AI consciousness is interesting and important. Some people reject AI consciousness because they accept biological naturalism, the thesis that some biological process is necessary for consciousness. This post argues against biological naturalism. We think that there are a number of strong arguments against biological naturalism and no good arguments for it.
This essay will criticize biological naturalism. This theory claims that some biological process is necessary for consciousness. The relevant process must be present in biological life, but absent in AIs.
Debates about biological naturalism often connect to * functionalism*. According to functionalism, mental states arise from functional roles, complex patterns of causes and effects. These profiles of cause and effect can include (i) features of the environment causing the mental state, (ii) the mental state causing and being caused by other mental states, and (iii) the actions of the entity that are caused by the mental state.
An alternative to functionalism is the “identity theory.” According to the identity theory, mental states are identical to specific physical processes. In humans, for example, dull bodily pain is associated with firing of a special kind of nerve called a C-fiber. An identity theorist might then claim that dull pain simply is C-fiber firing.
Biological naturalism is compatible with either identity theory or functionalism. Rosa Cao, for example, argues that the best version of functionalism will involve extremely specific functional roles. These functions are so specific that our best scientific laws imply they can only be realized by biological processes, such as literal neurons, glial cells, and neurotransmitters. Since AIs don’t use these biological processes, they don’t have states playing these hyperspecific functional roles.
One competitor to biological naturalism is computational* *functionalism. According to this view, the relevant kinds of functional profiles are computational. This means that the relevant causes and effects have the structure of mathematical computations, or computer programs. For example, maybe they implement basic logical operations like and, or, and not.
Which functional role is associated with consciousness? This is an active debate. On one approach, an entity is conscious when it has a global workspace. This means that it processes information through a series of parallel modules, which pool their information into a global workspace, which processes this information, makes it consistent, and broadcasts it across the whole brain. On another approach, an entity is conscious when it has higher order representations: mental states that represent its own mental states. On another approach, consciousness requires *recurrent *processing: feedback loops where later perceptual areas send signals back to earlier ones.
We’ll now lay out some arguments against biological naturalism.
The most famous argument for functionalism is * multiple realizability*. This is the idea that many different entities made of many different things can have minds. We worry that biological naturalists struggle with multiple realizability.
One application of multiple realizability concerns animals. Octopi, sharks, eagles, iguanas, dogs, and humans all have minds. Each has perception, belief, desire, intentional action, consciousness, and pain. But there are all sorts of differences in the neurology of these organisms. For example, bodily pain in humans is realized by the firing of C-fibers. But not all of these various organisms have C-fibers. If pain just was C-fiber firing, then only a small group of entities would have pain.
So the question for biological naturalists is *which *biological process is necessary for consciousness. What is the guarantee that all these different animals use the same biological process? For example, Cao 2022 suggests that the functional roles associated with cognition are extremely fine-grained. Cao surveys the ways in which the input/output patterns of neurons in humans depend on a host of non-neuronal factors, including chemicals and glial cells.
But the specific pattern of neuronal firing in each type of animal will differ from one another. Different animals will use different chemicals in different ways: insects use octopamine in the way we use noradrenaline. Birds lack a layered neocortex, and yet some birds match primates on cognitive tasks. We worry that any approach which claims that very specific functional roles are necessary for possessing various mental states will have trouble allowing many different animals to satisfy that role. As we’ll see below, a different approach is to instead simply superadd a biological condition (for example: the organism must metabolize) on top of relatively coarse functional roles. This approach allows various animals to have mental states. But now the biological condition appears to be a spandrel (as we’ll argue below).
Another application of multiple realizability concerns aliens. Imagine that we went to Alpha Centauri, and found an alien civilization similar to our own. Imagine that we coexisted with this civilization for centuries, and then suddenly learned that the aliens are made of silicon rather than carbon, or in various other ways lacked the relevant biological process. The aliens would plausibly be conscious despite being made of different stuff. But it looks like biological naturalists will have to deny this.
Our second concern with biological naturalism is that it makes false predictions about other mental states besides consciousness.
Here, biological naturalists face a choice. Are biological processes only necessary for conscious mental states, or are they required for other kinds of mental states too? This will depend on why someone is a biological naturalist in the first place.
Again, Cao 2022 motivates biological naturalism as a general thesis about the mind. The argument is that the functional roles associated with neuronal processing are too fine grained to be realized by computer chips, because they involve all sorts of chemicals and other cells besides neurons. But this has nothing to do with consciousness specifically, it is about cognition more generally.
So there is some pressure for biological naturalists to deny any mental states to AIs. On this view, Claude doesn’t see. Claude doesn’t believe, think, or wonder. Claude doesn’t introspect. Claude doesn’t desire, want, intend, plan, or act intentionally.
These further claims are harder to swallow than the claim that Claude lacks consciousness. Unlike consciousness, these other mental states come with fairly clear patterns of behavior and information process. Computational approaches have been amazingly empirically successful at explaining many distinct features of the mind.
An individual is conscious if and only if they have conscious mental states. A theory of consciousness should not only say which individuals are conscious. It should explain this by explaining which mental states are conscious.
Humans have unconscious desires and conscious desires. Humans have conscious perceptual states and unconscious perceptual states. For example, if you flash an image very quickly, humans will not verbally report seeing the image. What explains this?
Functionalist theories of consciousness are in the business of explaining these differences. According to higher order theories, the conscious mental states are those that are represented in a special way. According to global workspace theories, the conscious mental states are those that are globally broadcast. So far, the main theories on the market that explains the differences between conscious and unconscious mental processes does so in terms of relatively general functional roles, which could be realized in principle by AIs.
We worry that biological naturalists struggle to explain which mental states are conscious. As we’ll discuss in the next section, biological naturalists often say that an entity is conscious only if they have some general biological property, such as metabolism or autopoiesis. But this property generally does nothing to distinguish conscious from unconscious mental states.
Here, biological naturalists might *conjoin *functional conditions to biological conditions. Maybe a mental state is conscious if and only if it is globally broadcast *and *the entity that has the mental state is biological.
In general, one way to test a theory is to evaluate its explanatory power, compared to its simplicity. How many successful predictions does the theory give us, and how many degrees of freedom does the theory have. We worry that the conjunctive approach fares poorly by these lights. It looks like all of the explanatory power is coming from the coarse functional roles. The biological components of the theory are dead weight.
Here, a broader point is that when we explain human behavior, we often appeal to the difference between conscious and unconscious mental states. Action explanations, for example, typically involve what the agent was aware of, what they noticed, how they felt, what things looked like, and so on. These explanations seem to all involve functional conditions. For example, global workspace theory says that the information relevant to action will be globally broadcast. We think that biological properties play no role in such explanations.
Next, we worry that the details of specific biological naturalist theories tend to face serious problems. The conditions offered do little to distinguish conscious and unconscious mental states. They have no obvious conceptual connection to consciousness. They are vulnerable to counterexamples from thought experiments. The biological conditions are insufficient for consciousness, raising the question of what must be added to them to get consciousness, and why this further fact isn’t itself sufficient for consciousness. In addition, the conditions tend to be vague, in a way that makes them hard to apply. Finally, each biological naturalist offers their own condition, without much explanation of why we should prefer this condition over a rival. Let’s look at a few examples.
First, Peter Godfrey-Smith has suggested that metabolism plays an important role in consciousness (he later clarified that he doesn’t think it’s strictly needed, but is more an instance of “here is what it is, in us and other animals, that might enable us to get a grip on it.”)1 Metabolism is the sequence of chemical reactions that convert food into energy. Here are some worries.
Metabolism does nothing to distinguish conscious from unconscious mental states.
Metabolism has no obvious conceptual connection with any feature of consciousness. For example, metabolism seems unrelated to the character of visual experiences, to the ability to access and broadcast information, to having a point of view, and so on.
Metabolism fares poorly with thought experiments. Imagine aliens that got power from fusion, without metabolizing, but otherwise acted relevantly like humans. Imagine that when we did cognitive science on these aliens, they appeared to have a global workspace, higher order thoughts, and various other computational functional properties associated with consciousness. It seems surprising that the lack of metabolism would be the barrier to consciousness for these entities.
Metabolism is insufficient for consciousness, since bacteria metabolize. So some other condition besides metabolism will also be necessary. But then what role is metabolism playing in the explanation of consciousness? Is it an independent, superadded condition, or is it related to this further condition in some way?
Second, Cao et al 2026 claim that an individual is conscious only if they are a “valuing subject,” with internalized optimization processes related to persistence that evaluate how well things are going for the individual. Again, similar worries apply.
Evaluation or valence does not distinguish conscious from unconscious mental states; you can look at a white wall without any kind of evaluation, and you can have unconscious mental states with evaluative content. Here, we disagree with Cao et al, who claim that a piece of information that is broadcast globally “must be evaluatively relevant for a valuing subject.” By contrast, we would say that often
evaluative content is relevant to broadcast, but this isn’t always the case.Evaluation seems conceptually independent from consciousness.
There are thought experiments that allow for consciousness without evaluation: imagine “
Vulcans,” a hypothetical being with conscious experiences but which lacks valenced experience altogether.This kind of evaluation is insufficient for consciousness: bacteria may have these kinds of evaluative representations; but bacteria aren’t conscious. Again, we wonder what further condition would be added to evaluation to get consciousness, and we wonder why this further condition would not itself be sufficient. For example, Cao et al suggest that the evaluations have to be internalized, integrated, and centrally broadcast; but a rival view will say that information like that is conscious regardless of whether it is evaluative.
Third, Anil Seth suggests that consciousness requires autopoiesis. Autopoietic systems are ones that “continually regenerate their own material components through a network of processes, and because this network of processes actively maintains a boundary between the system and its surroundings.” Our remarks are again similar.
Autopoiesis doesn’t distinguish conscious from unconscious mental states.
Autopoiesis has no obvious conceptual connection to consciousness: to having experiences, to having a point of view, and so on. Whether a thing has experience doesn’t seem to be affected by whether it draws a boundary between it and its surroundings.
Thought experiments are again troublesome for autopoiesis. Imagine we encountered aliens that lacked a clear distinction between themselves and the world, but nonetheless had extremely complex cognitive capabilities and perceptual faculties, and reported introspective access with qualitative details. As usual, we don’t think that the presence or absence of autopoiesis would be especially relevant to assessing whether they were conscious. If we saw aliens that were like people on LSD highs, where they didn’t have a distinction between themselves and the world, would we deny they were conscious? Or again, if there was a life form that was intelligent, acted like it was in pain, but was made out of metal and so didn’t regenerate its material, would we confidently deny that it was conscious?
Autopoiesis is surely insufficient for consciousness, and so again we wonder what else must be added to autopoiesis to get consciousness, and why this further ingredient wouldn’t itself be sufficient.
Finally, when we lo…
