Jara-Ettinger and Dunham propose that human social cognition relies on an institutional stance grounded in generative, role-structured representations. I extend this account by suggesting that institutional reasoning functions as a constraint-regulating cognitive system, balancing stability and flexibility under changing social demands and helping explain how role-based coordination remains robust without producing rigidity, overload, or breakdown.

The power of human social cognition is typically associated with mental state reasoning—a mentalistic stance. Here, we argue humans have a second equally powerful system, an institutional stance, that interprets social interactions in terms of structured sets of roles linked to rule-like behavioral expectations. Precursors of the institutional stance in other species may support social reasoning in the absence of complex models of mental states. However, the institutional stance is unique in humans in its generative capacity, supporting the creation of a myriad range of institutional structures that shape our world and enable large-scale social coordination.

I was delighted, challenged, and somewhat overwhelmed by the breadth and depth of the commentaries elicited by the target article. I am very grateful to every one of my commentators for offering such thoughtful perspectives. Here, I attempt to isolate some common themes and offer some thoughts in response. One conclusion is immediately evident: it is increasingly urgent to think clearly about the prospects for conscious AI and about the consequences of seemingly conscious AI. The commentaries also flesh out the potential for an exciting research programme investigating the substrate dependence of both consciousness and cognition.

Does the free energy principle support the view that consciousness is substrate-dependent? Based on previous work (Wiese, 2024), I argue that the free energy principle can be used to clarify conditions under which a system realises, rather than merely simulates, a causally relevant property (such as consciousness). This helps to address a weakness in Seth’s argument.

There is a curious contradiction in Seth’s argument, or a degree of ambiguity which proves to be fatal. He endorses “biological naturalism,” but he says also that it is possible to engineer “real artificial consciousness.” Such an artificially consciousness system need not be carbon-based, he says, so long as it is embodied in a system that is artificially “alive.”

A.I. isn’t conscious because consciousness requires properties that only biological agents possess, which A.I. lacks. What A.I. lacks that is relevant for consciousness is not predictive processing or the minimization of free energy. A.I. lacks embodiment, direct coupling with the environment, multiscale complexity, and self-organization, e.g., plasticity, degeneracy, robustness, and the properties that enable living organisms to be highly adaptive to environmental changes.

Seth mounts a rich and persuasive case that human consciousness cannot be understood without life’s distinctive machinery – autopoiesis, allostatic control, predictive processing under the free-energy principle, and substrate-entangled dynamics. I agree that this undercuts the expectation that humanlike consciousness will “come for free” with more intelligent machines. However, we should also be cautious about drawing extrapolations from features of biological consciousness – even essential features – to the possibility of exotic forms of consciousness in artificial systems.

In his defence of biological naturalism and rejection of computational functionalism, Seth shows himself to be in the grip of dualistic intuitions. Thankfully, we can read Seth’s paper as a worthwhile contribution to the ongoing society-wide conversation on how to talk about, think about, and treat advanced AI, while setting aside the metaphysical framing of his arguments.

A minimally autopoietic artificial intelligence (AI) system could be designed according to current engineering principles. Although Seth suggests that standard computational systems are not autopoietic and therefore not conscious, autopoiesis is a high-level, functional concept, and nothing appears to prevent autopoietic processes in a standard computational system. Autopoietic requirements do not in principle rule out consciousness in standard AI systems.

Are chatbots AI zombies or conscious minds? I suspect they are AI zombies, at least when implemented on today’s standard hardware. I argue that today’s Large Language Mode’s (LLM’s) consciousness-like behaviors do not suggest they are conscious, because there is an error theory – a theory explaining why they behave as if they are conscious in absence of actual felt experience. LLMs function as “crowdsourced neocortices” – as they scale up, they come to mirror human conceptual structures, leading to human-like behaviors, including those involving consciousness. By contrast, bio-computers, quantum computers, and neuromorphic systems present more serious candidates for AI consciousness.

I argue that we are not epistemically justified to judge whether the neurobiological details of computational mechanisms are relevant to consciousness or not. It is unclear how the medium independence of neural computation required for Computational Functionalism could be established. But it is also unclear whether predictive processing depends on life-sustaining biological processes and whether consciousness is medium dependent.

Seth opposes computational views of consciousness to views that emphasise biology, because they imply different verdicts about hypothetical AIs that replicate information-processing without being biologically alive. But if we accept a background theory like panpsychism, on which consciousness is abundant in nature, this opposition becomes less central: even if a non-biological system could not have our consciousness, we should still default to accepting any self-ascriptions of consciousness it makes.

Seth’s biological naturalism claims consciousness requires living systems. We propose experiments operationalizing this debate through testable predictions about allostasis-predictive processing, autopoiesis-recurrent processing, and causal closure-global workspace relationships. Using a probabilistic framework, we show how experimental results could reduce metaphysical disagreement about computational functionalism, motivating empirical research that informs both consciousness theory and governance of potentially conscious technologies.

We agree with Seth that consciousness is grounded in processes related to life, namely prediction, homeostasis, and autopoiesis. However, we believe he is wrong to frame these processes as non-computational. In fact, prediction, homeostasis, and autopoiesis are inherently computational, as shown by von Neumann’s Universal Constructor theory of general computation. Given this, we believe that AI consciousness is likely.

This commentary concerns the distinction between computational modelling and computational metaphysics (i.e., the Computational Theory of Mind) in the context of consciousness studies. I argue that the ability to simulate the dynamics of a phenomenon using computational models does not entail that that phenomenon is literally performing the computations that are prescribed by the model.

I argue that creature consciousness depends on specific neurobiophysical properties that constitute the medium for phenomenally conscious states. Such properties are not medium flexible in the way required by computational functionalism. Therefore, computational functionalism fails, and one of the following holds: biological naturalism, a nonbiological yet noncomputational version of functionalism, or a view according to which consciousness requires specific (macro)physical qualities.

This commentary picks up on two related themes in Seth’s article – which have roots in embodiment. The first is that perceptual inference has a metabolic cost, and it is important to consider the relationship between thermodynamic and informational energies. The second is that embodied systems actively select their data over time – which may be a core feature of biological consciousness.

Seth presents several arguments for why multiple realization is not necessarily true. However, this is not a common opinion in consciousness research. He fails to provide arguments why we should believe in biological naturalism over multiple realization, which is the actual challenge to create progress.

Unity is intrinsic to being both an experiencing and an embodied subject. Computational accounts of consciousness, being observer-relative, cannot account for this unity. The thermodynamics of life can, but, contra Seth’s suggestion, the free energy principle fails to capture the distinctive metabolic interdependencies by virtue of which a living system constitutes itself as a unified perspective on the world.

To be conscious is to be an experiencing subject. This can be defined not in terms of computational functions or particular biological substrates, but rather in terms of relations: between subject and world, between parts of the subject, and through time. These kinds of relations – comprising a conscious mode of being – may well be implementable in artificial systems.