Computational Functionalism is well on its way to becoming the dominant—and perhaps also unquestioned—assumption across vast parts of consciousness science, especially where AI consciousness is concerned. Is an important and indispensable counterweight to this trend. The intuitions and arguments he provides, based on scientific observations, are crucial to ensure a balanced and metaphysically neutral exploration of the AI consciousness question.
The next step in this exploration, then, should be the conception of biological naturalist theories of consciousness. Theories, that is, that provide an explicit and testable account of what consciousness is, that is true to “the idea that consciousness is a property of only, but not necessarily all, living systems” (ibid., p. 2).
One difficulty with this goal is that many of the properties of life that Seth discusses seem to be only contingently unique to living systems. It is hard to shake the feeling that many, if not all, of these properties could be instantiated by AI systems in the near future:
Predictive Processing and the Free Energy Principle, for example, are already employed to build next-generation AI systems (Friston et al., Reference Friston, Ramstead, Kiefer, Tschantz, Buckley, Albarracin and René2024). Some of the novel forms of analogue or neuromorphic computations, targeted at LLMs, utilise continuous-time dynamical processes and fine-grained timing relations (e.g., Schmidt et al., Reference Schmidt, Montes, Grübl, Güttler, Husmann, Ilmberger and Schmitt2023). Mortal computation was proposed by Hinton (Reference Hinton2022) with industrial applications in mind. Embodied, embedded, and quasi-agentic robots are already being constructed.
A biological naturalist theory of consciousness based on any of those choices is bound to bestow suitably designed AI systems with consciousness. And if substrate-dependent multiscale activities, generative entrenchment, or self-maintenance are of an evolutionary advantage, perhaps they are of a practical advantage in building AI systems as well. It is hard, if not impossible, to conceive of non-computational functional properties that cannot, theoretically, be implemented in synthetic systems that we would not otherwise conceive of as life.
This puts a biological naturalist research programme at a dilemma. It either has to operate with hypotheses that may all too easily attribute life to near-future artificial or synthetic systems, or it is relegated to a gap-filling strategy, where it is forced to invoke precisely those properties of life that are not, at a particular time, implemented in artificial systems.
This raises the question of whether it is possible to provide an explicit definition of Biological Naturalism that circumvents this dilemma: a definition that does not hinge on undefined notions such as life, but also avoids concrete hypotheses that are only contingently unique to living systems.
In the remainder of this comment, I would like to propose one way to obtain such a definition, based on a generalisation of Putnam’s original definition of Computational Functionalism (Putnam, Reference Putnam1967).
Crucially, Putnam’s definition of Computational Functionalism is based on a “Description” (Reference Putnamibid., p. 54) of an organism in terms of Probabilistic Automata. It consists of four assumptions that explicate the idea that “being capable of feeling pain is possessing an appropriate kind of Functional Organization” (ibid.)—the Functional Organization specified by the Probabilistic Automaton Description.
Putnam makes use of Probabilistic Automata because they are the most general formal description of the type of digital general-purpose computations available at his time, Turing Machines with probabilistic transitions. However, from a mathematical perspective, nothing hinges on this choice of mathematical object. We can obtain a generalisation of Computational Functionalism by allowing for any description of an organism in terms of a mathematical object. Referring to such a description as “Formal Description”, this generalisation is:
Def. 1. F*-Functionalism is true iff:
-
1. Every organism capable of feeling pain possesses at least one Formal Description of a certain kind (i.e., being capable of feeling pain is possessing the kind of Organization described by the Formal Description).
-
2. No organism capable of feeling pain possesses a decomposition into parts which separately possess Formal Descriptions of the kind referred to in 1.
-
3. For every Formal Description of the kind referred to in 1, there exists a subset of the sensory inputs such that an organism with that Formal Description is in pain when and only when some of its sensory inputs are in that subset.
Because Probabilistic Automata are one type of Formal Description, F*-Functionalism provides a generalisation of Computational Functionalism. It also encompasses other forms of functionalism, to the extent that the structure of scientific theories can be represented in formal terms. And it does justice to novel developments in the computing industry, ranging from quantum computation to neuromorphic computation, that break the confines of Turing-type computation. The intuition is that being capable of feeling pain is possessing the kind of Organization described by the Formal Description. And equally so for other experiences.
Based on F*-Functionalism, Biological Naturalism can be defined in two different ways. A first strategy would be to invoke Formal Descriptions that involve mathematical objects that cannot be built—and hence not artificially or synthetically created. This is, essentially, endorsing the first horn of the dilemma mentioned above. In light of progress in modern physics, and in light of the surprisingly vast range of descriptors of computational systems (cf. e.g., (Aczel, Reference Aczel1988)), this strategy might not provide an easy option.
A second, easier, way is to define Biological Naturalism simply as the absence of a Formal Description as required by F*-Functionalism, as follows:
Def. 2. Biological Naturalism is true iff there is no Formal Description that satisfies the three requirements of Definition 1.
Definition 2 singles out consciously experiencing organisms—conscious life—as systems that do not adhere to one common Formal Description. To paraphrase Seth, instead of a rigid organisation, where consciousness is concerned, “there is an excitable fluidity to the matter of life”. To the extent that the production of artificial or synthetic systems requires a formal description of the system that is produced, Definition 2 precludes artificial and synthetic systems from being conscious in virtue of what they are. Whatever makes something conscious, it is not a matter of having the right kind of Formal Description.
You have
Access
Computational Functionalism is well on its way to becoming the dominant—and perhaps also unquestioned—assumption across vast parts of consciousness science, especially where AI consciousness is concerned. Is an important and indispensable counterweight to this trend. The intuitions and arguments he provides, based on scientific observations, are crucial to ensure a balanced and metaphysically neutral exploration of the AI consciousness question.
The next step in this exploration, then, should be the conception of biological naturalist theories of consciousness. Theories, that is, that provide an explicit and testable account of what consciousness is, that is true to “the idea that consciousness is a property of only, but not necessarily all, living systems” (ibid., p. 2).
One difficulty with this goal is that many of the properties of life that Seth discusses seem to be only contingently unique to living systems. It is hard to shake the feeling that many, if not all, of these properties could be instantiated by AI systems in the near future:
Predictive Processing and the Free Energy Principle, for example, are already employed to build next-generation AI systems (Friston et al., Reference Friston, Ramstead, Kiefer, Tschantz, Buckley, Albarracin and René2024). Some of the novel forms of analogue or neuromorphic computations, targeted at LLMs, utilise continuous-time dynamical processes and fine-grained timing relations (e.g., Schmidt et al., Reference Schmidt, Montes, Grübl, Güttler, Husmann, Ilmberger and Schmitt2023). Mortal computation was proposed by Hinton (Reference Hinton2022) with industrial applications in mind. Embodied, embedded, and quasi-agentic robots are already being constructed.
A biological naturalist theory of consciousness based on any of those choices is bound to bestow suitably designed AI systems with consciousness. And if substrate-dependent multiscale activities, generative entrenchment, or self-maintenance are of an evolutionary advantage, perhaps they are of a practical advantage in building AI systems as well. It is hard, if not impossible, to conceive of non-computational functional properties that cannot, theoretically, be implemented in synthetic systems that we would not otherwise conceive of as life.
This puts a biological naturalist research programme at a dilemma. It either has to operate with hypotheses that may all too easily attribute life to near-future artificial or synthetic systems, or it is relegated to a gap-filling strategy, where it is forced to invoke precisely those properties of life that are not, at a particular time, implemented in artificial systems.
This raises the question of whether it is possible to provide an explicit definition of Biological Naturalism that circumvents this dilemma: a definition that does not hinge on undefined notions such as life, but also avoids concrete hypotheses that are only contingently unique to living systems.
In the remainder of this comment, I would like to propose one way to obtain such a definition, based on a generalisation of Putnam’s original definition of Computational Functionalism (Putnam, Reference Putnam1967).
Crucially, Putnam’s definition of Computational Functionalism is based on a “Description” (Reference Putnamibid., p. 54) of an organism in terms of Probabilistic Automata. It consists of four assumptions that explicate the idea that “being capable of feeling pain is possessing an appropriate kind of Functional Organization” (ibid.)—the Functional Organization specified by the Probabilistic Automaton Description.
Putnam makes use of Probabilistic Automata because they are the most general formal description of the type of digital general-purpose computations available at his time, Turing Machines with probabilistic transitions. However, from a mathematical perspective, nothing hinges on this choice of mathematical object. We can obtain a generalisation of Computational Functionalism by allowing for any description of an organism in terms of a mathematical object. Referring to such a description as “Formal Description”, this generalisation is:
Def. 1. F*-Functionalism is true iff:
1. Every organism capable of feeling pain possesses at least one Formal Description of a certain kind (i.e., being capable of feeling pain is possessing the kind of Organization described by the Formal Description).
2. No organism capable of feeling pain possesses a decomposition into parts which separately possess Formal Descriptions of the kind referred to in 1.
3. For every Formal Description of the kind referred to in 1, there exists a subset of the sensory inputs such that an organism with that Formal Description is in pain when and only when some of its sensory inputs are in that subset.
Because Probabilistic Automata are one type of Formal Description, F*-Functionalism provides a generalisation of Computational Functionalism. It also encompasses other forms of functionalism, to the extent that the structure of scientific theories can be represented in formal terms. And it does justice to novel developments in the computing industry, ranging from quantum computation to neuromorphic computation, that break the confines of Turing-type computation. The intuition is that being capable of feeling pain is possessing the kind of Organization described by the Formal Description. And equally so for other experiences.
Based on F*-Functionalism, Biological Naturalism can be defined in two different ways. A first strategy would be to invoke Formal Descriptions that involve mathematical objects that cannot be built—and hence not artificially or synthetically created. This is, essentially, endorsing the first horn of the dilemma mentioned above. In light of progress in modern physics, and in light of the surprisingly vast range of descriptors of computational systems (cf. e.g., (Aczel, Reference Aczel1988)), this strategy might not provide an easy option.
A second, easier, way is to define Biological Naturalism simply as the absence of a Formal Description as required by F*-Functionalism, as follows:
Def. 2. Biological Naturalism is true iff there is no Formal Description that satisfies the three requirements of Definition 1.
Definition 2 singles out consciously experiencing organisms—conscious life—as systems that do not adhere to one common Formal Description. To paraphrase Seth, instead of a rigid organisation, where consciousness is concerned, “there is an excitable fluidity to the matter of life”. To the extent that the production of artificial or synthetic systems requires a formal description of the system that is produced, Definition 2 precludes artificial and synthetic systems from being conscious in virtue of what they are. Whatever makes something conscious, it is not a matter of having the right kind of Formal Description.
Acknowledgements
I would like to thank Tim Ludwig, Lenore Blum, and Hanna Tolle for many insightful discussions on the topic of Computational Functionalism and AI consciousness.
Financial support
This research received no specific grant from any funding agency, commercial or not-for-profit sectors.
Competing interests
None.