In the Abstract to Seth’s article, he proposes to answer, “what it would take for conscious AI [artificial intelligence] to be a realistic prospect” by (a) challenging “the assumption that computation provides a sufficient basis for consciousness” and by (b) making a case “that consciousness depends on our nature as living organisms – a form of biological naturalism.”
Due to a word limit, we focus mainly on (b), contending that in making the case for biological naturalism, Seth likewise makes a case for the inevitability of AI consciousness.
To do this, we consider features that Seth views as essential for biological consciousness and show how these features are also realized in a simple machine model of consciousness, the Conscious Turing Machine (CTM). At the same time, we highlight how CTM aligns with underlying principles of theories that Seth views as promising for “real artificial consciousness.” These theories include global workspace theory (GWT); attention schema theory (AST); the embodied, embedded, enactive, and extended (4E) theories borrowed from cognitive science; and integrated information theory (IIT).
To support a case for biological naturalism, Seth enumerates features that he views as key to consciousness, features that presumably evolved to keep living beings alive. He comments often on the richness of biological systems and human experiences. No argument here. But to conclude, as he often does, that AI systems could not entertain these very features is a non-sequitur.
The focus of Seth’s argument for biological naturalism (section 4.0) relies primarily on two related notions: predictive processing and the free energy principle, both of which highlight the important cybernetic process of prediction error minimization. According to Seth, prediction error minimization is key to the entity making “best guesses” about sensory signals (both external and internal). These best guesses are what generate controlled hallucinations, which Seth describes as providing our conscious perception and sense of self. Additionally, by minimizing sensory error predictions, Seth asserts that the free energy principle enables the system to maintain boundaries between internal and external states, further supporting conscious experience via controlled hallucination.
To comment on the realization of prediction error minimization and controlled hallucinations in CTM, we start with a brief overview of CTM, highlighting features that align with underlying principles of major theories of consciousness. See Blum and Blum (Reference Blum and Blum2022) and Blum and Blum (Reference Blum and Blum2025).
CTM’s architecture is inspired by the GWT “theater model” of consciousness. For simplicity, its stage, STM, holds one (formally defined) chunk at a time. Its audience, LTM, consists of an enormous collection of initially independent powerful processors. Over time, links are created between like-minded processors. While processors each have their own internal language, they communicate with each other in Brainish, CTM’s self-generated multimodal language.
For simplicity, all CTM processors are in LTM, including its Input (sensory) and Output (motor) processors and Gauge processors (embodied homeostats, e.g., fuel and temperature gauges). Initially, all other processors are generic, their functionality emerging over time. The CTM is embedded and enacts in its outer world via its Input sensors and Output actuators. Its functionality is extended through interactions with its outer world and by augmenting LTM with “off-the-shelf” processors. As we shall see, global integration and extensive feedback in CTM align with IIT’s underlying principles for achieving high phi.
At each (discrete) moment in time, processors submit chunks into a well-defined fast competition to get on stage. The winning chunk is then globally broadcast to all LTM processors triggering CTM’s attention and, by CTM’s Axiom 1, a unitary experience.
For awareness and feelings of consciousness in CTM, the co-evolution of Brainish and CTM’s world models enter the picture. A sketch of a rose in a world model might be labeled with the multisensory Brainish gist, “bright red color, sweet smell, soft touch.” When CTM inspects a broadcasted chunk whose Brainish gist denotes this labeled sketch, CTM will become aware of the labeled sketch. A unitary and, by Axiom 2, a fused subjective experience will be evoked. CTM will experience the rose’s bright red color, sweet smell, soft touch of its petal.
This notion of awareness aligns with AST and with the role of Seth’s hallucinations.
CTM’s predictive dynamics (cycles of prediction, feedback and learning, which work to minimize prediction error, locally and globally) are analogous to both predictive processing and the free energy principle and play a major role in the co-evolvement of Brainish and CTM’s world models. We also note that CTM’s Input sensors and Output actuators serve as CTM’s Markov blanket.
Seth argues that while predictive processing may seem like a computational process, its continuity with the free energy principle makes it “a dynamical, non-computational theory.” We beg to differ.
This takes us to computation and Seth’s challenge (a).
We note that in CTM, time is discrete. Additionally, CTM’s processes employ digital computations. Utilizing predictive dynamics, these processes evolve over time, affected by and responding to circumstances. However, Seth seems to assume that a machine that uses digital computations must be rigid. Indeed, to reinforce his case for biological naturalism, Seth proclaims, “Instead of a computational rigidity, there is an excitable fluidity to the matter of life.”
On numerous occasions, Seth surmises that while digital computations might approximate neural processes, “it would seem an open question whether these approximations would be good enough to realize consciousness.” On the other hand, he notes that since “exact Bayesian inference is generally analytically intractable, the brain implements an approximation.”
To explore the feasibility of alternate computing possibilities, Seth speculates about non-computable functions and the potential of rectifying them with quantum, analogue, neuromorphic, and mortal computation. Of course, some are “straw men.” For example, according to the mathematical theory of quantum computation, the class of quantum and Turing computable functions is identical. As for speed-ups, although algebraic problems like factoring are quantum computable in polynomial time, there is evidence that this is not the case for NP-complete problems. See Bernstein and Vazirani (Reference Bernstein and Vazirani1997). We refer to Nayebi (Reference Nayebi2014), which likewise disputes the claim that analogue computing could be more powerful than digital computing.
Conscious AI is inevitable.
You have
Access
In the Abstract to Seth’s article, he proposes to answer, “what it would take for conscious AI [artificial intelligence] to be a realistic prospect” by (a) challenging “the assumption that computation provides a sufficient basis for consciousness” and by (b) making a case “that consciousness depends on our nature as living organisms – a form of biological naturalism.”
Due to a word limit, we focus mainly on (b), contending that in making the case for biological naturalism, Seth likewise makes a case for the inevitability of AI consciousness.
To do this, we consider features that Seth views as essential for biological consciousness and show how these features are also realized in a simple machine model of consciousness, the Conscious Turing Machine (CTM). At the same time, we highlight how CTM aligns with underlying principles of theories that Seth views as promising for “real artificial consciousness.” These theories include global workspace theory (GWT); attention schema theory (AST); the embodied, embedded, enactive, and extended (4E) theories borrowed from cognitive science; and integrated information theory (IIT).
To support a case for biological naturalism, Seth enumerates features that he views as key to consciousness, features that presumably evolved to keep living beings alive. He comments often on the richness of biological systems and human experiences. No argument here. But to conclude, as he often does, that AI systems could not entertain these very features is a non-sequitur.
The focus of Seth’s argument for biological naturalism (section 4.0) relies primarily on two related notions: predictive processing and the free energy principle, both of which highlight the important cybernetic process of prediction error minimization. According to Seth, prediction error minimization is key to the entity making “best guesses” about sensory signals (both external and internal). These best guesses are what generate controlled hallucinations, which Seth describes as providing our conscious perception and sense of self. Additionally, by minimizing sensory error predictions, Seth asserts that the free energy principle enables the system to maintain boundaries between internal and external states, further supporting conscious experience via controlled hallucination.
To comment on the realization of prediction error minimization and controlled hallucinations in CTM, we start with a brief overview of CTM, highlighting features that align with underlying principles of major theories of consciousness. See Blum and Blum (Reference Blum and Blum2022) and Blum and Blum (Reference Blum and Blum2025).
CTM’s architecture is inspired by the GWT “theater model” of consciousness. For simplicity, its stage, STM, holds one (formally defined) chunk at a time. Its audience, LTM, consists of an enormous collection of initially independent powerful processors. Over time, links are created between like-minded processors. While processors each have their own internal language, they communicate with each other in Brainish, CTM’s self-generated multimodal language.
For simplicity, all CTM processors are in LTM, including its Input (sensory) and Output (motor) processors and Gauge processors (embodied homeostats, e.g., fuel and temperature gauges). Initially, all other processors are generic, their functionality emerging over time. The CTM is embedded and enacts in its outer world via its Input sensors and Output actuators. Its functionality is extended through interactions with its outer world and by augmenting LTM with “off-the-shelf” processors. As we shall see, global integration and extensive feedback in CTM align with IIT’s underlying principles for achieving high phi.
At each (discrete) moment in time, processors submit chunks into a well-defined fast competition to get on stage. The winning chunk is then globally broadcast to all LTM processors triggering CTM’s attention and, by CTM’s Axiom 1, a unitary experience.
For awareness and feelings of consciousness in CTM, the co-evolution of Brainish and CTM’s world models enter the picture. A sketch of a rose in a world model might be labeled with the multisensory Brainish gist, “bright red color, sweet smell, soft touch.” When CTM inspects a broadcasted chunk whose Brainish gist denotes this labeled sketch, CTM will become aware of the labeled sketch. A unitary and, by Axiom 2, a fused subjective experience will be evoked. CTM will experience the rose’s bright red color, sweet smell, soft touch of its petal.
This notion of awareness aligns with AST and with the role of Seth’s hallucinations.
CTM’s predictive dynamics (cycles of prediction, feedback and learning, which work to minimize prediction error, locally and globally) are analogous to both predictive processing and the free energy principle and play a major role in the co-evolvement of Brainish and CTM’s world models. We also note that CTM’s Input sensors and Output actuators serve as CTM’s Markov blanket.
Seth argues that while predictive processing may seem like a computational process, its continuity with the free energy principle makes it “a dynamical, non-computational theory.” We beg to differ.
This takes us to computation and Seth’s challenge (a).
We note that in CTM, time is discrete. Additionally, CTM’s processes employ digital computations. Utilizing predictive dynamics, these processes evolve over time, affected by and responding to circumstances. However, Seth seems to assume that a machine that uses digital computations must be rigid. Indeed, to reinforce his case for biological naturalism, Seth proclaims, “Instead of a computational rigidity, there is an excitable fluidity to the matter of life.”
On numerous occasions, Seth surmises that while digital computations might approximate neural processes, “it would seem an open question whether these approximations would be good enough to realize consciousness.” On the other hand, he notes that since “exact Bayesian inference is generally analytically intractable, the brain implements an approximation.”
To explore the feasibility of alternate computing possibilities, Seth speculates about non-computable functions and the potential of rectifying them with quantum, analogue, neuromorphic, and mortal computation. Of course, some are “straw men.” For example, according to the mathematical theory of quantum computation, the class of quantum and Turing computable functions is identical. As for speed-ups, although algebraic problems like factoring are quantum computable in polynomial time, there is evidence that this is not the case for NP-complete problems. See Bernstein and Vazirani (Reference Bernstein and Vazirani1997). We refer to Nayebi (Reference Nayebi2014), which likewise disputes the claim that analogue computing could be more powerful than digital computing.
Conscious AI is inevitable.
Acknowledgments
We thank Johannes Kleiner, Aran Nayebi, and Umesh Vazirani for extremely helpful discussions. We take responsibility for our arguments and conclusions.
Financial support
We are grateful to UniDT for helping to support our research.
Competing interests
Neither author has competing interests.