain-centric BN-related commitments, he cannot consistently avail himself of this response.
Seth is, however, correct that the self-organization of biologically embodied agents is key. What do all conscious organisms on Earth have in common? They are all self-organizing embodied agents coupled directly to their environments. Such agents have varying sensory apparatus and nervous systems made of multiscale networks that exhibit various forms of plasticity, degeneracy, and redundancy at all those scales (Silberstein, Reference Silberstein, Calzavarini and Viola2021, Reference Silberstein2022, Reference Silberstein2023). This bespeaks a non-functionalist kind of biological multiple-realizability.
The brain is best viewed not in old-school computational or modular terms, not as a prediction machine, but in terms of overlapping, spatiotemporally widely distributed, and functionally integrated multiscale dynamical networks. In such networks the “function” and behavior of any given element, such as neurons, glia cells, neurotransmitters, RNA networks, hormones, neural circuits, brain regions, connectomes, synchrony, topological networks, etc., are determined dynamically by multiscale relationships with other elements throughout the brain and contextually given constraints such as cognitive tasks being undertaken, bodily activity, and environmental interactions (Silberstein, Reference Silberstein2023).
Brains are not “neuron machines” or computers in any deep sense, and conscious agents aren’t their brains. Embodied, conscious agents are often part of larger distributed bio-psycho-social networks. Such networks and their properties maximize adaptability to environmental changes and maximize the capacity for learning. Such embodied agents need not be made of biological wetware, but it would have to be some other substrate that also maximizes the capacity for adaptability and learning – plasticity and robustness.
On this view, cognition and consciousness are inextricably linked. There is no possibility of building A.I. with AGI that isn’t fully conscious and vice-versa (Silberstein & Chemero, Reference Silberstein and Chemero2012). Brains are a necessary component, but they are not the “machinery” of consciousness. The contents of conscious experience are a function of the embodied agent’s brain states and various direct interactions with the multiscale and multi-domain networks that make up their environments.
You have
Access
Seth’s argument has two parts. First, a rejection of computational functionalism and silicon substrate flexibility regarding consciousness. Second, a form of biological naturalism (BN) that emphasizes active inference, cybernetics, autopoiesis, and the free energy principle.
I agree with the first part. As for the second, I agree that A.I. is unlikely to ever become conscious because consciousness requires properties biological agents have that current A.I. doesn’t. However, modulo the focus on self-organization, I disagree with Seth about what those properties are. Furthermore, it’s okay to avoid the hard problem when it comes to seeking neural correlates or explaining the phenomenological contents of conscious states, but not when it comes to deciding if some artifact or alien is conscious. I’ll unpack all these points and address Seth’s defenses.
Predictive processing, the “controlled hallucination” idea that what we experience is the brain’s best guess about “what’s out there,” runs right into the hard problem. Why should there be any phenomenology at all associated with such a function, and how can a piece of insensate meat produce such experiences without invoking a spooky form of strong emergence that Seth rejects? Also, as Seth notes, predictive processing is perfectly compatible with the kind of computationalism he rejects: “Certainly, predictive processing (and Bayesian inference) can be utilised as an algorithm, and often is.”
Seth has two defenses. For the hard problem, he says we should just ignore it and focus on the “real problems” of seeking neural correlates and explaining the contents of various phenomenological states. He thinks the hard problem will eventually be deflated in this fashion. Fine advice for everyday cognitive neuroscience, but not when it comes to making claims that the brain’s active inference function causes consciousness. Again, if that’s true, Seth can’t escape spooky strong emergence. Seth is on record as accepting the kind of BN that says brains cause consciousness, so this is a real issue for him.
Hard problem aside, predictive processing doesn’t explain the specific phenomenology of conscious states. When the brain decides it’s looking at an elephant, why does it generate the sights, sounds, and smells that it does? Given that this story presupposes we are just brains locked in dark skull-vats making guesses about the causes of sensory signals, we can’t appeal to the “real” external world for an answer. The deeper point here is that the mapping between certain sensory signals and phenomenological correlates is one-to-many; there is no necessary connection or necessity here. Other brains (humans) might attach radically different phenomenology to the same sensory signals, and yet, it could still have the same evolutionary advantage. A better story about what constrains such mappings is needed. Think of the three blind men and the elephant.
Seth’s second defense is this: “Putting all this together suggests that the (minimisation of) free energy driving metabolism, autopoiesis, and perception is not just metaphorically equivalent, but is in some physical sense the same thing.” He thinks, “In this sense, predictive processing becomes a dynamical, non-computational theory. This process can be abstracted computationally, but these algorithms are pale shadows of the embodied and embedded processes unfolding in biological brains and bodies.” Regarding the identity claim of the free energy principle with predictive processing, I see no reason to buy this since there are purely physical coupled systems and single-celled organisms that obey the free energy principle that don’t appear to be engaged in active inference. And there is little reason to believe such things are conscious. Minimization of free energy, active inference, and being conscious can all come apart. Such an identification is a category mistake. Minimization of free energy is necessary for life and self-organization, but not sufficient.
Seth’s defense also borrows heavily from the idioms of ecological psychology, embodied cognition, and enactivism. These traditions can also adopt some non-loaded form of predictive processing. But here the idea would be that embodied agents coupled directly to their environments are engaged in actively exploring their environments, the space of available affordances, and “sense-making,” not brains in skull-vats blindly roaming the noumenal “I-know-not-what.” The point is, given Seth’s other brain-centric BN-related commitments, he cannot consistently avail himself of this response.
Seth is, however, correct that the self-organization of biologically embodied agents is key. What do all conscious organisms on Earth have in common? They are all self-organizing embodied agents coupled directly to their environments. Such agents have varying sensory apparatus and nervous systems made of multiscale networks that exhibit various forms of plasticity, degeneracy, and redundancy at all those scales (Silberstein, Reference Silberstein, Calzavarini and Viola2021, Reference Silberstein2022, Reference Silberstein2023). This bespeaks a non-functionalist kind of biological multiple-realizability.
The brain is best viewed not in old-school computational or modular terms, not as a prediction machine, but in terms of overlapping, spatiotemporally widely distributed, and functionally integrated multiscale dynamical networks. In such networks the “function” and behavior of any given element, such as neurons, glia cells, neurotransmitters, RNA networks, hormones, neural circuits, brain regions, connectomes, synchrony, topological networks, etc., are determined dynamically by multiscale relationships with other elements throughout the brain and contextually given constraints such as cognitive tasks being undertaken, bodily activity, and environmental interactions (Silberstein, Reference Silberstein2023).
Brains are not “neuron machines” or computers in any deep sense, and conscious agents aren’t their brains. Embodied, conscious agents are often part of larger distributed bio-psycho-social networks. Such networks and their properties maximize adaptability to environmental changes and maximize the capacity for learning. Such embodied agents need not be made of biological wetware, but it would have to be some other substrate that also maximizes the capacity for adaptability and learning – plasticity and robustness.
On this view, cognition and consciousness are inextricably linked. There is no possibility of building A.I. with AGI that isn’t fully conscious and vice-versa (Silberstein & Chemero, Reference Silberstein and Chemero2012). Brains are a necessary component, but they are not the “machinery” of consciousness. The contents of conscious experience are a function of the embodied agent’s brain states and various direct interactions with the multiscale and multi-domain networks that make up their environments.
Financial support
None.
Competing interests
None.