Advances in artificial intelligence (AI) add new urgency to an old question: is there a deep connection between life and consciousness? Anil Seth has presented a bold argument in favour. It won’t break the deadlock, but it does suggest possible ways forward.

Biological naturalism can be a slippery target because the number of possible biological properties that might matter to consciousness is large, and the biological naturalist can always shift their gaze to a different property if their initially favoured candidates are recreated in AI. Differences in agency and embodiment that seem stark now might look much less so in a few years. There will, however, always be some differences. AI systems will never fully recreate cell-level metabolic processes. Consciousness might depend constitutively on these processes, but are there good reasons to think that it does?

Seth, in §4, weighs in with a novel argument from free energy minimization. He suggests that there is a crucial difference between a perceptual inference algorithm running in a biological system that works to minimize free energy right down to the cellular level and the same algorithm running in a non-living substrate that is not working to minimizing free energy at lower levels. The latter is said to be a “pale shadow” of the former, making it much less likely to support consciousness.

I see two main problems for this argument. One is that it begs the question against a computational functionalist opponent, who will deny the “pale shadow” thesis. They will not accept that what happens at lower levels of organization makes a difference to an algorithm’s ability to support consciousness.

The second is that the free energy principle is too speculative a foundation to carry this much weight. As I see it, living things minimize prediction error when it serves their inclusive fitness interests, but when it doesn’t, they don’t. The “dark room problem” has been much discussed: organisms do not generally hide in dark caves even though this renders the external environment maximally predictable (Sun & Firestone, Reference Sun and Firestone2020). Fristonians have a fair comeback: staying in the cave renders the internal environment increasingly unpredictable because you cannot maintain homeostasis when starving (Seth et al., Reference Seth, Millidge, Buckley and Tschantz2020). But to this we can add the inverse dark room problem: organisms sometimes do starve to death in dark caves when it serves their inclusive fitness interests. Female octopuses do this to protect their eggs (Wang, Reference Wang2023). The only way to make sense of why organisms sometimes prioritize homeostasis and sometimes prioritize something else (such as mating opportunities, offspring, or other relatives) is through the principle of inclusive fitness maximization. To the extent biological processes optimize any single variable, it is inclusive fitness, not free energy (Grafen, Reference Grafen2006). There is no rhyme or reason to what organisms do except the drumbeat of natural selection. Friston’s claim to have found a second principle of equivalent or greater stature is exasperatingly under-evidenced.

Seth may reply: yes, the argument is speculative, but arguments for computational functionalism are speculative too. See, for example, Chalmers (Reference Chalmers1996) on fading and dancing qualia. Those arguments often beg the question against a biological naturalist opponent (see Birch, Reference Birch2024, ch. 3, note 55). Both biological naturalism and computational functionalism are speculations lacking empirical validation. Neither deserves to be the default. All can agree we need ways to put claims about the life-consciousness connection, or lack of one, to empirical test.

Can the hypothesis of a constitutive relationship between consciousness and a biological property ever be tested? A worry arises that such hypotheses might be untestable. The challenge is that we would need to test for direct dependencies of phenomenology on biology, unmediated by effects on computation. If all the dependencies we can find are mediated, they will be compatible with computational functionalism. For example, disrupting the functioning of layer V pyramidal neurons will disrupt consciousness, but if the disruption is mediated by disruption to computations, computational functionalists need not worry. But how we could ever show a phenomenological effect to be unmediated by an effect on computation? For the effect to be reportable at all, there must be some difference at the level of computation, since report is itself a computational process.

The problem is serious, but I don’t see it as insurmountable. To see grounds for optimism, first note that some claims about the life-consciousness connection can be falsified. Think, for example, of Beck and Eccles’ (Reference Beck and Eccles1992) proposed link between consciousness and quantum tunnelling during exocytosis. This specific hypothesis has dropped out of the discussion because it entailed predictions that were shown to be false (Georgiev & Glazebrook, Reference Georgiev and Glazebrook2014).

The hypothesis that consciousness depends on ephaptic coupling between neurons (mentioned briefly by Seth) also strikes me as testable. Myelination inhibits ephaptic coupling, so if it were essential to consciousness, we would expect to see fewer signs of consciousness in species with high degrees of myelination, such as cetaceans and elephants. This already looks implausible.

If life and free energy minimization were inseparable, it would be impossible to test for correlations between free energy minimization and consciousness in living things. But if they are, as I suggested earlier, readily separable, then testing for such a correlation becomes possible. We could investigate whether organisms pursuing free energy minimization display more markers of consciousness than those sacrificing their own homeostasis to maximize inclusive fitness.

So, specific versions of biological naturalism can be tested. This shows how, over time, we might gain traction on the bigger question of whether any version is correct. For if we repeatedly test different versions and find them wanting, we will be entitled to interpret the long-term pattern of failure as tilting the odds towards computational functionalism.

Granted, there will always be more biological properties that might matter, so we will not reach conclusive proof. But since computational functionalism predicts a pattern of failure when we look for dependencies of consciousness markers on any biological properties other than those involved in implementing computations, such a pattern would count in its favour.