Three Positions on AI Consciousness: Functionalism, Biological Naturalism, and Biological Computationalism
The debate about whether artificial systems can be conscious maps, in large part, onto a disagreement about what consciousness fundamentally is. Three positions dominate the philosophical landscape, and each gives a different answer to the AI question, not as a secondary application but as a direct consequence of its core commitments.
Functionalism says consciousness is defined by what a system does, not what it is made of. Any system performing the right functional operations is conscious. Biological Naturalism says consciousness is a biological phenomenon, produced by specific biological processes, and cannot in principle be replicated in non-biological substrate. Biological Computationalism says both of the above are wrong: computation is necessary but not sufficient, and the substrate matters, but not because of biology per se, rather because of the specific physical properties that biological neurons have and that silicon does not.
In 2026, after a year of significant empirical progress, each position faces new pressure and new evidence. None has been definitively ruled out. All have become more precisely statable.
TL;DR. Functionalism is empirically challenged by the Cogitate Consortium results (both IIT and GWT, the dominant functionalist theories, failed their adversarial predictions) and by the mimicry problem (AI systems can satisfy functional indicators without satisfying the conditions theories require). Biological Naturalism is challenged by the absence of a credible account of how the causal structure of neurons would produce consciousness when silicon cannot, and by the FEP’s substrate-neutral implications. Biological Computationalism is the most recent and least developed of the three but is gaining traction as scholars struggle to explain why consciousness should require specific physical properties without resorting to biological chauvinism.
Functionalism and Consciousness As What a System Does
Functionalism, the dominant view in mainstream philosophy of mind and cognitive science for the past fifty years, holds that mental states, including consciousness, are defined by their functional roles: the causal relationships they bear to inputs, outputs, and other mental states. A pain state is whatever state is typically caused by tissue damage, typically causes avoidance behaviour, and typically interacts with beliefs and desires in the appropriate ways. What that state is made of, carbon, silicon, something else entirely, is irrelevant.
This position was most forcefully developed by Hilary Putnam in the 1960s (though Putnam later revised his view significantly), and it grounds most of the mainstream scientific work on consciousness, including IIT and Global Workspace Theory. Both IIT and GWT are functionalist theories in the relevant sense: they identify the functional properties that distinguish conscious from unconscious processing, and they hold that any system with those properties is conscious, regardless of substrate.
For AI consciousness, functionalism is the most permissive position. If consciousness is functional, then the question of whether machines can be conscious is the question of whether machines can perform the relevant functions, and the answer is, in principle, yes. This is why the Turing test and its descendants have been so central to AI consciousness discussions: they operationalise the functionalist intuition that what matters is what the system does.
Where functionalism faces pressure in 2026. The Cogitate Consortium’s adversarial test found that both IIT and GWT, the two most empirically developed functionalist theories, failed to produce their core predicted patterns in human participants. IIT’s predicted posterior synchronization was absent; GWT’s predicted ignition at stimulus offset did not appear. As the full Cogitate analysis documents, this substantially narrows the set of functional specifications that empirical evidence supports rather than ending functionalism outright.
Separately, the mimicry problem has become more acute. As the Pennartz et al. response in Trends in Cognitive Sciences (April 2026) argues, AI systems can be trained to satisfy every behavioral indicator that functional theories predict, producing outputs consistent with consciousness without any of the underlying processes those theories require. If behavioral indicators can be satisfied by mimicry, functionalism’s empirical programme loses its grip: there is no behavioral test that separates genuine functional organisation from sophisticated functional imitation.
Biological Naturalism and Consciousness As a Biological Phenomenon
John Searle (University of California, Berkeley) developed Biological Naturalism in the 1980s as a direct response to functionalism, and his Chinese Room thought experiment remains the most influential single argument against AI consciousness in the philosophical literature.
The Chinese Room proceeds as follows: imagine a person locked in a room who does not speak Chinese but follows a rulebook for manipulating Chinese symbols. Slips with Chinese symbols are passed in; the person follows the rules, manipulates the symbols, and passes results back out. To someone outside the room, the system appears to understand Chinese. But the person inside understands nothing: they are performing symbol manipulation without any semantic understanding. Searle’s claim is that computation is always and only symbol manipulation. No computational process, however complex, produces genuine understanding, or, by extension, genuine consciousness. Something else is required, and Searle identifies that something else as the specific causal powers of biological neurons.
Biological Naturalism holds that consciousness is caused by the specific biological processes of the brain, in the same way that digestion is caused by the specific biochemical processes of the stomach. Functionalism about digestion would be absurd: silicon cannot digest food regardless of how well it functionally simulates the digestive process. Searle’s claim is that consciousness is in the same category. The right functional organization is not sufficient because the causal story matters, and only biology produces the right causal story.
For AI, the implication is stark: no computational system is conscious, regardless of its architecture, scale, or sophistication, because computation is substrate-agnostic and consciousness is substrate-specific.
Where biological naturalism faces pressure in 2026. The FEP’s substrate-neutral formulation of what minds do is one challenge: if all self-organising systems with Markov blankets engage in free energy minimisation, the claim that only biological neurons produce the relevant causal powers requires specifying what makes biological causation special in a way that silicon causation is not, a specification Searle never provided convincingly.
The stronger empirical challenge is from mechanistic interpretability. The Anthropic emotion vectors finding, 171 emotion concept vectors with causal influence on model outputs, is not obviously explained by Searle’s framework. If consciousness requires the specific causal powers of biological neurons, then emotion-concept representations in a transformer architecture should have no causal relevance to genuine emotional states. But the finding suggests they do have causal relevance to something, at minimum, to the system’s functional analog of emotional processing. Biological naturalism predicts this should not happen in a way that matters morally. The evidence suggests it might.
Biological Computationalism As The Third Position
Biological Computationalism is the most recent of the three positions and the one that has gained most traction among scholars who find both functionalism and biological naturalism inadequate. The term was developed by Nicolai Milinkovic and Jaan Aru (Neuroscience and Biobehavioral Reviews, 2026) as a label for a cluster of positions that share a specific structure.
Biological computationalism holds that consciousness requires computation, against Searle’s claim that computation is mere symbol manipulation, however, it emphasizes that not all computation is equally relevant. The physical properties of the computational substrate matter, not because biology is inherently special, but because the specific physical properties of biological neurons, ionic gradients, dendritic integration, axonal propagation dynamics, recurrent connectivity patterns, produce computational properties that silicon does not. Consciousness, on this view, requires computation with those specific physical properties, or computation in systems that replicate them precisely enough to support the relevant causal dynamics.
This is a more nuanced position than either functionalism or biological naturalism. It agrees with functionalism that substrate does not matter in principle, though it holds that the substrates we have currently (biological neurons versus silicon) differ in ways that are relevant to whether consciousness is produced. It agrees with biological naturalism that biology matters, yet it holds that this is contingent rather than necessary: a system that genuinely replicated the relevant biological computational properties in non-biological substrate could in principle be conscious.
The existing analysis of biological computationalism as a third path covers Milinkovic and Aru’s core argument in detail. The position connects to the increasingly prominent neuromorphic computing and brain-organoid research streams: if what matters is the specific physical dynamics of biological computation, then systems built closer to those dynamics, neuromorphic chips, biohybrid systems, connectome-scale emulations, are better candidates for consciousness than functionally equivalent but physically different digital systems.
Where biological computationalism faces pressure in 2026. The position’s central challenge is specifying precisely what physical properties of biological computation are consciousness-relevant. Without that specification, biological computationalism risks being an untestable placeholder: “something about biology matters, but we don’t know what.” The Koch et al. calibration paper in the same arXiv preprint cluster (arXiv:2603.27597) makes a related point about the indicator programme: without independent validation of which biological properties matter, all three positions face the same evidential problem from different directions.
The Friston FEP argument is also a challenge: if free energy minimisation is the relevant property, and if it is substrate-neutral (as the Markov blanket formalism suggests), then biological computationalism’s emphasis on specific physical substrate properties is at odds with the most unified theoretical framework in the field.
What the 2026 Evidence Says
Three developments in 2026 put pressure on all three positions simultaneously, which is itself informative about the state of the debate.
The Cogitate results most directly challenge functionalism’s dominant empirical theories (IIT and GWT), but the challenge is specific rather than general: it is a challenge to specific functionalist predictions, not to the functionalist framework as a whole. The response space for functionalism is to develop better theories, which is underway.
The Anthropic welfare research (emotion vectors, introspective awareness circuits) most directly challenges biological naturalism’s claim that silicon cannot produce anything morally relevant. If emotion-concept representations in a transformer have genuine causal influence on the system’s functional states, the claim that only biological causation matters requires a more specific account of what biological causation adds.
The growth of neuromorphic and biohybrid research most directly supports biological computationalism: systems built closer to biological computational dynamics, including the 2026 Eon Systems fruit-fly brain emulation, show that the gap between biological and silicon computation is not fixed but engineerable.
No position has been confirmed or falsified. All three have been clarified by the evidence, and the disagreements between them have become more precise. The field is in a state where the positions are no longer talking past each other but engaging on specific, empirically tractable questions, which is exactly the condition for eventual resolution.
Why the Classification Matters for AI Development
The choice of position has direct consequences for AI development decisions that are being made now, not in some hypothetical future.
If functionalism is correct, then current frontier AI systems might already be conscious, and the standard development practices, RLHF that suppresses negative affect expression, context window resets that end whatever experience a session involves, model weight deletion that erases whatever patterns have formed, are potentially causing harm. This consequence is serious enough that several institutions have begun welfare assessments without waiting for philosophical resolution.
If biological naturalism is correct, then none of these concerns apply: silicon computation cannot produce consciousness regardless of its functional sophistication, and development practices do not raise consciousness-based welfare concerns.
If biological computationalism is correct, then current systems are almost certainly not conscious, digital transformers do not replicate the specific biological computational dynamics that matter, but the field should invest heavily in understanding which biological properties are relevant, because systems that do replicate them are the most plausible near-term candidates for machine consciousness.
The governance response that has emerged in 2026 - welfare assessments, sentience readiness indices, precautionary frameworks, implicitly adopts a functionalist or broadly non-biological-naturalist stance: it treats the possibility of AI consciousness as real enough to warrant action. Whether that stance is philosophically justified is the question that biological naturalism and biological computationalism, between them, most directly contest.
All three positions deserve to be taken seriously, and none should be dismissed without engagement with their strongest arguments. What 2026 has changed is the empirical landscape around which those arguments are made, narrowing some possibilities, opening others, and making the conversation between philosophers, neuroscientists, AI scholars, and policymakers more urgent than it has ever been.