We are not epistemically justified to judge whether Biological Naturalism or Computational Functionalism is true.

Seth correctly emphasizes the potential relevance of biochemical processes for consciousness: “some properties of life may be necessary for consciousness” (p. 13), and conscious artificial systems may be impossible. But it is far from established that consciousness is anchored in life and that “the (higher-level) mechanisms of predictive processing are […] inextricably embedded in their (lower-level) biological context” (p. 8). Seth’s interpretation of predictive processing seems idiosyncratic. Other proponents may disagree, regarding it as a general computational model of cognition and consciousness, without metaphysical commitments like making its explanatory merit dependent on the prediction error mechanism being tied to the self-sustaining mechanisms in our bodies (Schlicht, Reference Schlicht2025).

Regarding the possibility of conscious AI, Seth engages with Butlin et al., (Reference Butlin, Long, Elmoznino, Bengio, Birch, Constant, Deane, Fleming, Frith, Ji, Kanai, Klein, Lindsay, Michel, Mudrik, Peters, Schwitzgebel, Simon and VanRullen2023), who advocate a research program that assumes Computational Functionalism and proposes to use neuroscientific theories to yield computational markers of consciousness and thereby assess (i.e., verify) consciousness in AI. Like other mental phenomena, consciousness is assumed to be functional and computational. Functionalism allows for multiple realizability but Computationalism requires medium independence (Piccinini, Reference Piccinini2009). For their combination, Computational Functionalism, to be vindicated, neural computation must be independent from, i.e., unconstrained by, properties of its implementation in the brain. But while digital computation may be medium independent by definition (Haugeland, Reference Haugeland1985), it is an open question (indeed, questionable) whether neural computation is medium independent. Maley (Reference Maley2023, Reference Maley2025) argues that it is analog and therefore medium dependent. Piccinini (Reference Piccinini2020) takes it to be different from digital and analog computation, yet medium independent (see also Drayson, Reference Drayson2025). But since Piccinini acknowledges that we do not know how this computation sui generis works, calling it medium independent is premature.

For Butlin et al.’s program to succeed, the theories they consider, Global Neuronal Workspace Theory (GNW), for example, are presented as computational theories, allowing for conscious AI in principle (Dehaene et al., Reference Dehaene, Lau and Kouider2017, Mudrik et al., Reference Mudrik, Boly, Dehaene, Fleming, Lamme, Seth and Melloni2025). GNW proposes four neuronal signatures of consciousness (Dehaene, Reference Dehaene2014), supported by evidence from human brains. But for it to provide computational markers, these would have to be medium independent and applicable not only to human brains but also to non-biological conscious systems. How could computational markers be empirically vindicated for non-biological systems? Phenomenal consciousness must be characterized functionally (contra Chalmers, Reference Chalmers1996), since discovering a computational mechanism requires it to have effects. But even characterized functionally, e.g., via identification with conscious access (Dehaene, Reference Dehaene2014, Cohen & Dennett, Reference Cohen and Dennett2011), a non-biological implementation of a workspace may not yield conscious but only “Ersatz broadcasting” (Block, Reference Block2008, p. 119). In non-biological artificial systems, all markers we use in human and non-human animals are either trivially absent (e.g., neuronal and first-person markers), ambiguous (e.g., reports, other behaviors), or question-begging (computational markers). Consider that the sophisticated outputs of large language models are not typically taken as evidence for consciousness. Contra Butlin et al.’s suggestion, because of redundancy and incompatibility, computational markers do not pile up and make conscious AI more likely.

Neuroscientific theories provide empirically supported neurobiological markers but not equally supported computational markers. They are supported as biological theories. Cao (Reference Cao2022) and Chirimuuta (Reference Chirimuuta2022, Reference Chirimuuta2025) highlight the difficulty of separating functional from structural properties in chemical signaling between neurons, listing numerous features of neural information processing that may make conscious AI impossible in practice. Possibility in principle is unverifiable if conscious AI requires data centers the size of Asia. A silicon-based system functionally equivalent to a brain may be technologically impossible, apart from the fact that we do not know which level of granularity is required for functional and computational equivalence. Given that the evidence favors biological accounts, Cao (Reference Cao2022, p. 25) is right that the burden of proof is on the functionalist to demonstrate that we can fully replicate the function of neurons. Birch (Reference Birch2024, p. 68) concurs that if equivalence of computations within neurons was required, this may be impossible to create.

I agree with Seth that we should thoroughly scrutinize the typically unquestioned assumption of Computational Functionalism. But Biological Naturalism only has a slight advantage. While I have stressed that neuroscientific theories of consciousness are empirically supported via evidence from human brains, capturing consciousness in other biological systems may require a more abstract functionalist account. Being unsure about whether the neurobiological and chemical details of biological systems are relevant for consciousness also means that they may not be. Functionalists could point to evidence for consciousness in other animals to support multiple realizability. However, this would not vindicate Computational Functionalism, and in the case of other humans or nonhuman animals, we remain in the realm of “messy” (Cao, Reference Cao2022, p. 26) biological implementation. An evolutionary account explaining how these biological mechanisms have evolved, say, in mammalian and avian brains (Godfrey-Smith, Reference Godfrey-Smith2024), is unavailable for artificial systems. But the simplified picture rooted in McCullough and Pitts’ (Reference McCulloch and Pitts1943) ideas may capture the essentials of neural computation, making the implementation properties of human brains dispensable. Analogously, the computational elements of the predictive processing framework, including the hierarchical generative model and prediction error minimization machinery, may be detachable from the life-sustaining biochemical processes of living organisms and still yield consciousness if implemented non-biologically. Combined with multiple realizability, this would support Computational Functionalism. However, there is no evidence for either Seth’s view or the opposition. We simply do not know, and it is difficult to see how we could decide the issue.

Verifying if an embodied AI can do a backflip is straightforward but verifying if it is conscious is not. Unless there are not only tests for consciousness (Bayne et al., Reference Bayne, Seth, Massimini, Shepherd, Cleeremans, Fleming, Malach, Mattingley, Menon, Owen, Peters, Razi and Mudrik2024) but also tests that inform us about medium (in-)dependence of neural computation, we remain epistemically limited about whether Biological Naturalism or Computational Functionalism is more plausible. Since neurobiological markers will always be better supported than computational markers, Seth’s biological story seems to be in the lead. But the moral of our epistemic situation is that we cannot assess consciousness in artificial systems. We can only attribute it – from the “intentional stance” (Dennett, Reference Dennett1987) – without verifying it, a much less ambitious claim.