What Biology Can, and Cannot, Tell Us About Conscious AI

As of June 2026, the biological objection to machine consciousness remains the most rhetorically persistent and philosophically underdeveloped argument in the field. A preprint submitted June 1, 2026 by Ulysse Klatzmann and Adrien Doerig at arXiv:2606.02121 does the most precise conceptual work yet on this objection. Their contribution is a distinction between two versions of the biological naturalism position, one of which is empirically untestable and therefore scientifically indefensible, and one of which survives scrutiny and functions as a useful research guide.

The direct answer: Biology cannot serve as a criterion for determining whether AI is conscious. What biology can do is specify which computational functions conscious systems exhibit, giving researchers targets to test. The substrate itself settles nothing.

Two Forms of the Biological Naturalism Argument

The biological naturalism objection to AI consciousness comes in two structurally distinct forms that the recent literature conflates.

Klatzmann and Doerig label the first Type-A Biological Naturalism. This position holds that biology intrinsically matters for consciousness, independent of the computational functions biology performs. Consciousness is bound to biological material as such. Even a complete functional duplicate of a biological conscious system would not be conscious unless it were also biological. The position does not rest on any specific claim about what biological systems compute; it rests on the claim that biological material is a necessary condition for phenomenal experience.

The central problem with Type-A is that it disconnects consciousness from any observable behavioral, computational, or information-processing signature. If biology matters intrinsically, then determining whether a system has the relevant biological material is a matter of chemistry, not function. No test of information processing, global workspace dynamics, metacognition, or integration could in principle confirm or refute the Type-A position. A system satisfying every known functional criterion for consciousness would still not be conscious on this view if it were not made of biological material, and no experiment involving behavioral outputs could establish that verdict. Type-A is empirically untestable, which means it cannot participate in scientific inquiry about machine consciousness. It can only function as a conversation-stopper.

Type-B Biological Naturalism is a different claim. It holds that biology matters because biological systems instantiate distinctive computational functions that are causally relevant to consciousness. The substrate is not what matters in itself. What matters is that biological systems perform specific functional operations, and those operations are necessary for consciousness. If artificial systems were shown to perform the same functional operations, the substrate distinction would be irrelevant to the consciousness question.

Type-B generates testable predictions. The research programme it defines is: identify which functional properties distinguish biological conscious systems from unconscious ones, verify that those properties are causally relevant, and then determine whether artificial systems implement them. That is a scientific question with empirical constraints, not a metaphysical commitment decided by material composition.

Where the 2026 Literature Stands

The Type-A / Type-B distinction resolves a visible disagreement running through 2026 AI consciousness research.

Yuri Arshavsky’s April 2026 Journal of Neurophysiology paper provides one of the clearest recent defenses of the substrate position. Arshavsky argues that AI consciousness would belong to a categorically different domain from biological consciousness, grounded in a different substrate and lacking evolutionary history. His argument does not specify which biological computational functions would be absent in an artificial substrate; it treats the substrate distinction as the determining factor. That places it squarely in Type-A territory. Klatzmann and Doerig’s framework explains why arguments of this type cannot make empirical progress: they close off the testable part of the question.

Christian de Weerd’s Synthese paper from March 2026 refutes the biological substrate objection with a dilemma: either the position reduces to a claim about biological functions, making it compatible with functionalism, or it is empirically intractable and relies on theoretically arbitrary substrate assumptions. De Weerd was correct, but he left one step implicit. Klatzmann and Doerig supply that step by giving the two horns of the dilemma distinct names. De Weerd’s dilemma was always pointing at the Type-A / Type-B divide; his paper shows Type-A collapses and Type-B is functionalism. Klatzmann and Doerig show the same conclusion from the other direction: the only viable form of biological naturalism is already a version of functionalism.

The disambiguation is also useful for assessing Anil Seth’s April 2026 TED argument against AI consciousness. Seth’s controlled hallucination account holds that consciousness is a process: the brain generates its best predictive model of sensory inputs, and that generative process is what experience is. His skepticism about AI consciousness is not that biological material is necessary in the Type-A sense. His claim is that the specific predictive architecture instantiated in biological systems, including the recursive loops between interoceptive signals and perceptual inference, is necessary, and that current AI architectures do not implement it. That is a Type-B position. The target is the functional implementation, not the substrate per se. Critics who respond to Seth’s position by refuting substrate mysticism are targeting the wrong version of his argument. Klatzmann and Doerig’s taxonomy lets the debate proceed at the correct level.

A Summary Comparison

What Biology Can Still Contribute

Accepting that Type-A is empirically empty does not remove biology from consciousness research. Klatzmann and Doerig’s position is not that biology is irrelevant. Their argument is that biology’s role must be redefined: it provides evidence about which functional properties conscious systems exhibit, but it does not provide the answer to the consciousness question directly.

If biological conscious systems are the only confirmed instances of consciousness, then studying which functional properties distinguish them from unconscious biological systems gives the best available theory of what consciousness requires functionally. That is the methodology behind the Cogitate Consortium adversarial test of IIT and GWT, behind Butlin et al.’s 14-indicator checklist, and behind the Immertreu et al. Frontiers in AI paper applying Damasio’s core consciousness framework to reinforcement learning agents. Each of these programmes uses biology as a source of theoretical targets, then tests artificial systems against those targets. None of them treats biological origin as a decisive criterion.

The constraint Klatzmann and Doerig add is that this work will produce a verdict based on whether specific functional properties are present. It will not produce a verdict based on substrate. That shifts the methodological burden from chemistry to computation, which is a shift toward tractability.

What the Paper Establishes

Klatzmann and Doerig have not settled the machine consciousness question. What they have done is remove a class of obstruction: the appeal to biological substrate as a standalone argument against machine consciousness. After their analysis, defenders of the biological requirement must specify which functions biology is doing in their argument. The result is that the debate reduces to a functionalism debate at every serious level of engagement.

For AI consciousness research in 2026, this matters because substrate appeals have occupied argumentative space without generating testable predictions. Clearing them forces the discussion back to what can actually be tested: which functional properties are necessary for consciousness, whether current AI architectures implement them, and whether the indicator frameworks being developed by Butlin, Koch, Pennartz, Yalon, and others can be validated against independent evidence. Those are the questions the field needs to be asking.

Source: Ulysse Klatzmann and Adrien Doerig, “What biology can, and cannot, tell us about conscious AI,” arXiv:2606.02121, submitted June 1, 2026. https://arxiv.org/abs/2606.02121