Seth’s biological naturalism restricts consciousness to living systems, a claim with profound implications for emerging technologies. We propose experiments that transform this metaphysical debate into testable predictions, offering both theoretical progress and evidence-based governance for potentially conscious technologies.

The implications extend beyond artificial intelligence (AI). When researchers replicate pain pathways in human assembloids (Kim et al., Reference Kim, Imaizumi, Jurjuț, Kelley, Wang, Thete, Hudacova, Amin, Levy, Scherrer and Pașca2025), we must trust that those assembloids lack subjective experience. Due to our limited understanding of consciousness, technologies whose conscious status divides expert opinion could emerge in the near term. We outline a research agenda aiming to reduce expert disagreement on the nature of consciousness by converting metaphysical debates into testable predictions.

Ongoing disagreements over metaphysical assumptions and scientific theories of consciousness suggest that assessments of nonhuman consciousness are best framed probabilistically. Long et al. (Long et al., Reference Long, Sebo, Butlin, Finlinson, Fish, Harding, Pfau, Sims, Birch and Chalmers2024) formalize an individual’s credence that a specific AI system is conscious, $P\left( C \right),$ as $P\left( C \right) = P\left( {CF} \right)\; \times \;P\left( {C|CF} \right)\;$ , with $P\left( {CF} \right)$ representing credence that computational functionalism is true, and $\left( {C|CF} \right)$ representing credence that the system is conscious if computational functionalism is true.

Previous work (Butlin et al., Reference Butlin, Long, Elmoznino, Bengio, Birch, Constant, Deane, Fleming, Frith, Ji, Kanai, Klein, Lindsay, Michel, Mudrik, Peters, Schwitzgebel, Simon and VanRullen2023) suggests several empirical indicators inspired by scientific theories of consciousness compatible with computational functionalism. The above equation can be expanded into the below:

$P\left( C \right) = P\left( {CF} \right)\; \times \;\mathop \sum \limits_{i = 1}^n P\left( {{T_i}|CF} \right) \times P\left( {C|{T_i},\;CF} \right)\;$

The indicators proposed by Butlin et al. impact $P\left( {C|{T_i},\;CF} \right)$ , meaning that detecting their presence in an AI system will actually increase collective uncertainty about the system’s consciousness. Evidence supporting computational indicators of consciousness will paradoxically increase disagreement unless we first address underlying metaphysical disputes about $P\left( {CF} \right).$

Using the recent adversarial collaboration testing the predictions of global neuronal workspace and integrated information theories of consciousness (Ferrante et al., Reference Ferrante, Gorska-Klimowska, Henin, Hirschhorn, Khalaf, Lepauvre, Liu, Richter, Vidal, Bonacchi, Brown, Sripad, Armendariz, Bendtz, Ghafari, Hetenyi, Jeschke, Kozma, Mazumder and Melloni2025) as a blueprint, we propose experiments that translate Seth’s arguments against computational functionalism into falsifiable predictions in order to reduce the individual-to-individual variance of $P\left( {CF} \right)$ .

We formalize credence in a specific theory of consciousness as $P(T_i) = P(Q) \cdot P({T_{i}} | Q) + P(\sim Q) \cdot P(T_{i} | \sim Q)$ , where $Q$ represents a metaphysical statement (e.g. mind is substrate independent) and $P\left( Q \right)$ represents one’s credence that the statement is true. By designing studies targeting the term $P(Q) \cdot P({T_{i}} | Q)$ , we aim to reduce individual-to-individual variance in $P\left( C \right)$ by condensing distributions of beliefs $P\left( {{T_i}} \right)$ and $P\left( Q \right)$ . The below experiments are designed to explicitly preclude a rationale for updating $P\left( {{T_i}|\sim Q} \right)$ . Therefore, when observers who inaccurately predict the experimental results adjust their confidence in $P({T_{i}} | Q)$ , maintaining coherent total probability requires these observers revise either their confidence in the experimentally relevant theory of consciousness, $P\left( {{T_i}} \right)\;$ , or in their metaphysical priors, $P\left( Q \right)$ and $P\left(\sim Q \right)$ .

1. 1. $Q:$ Consciousness’s function is allostasis, ${T_i}:$ Predictive processing theories: Researchers have successfully taught brain organoids—three-dimensional cell cultures—to perform reinforcement learning tasks such as Pong by applying unpredictable “punishment” stimuli for undesirable behavior (Jordan et al., Reference Jordan, Kutter, Comby, Brozzi and Kurtys2024; Kagan et al., Reference Kagan, Kitchen, Tran, Habibollahi, Khajehnejad, Parker, Bhat, Rollo, Razi and Friston2022). Seth suggests that the biological imperative for allostasis may have been “the evolutionary driver for predictive processing as a general mechanism underlying conscious perception.” To test this relationship, researchers could analyze organoid computing performance across a range of punishment-stimuli entropy values. If predictive processing serves allostasis, learning should improve with punishment-stimuli entropy until excessive unpredictability disrupts homeostasis, degrading performance. Conversely, if predictive processing operates independently of allostatic imperatives, entropy should minimally affect performance. Results favoring the former would require observers either to update toward biological naturalism or away from predictive processing as an explanation for conscious perception.

2. 2. $Q:$ Consciousness emerges from autopoiesis, ${T_i}:$ Recurrent processing theories: Researchers have produced human brain organoids that structurally mimic the reciprocal connections that form “reentry circuits” in human brains, and have demonstrated that connected organoids produce spontaneous oscillations which are more complex than single or fused organoids (Kirihara et al., Reference Kirihara, Luo, Chow, Misawa, Kawada, Shibata, Khoyratee, Vollette, Volz, Levi, Fujii and Ikeuchi2019; Osaki et al., Reference Osaki, Duenki, Chow, Ikegami, Beaubois, Levi, Nakagawa-Tamagawa, Hirano and Ikeuchi2024). Experiments could compare matched pairs of organoids—one group self‑generating energy (ATP) through metabolism, the other supported by external perfusion of metabolic substrates. Seth’s theory predicts that self-generating organoids should display more complex structures and oscillatory patterns. If instead the structural and functional complexity of matched pairs proves similar when provided equivalent energy, observers must either weaken their credence that consciousness emerges from autopoiesis or their belief in recurrent processing theories of consciousness.

3. 3. $Q:$ Causal closure, ${T_i}:$ Global neuronal workspace theory: Current neuroprostheses are capable of rapidly decoding speech from invasive brain measurements (Metzger et al., Reference Metzger, Littlejohn, Silva, Moses, Seaton, Wang, Dougherty, Liu, Wu, Berger, Zhuravleva, Tu-Chan, Ganguly, Anumanchipalli and Chang2023), and advances in decoding algorithms suggest noninvasive decoding may be feasible (Défossez et al., Reference Défossez, Caucheteux, Rapin, Kabeli and King2023; Tang et al., Reference Tang, Du, Vo, Lal and Huth2023). Noninvasive neuroprosthesis could enable experiments testing the inseparability of the brain’s “hardware” and “software” by recording subjects in three conditions: speaking aloud, having a neuroprosthesis decode silently mouthed speech, and having a neuroprosthesis decode imagined speech from brain activity alone. If ignition collapses in the final condition, such that semantic content associated with speech production no longer appears distributed throughout the brain as the motor components are offloaded to the neuroprosthesis, it would suggest either that substrate-specific implementation details play a crucial role in the mind’s function, or that global neuronal workspace theory requires revisions. Maintained integration across conditions would support substrate flexibility.

We recognize that committed adherents to either position could interpret these results flexibly. Biological naturalists might argue that negative results inadequately capture relevant biological properties, while functionalists might claim that positive results demonstrate necessary but insufficient conditions. Pre-registered predictions and adversarial collaboration protocols can minimize post-hoc rationalization, though experimental results rarely force belief revision when theories serve broader theoretical frameworks.

But history is littered with regrettable and discriminatory policies justified by scientific theories that were, at the time, contested (Naomi Oreskes, Reference Oreskes2021). We suggest that policymakers exercise epistemic humility when considering the moral and legal standing of artificial minds and ground policy decisions in the plausibility of all mainstream theories of consciousness.

Our proposed experiments offer a pathway to operationalize this debate and move policy beyond epistemic humility. If results favor computational or hybrid accounts, policymakers must prepare for conscious AI sooner than Seth predicts; if results uphold biological naturalism, regulators gain clearer limits. Whether the proposed experiments vindicate or falsify biological naturalism, they convert Seth’s largely metaphysical debate into a concrete research agenda.