The target article draws a sharp contrast between intelligence and consciousness. Intelligence, Seth maintains, is largely indifferent to the physical stuff in which it is realised, so artificial general intelligence (AGI) may emerge in present-day silicon. Consciousness, by contrast, might be biologically entrenched: its appearance in machines is correspondingly doubtful. Much in the paper builds on this proposed asymmetry, yet the gap narrows once we examine a concession Seth makes briefly in §3.8. There he allows that some intellectual capacities might themselves resist substrate flexibility. To preserve his thesis, he limits “intelligence” to abilities that can be exhaustively characterised as mappings from sensory inputs to behavioural outputs. That restriction, we suggest, excludes some of the capacities – goal‑directed planning, reasoning and linguistic competence – that initially motivated the cleaving off of consciousness from intelligence.

Consider planning. A system that selects means to an end does more than correlate inputs with outputs: it represents a goal and evaluates prospective actions in its light. Having a genuine goal may furthermore require having genuine needs, perhaps even metabolism (Godfrey-Smith, Reference Godfrey-Smith2019; Westlund, Reference Westlund and Ferrero2022; see also Jeuchems & Summerfield, Reference Juechems and Summerfield2019). Similarly, logical reasoning operates over contentful states. Whether the content of a belief or of a proposition can be reduced to input-output characteristics remains controversial. Finally, linguistic competence is not just a matter of outputting sounds or tokens. It plausibly requires communication and perhaps understanding. With respect to all of these mental properties, we currently remain unsure as to the appropriate characterisation in terms of function, algorithm, or implementation, and whether the capacity has sufficient substrate flexibility to be realised in silico.

These points already cast doubt on the postulated asymmetry, but a further difficulty arises when we notice that several capacities Seth labels as “intelligent” are empirically entangled with consciousness. Metacognition, for instance, shows higher efficiency when the underlying task representations are conscious rather than non-conscious, even under conditions when first-order (input-output) performance is matched (Michel, Reference Michel2023). Some forms of instrumental learning and control appear to depend on being conscious of the stimulus (Skora et al., Reference Skora, Yeomans, Crombag and Scott2021). The psychological functionality enabled by internal states such as mind-wandering and imagination may be hard to divorce from the subjective aspects of those states (Christoff et al., Reference Christoff, Irving, Fox, Spreng and Andrews-Hanna2016; Dijkstra, Bosch, & van Gerven, Reference Dijkstra, Bosch and van Gerven2019). If a capacity depends on consciousness, it inherits whatever substrate dependencies consciousness has. The fortunes of intelligence and consciousness may be interwoven more tightly than Seth’s framing implies.

So far the moral is that for many aspects of intelligence, as for consciousness, the details of implementation and their substrate flexibility remain open empirical questions. How then should research proceed? Seth’s §5 lists several theoretical options, and §4 canvasses properties that might be characteristic of consciousness – predictive processing, autopoiesis, affective dynamics, and various informational measures. What is missing is an explicit test for exclusivity: are these properties present only when consciousness is present, or do they covary with other cognitive capacities as well? Predictive processing, for example, is plausibly a ubiquitous principle of neural computation (Keller & Mrsic-Flogel, Reference Keller and Mrsic-Flogel2018); autopoietic organisation characterises biological organisms in general (Di Paolo, Reference Di Paolo2005). To decide such matters, we need a research strategy capable of tracking which properties travel together across a suitably broad sample of systems.

We propose that the appropriate strategy is the natural kind method (Shea & Bayne, Reference Shea and Bayne2010; Shea, Reference Shea2012; Bayne & Shea, Reference Bayne and Shea2020). The guiding question is whether properties that are potentially connected with consciousness do in fact reliably co‑occur. Do consensus measures of conscious state – such as subjective vividness, qualitative similarity relations, and temporal structure – cluster together? A finding that some collection of putatively consciousness-related properties cluster together provides a prima facie case that they are due to the presence of an underlying natural kind. We can then ask about the contours of that kind and how they are related to other aspects of psychological, neural, and biological function. Does this cluster relate to psychological functions such as learning, planning, or metacognition? Do specific patterns of predictive coding (for example, a distinctive balance of feedforward and feedback error signals) accompany conscious perception across conditions (across experimental paradigms, types of people, and other species)? Are metabolic signatures such as adenosine triphosphate consumption or mitochondrial activity part of the cluster? By mapping such covariances, we may study the underlying kind in advance of understanding its metaphysical underpinnings.

A signal advantage of the natural‑kind approach is that it is entirely compatible with computational modelling. Indeed, computational tools are indispensable. They allow us to articulate precise hypotheses about the algorithms that mediate conscious and non‑conscious processing, to simulate the effects of focal lesions or altered connectivity, and to quantify the putative signatures in large datasets. More generally, computational characterisations of conscious inference, global broadcasting, or affective appraisal provide testable targets for neuroimaging and for attempts to implement the same functions in neuromorphic hardware or deep learning architectures. Such findings neither require nor refute a biological theory of consciousness; they simply add a dimension along which conscious and non‑conscious states plausibly diverge. Crucially, computational investigations contribute explanatory traction whether or not computational functionalism is the correct metaphysical thesis.

Seth is surely right that we should aspire to integrate computational descriptions with biological ones. Yet the moral is symmetrical: it applies just as forcefully to advanced cognition as to subjective experience. Goal pursuit, reasoning, and natural language understanding all invite a joint analysis in which computational structure, energetic demands, and neural implementation are examined together. Until such integrative work is done, we cannot know whether any particular cognitive architecture – silicon‑based or otherwise – supports the requisite internal organisation.

The upshot is that an alleged asymmetry between intelligence and consciousness has not been established. When Seth restricts “intelligence” to capacities definable solely by input–output relations, he quietly excludes many of the very abilities that matter for AGI. Once those abilities are restored to the picture, the methodological terrain is level: both consciousness and intelligence require integrated inquiry that keeps an open mind about substrate dependence and that exploits computational modelling as a primary source of evidence. The most promising path forward is to develop computational models of consciousness with the same energy that cognitive science now devotes to models of aspects of intelligence and to test whether the resulting signatures line up with existing data. The natural‑kind method offers a principled way to conduct that inquiry, seeking lawful clusters among behavioural, computational, and biological properties without assuming from the outset which of them is essential.