What is consciousness? Seth characterizes it in terms of feeling (“what it’s like”) rather than function (“what it does”). The conceptual distinction between feelings and doings has seemed clear to philosophers of mind, who have relied on it to generate many thought experiments. Seth mercifully pays little attention to such thought experiments, thereby signaling that he sets little store in them.

Despite their conceptual separability, feelings often do things (epiphenomenalism notwithstanding). Painful stimuli lead conscious organisms to avoid situations that cause them and to ingest pain-relievers when available. Dichotomous characterizations of consciousness in terms of feelings independently of functions can only be regarded as partial at best.

Many of the arguments others have given in support of the idea that conscious AI is imminent are diagnosed by Seth as presupposing computational functionalism about consciousness. Because this presupposition begs the question against Seth’s preferred bionaturalist view, it gives him a reason to be unmoved by those arguments.

Previous versions of biological naturalism leave the connection between biology and consciousness mysterious. Seth attempts to rectify this by adopting the proposal that the free energy principle and predictive processing views of conscious mental contents are linked. This link goes via the shared mathematics underlying, on the one hand, the requirement that biological “autopoietic” systems minimize free energy to combat entropy, and on the other hand, the minimization of perceptual prediction errors via active inference.

Seth faces the obvious objection that something approximating active inference can be (and has been) implemented computationally. Because active inference seems like a “doing,” something also needs to be said about why it becomes a feeling when implemented biologically but not when given a computational-functional implementation. His suggestion is that computational implementation of active inference is just a crude facsimile of what brains do. The idea seems to be that the processes involved in active inference and the energetics involved in combating entropy are tightly coupled in living organisms, right down to the level of individual cells. Computers instead instantiate a kind of “eternal” computing that is not strongly coupled to any energetic or temporal demands.

The relative detachment of Turing-style computation from the battle against entropy may be characteristic of current computational implementations of active inference. But is this a principled difference between computational and biological systems, or just a difference of degree? Because Bayesian calculations are computationally intractable, they can only be approximated in both brains and computers. The dilemma for Seth’s version of biological naturalism is that either consciousness is at least partly due to (tractably) approximating a computable function, or it isn’t. If it isn’t, then the free energy principle doesn’t explain how to get from approximating active inference to feelings. If it is, then it needs to be explained why some Bayesian approximators (biological organisms) are conscious but some (computers) aren’t. Without such an account, the FEP-bionaturalist equally begs the question against the computational functionalist.

Perhaps two sides begging questions against each other reflects both normal metaphysics and pre-paradigmatic science. Any attempt to resolve questions about what something “really” is requires interpretation of evidence through the lens of some theory or other, and how things are interpreted from one theoretical perspective will beg questions against the others. There are no theory-independent standpoints on the evidence, although there may be relatively theory-neutral descriptions of the evidence (such as that physicists can agree on the radiation measured coming from a black hole without agreeing on a theory of black holes).

But where does this leave the science of consciousness? Without a consensus theory, or even an agreed upon characterization of consciousness, the best we can do collectively is to encourage computationalists and biological naturalists – indeed, anyone who agrees on some basic scientific precepts about holding their theories accountable to evidence – to continue refining experiments and measurements to probe the putative cases of consciousness that confront us: human, animal, or (most controversially) machine.

As Seth humbly admits, no one is in a position to say what really matters for consciousness or how fine-grained the analysis of the underlying mechanisms needs to be. Hence, we are not in a position to rule any theory off the table so far, either a priori or on the basis of the limited evidence that we have. This makes it okay and perhaps inevitable for different scientists to play their own hunches: different predilections will lean toward different attempts to unravel the mystery of how and why some systems are conscious.

What would not be okay, as Seth surely agrees, is to force hegemony at this early stage of investigation. The horizon for the development of field-defining theories is long – typically on the order of a century or more – and it is plain that the scientific study of consciousness is nowhere near such consensus. However, even incorrect theories can yield good techniques and measurements. The current pluralism among consciousness researchers is thus to be encouraged.

With this target article, Seth has done a great service to the field, making an excellent case against taking computational functionalism for granted, and doing a similarly excellent job of explaining the lines of thought in favor of taking the complexity of biological systems more seriously than computationalists typically do. I agree with his hunch that the degree of complexity found in brains matters for both intelligence and consciousness, but better measurements and better theories will need to be invented before anyone can say which details matter and whether implementing such complexity requires more than Turing computation.