A key opens a lock, and that is what makes it a key. It does not matter whether it is cut from brass, molded from plastic, or stored as a string of digits that a smart lock reads, because the thing that makes it a key is the role it plays, not the stuff it is made of. Functionalism applies that same move to the mind. It says a mental state is whatever occupies a certain causal role: a particular pattern of what tends to cause it, what it tends to cause in turn, and how it connects to other mental states along the way. On this view, asking what a mind is made of is the wrong question, the way asking what metal a key “really” is misses that any metal will do as long as it turns the lock.
The idea
Functionalism holds that mental states are individuated by their causal or functional role, what they DO, not by their physical realization, what they are MADE OF. Take pain: it is the state typically caused by bodily damage, which then produces beliefs and desires (something is wrong, I want out of this state) and issues in avoidance behavior. Anything wired so that injury-type input drives that mental-and-behavioral pattern thereby realizes pain. Hilary Putnam pushed this with multiple realizability: the same mental state can be realized by very different physical systems, a centralized vertebrate brain, an octopus’s distributed nervous system, perhaps a Martian’s green slime or a silicon robot. If one state has many physical realizations, then function, not material, is what makes a state the state it is. That premise is exactly what the optimistic side of the AI-consciousness debate needs.
A state is defined by what it does
The Stanford Encyclopedia puts the core thesis plainly: what makes something a mental state of a particular type “does not depend on its internal constitution, but rather on the way it functions, or the role it plays” in the system it belongs to. The contrast is with type-identity theory, which says a given mental state just is some specific physical state, pain just is a particular kind of neural firing. Functionalism backs off that claim. It locates the mental one level up, at the pattern of causes and effects, and treats the physical hardware as whatever happens to implement that pattern.
Pain is the standard worked example, and it is worth seeing why it fits the role-based picture so cleanly. The encyclopedia describes pain as the state that tends to be caused by bodily injury, to produce the belief that something is wrong with the body and the desire to be out of that state, and to cause behavior like wincing or withdrawal. That is a complete functional organization in three beats: an input (tissue damage, bodily harm), a set of internal mental connections (the belief plus the desire), and an output (avoidance behavior). Functionalism’s claim is that this organization is what pain is. Build a system, out of anything, in which damage-signaling input drives that same belief-desire-avoidance loop, and you have built something that is in pain, because there is nothing more to being in pain than occupying that role. This is the same architecture that the access side of consciousness is happy to grant a machine, and the question of whether the felt side comes along for free is exactly where the fight starts.
The same state, many bodies
The argument that turned functionalism from a tidy definition into a live position is multiple realizability, and it is Hilary Putnam’s. In his 1967 paper, originally titled “Psychological Predicates” and later reprinted as “The Nature of Mental States,” he challenged the brain-state theorists with a simple observation: pain is not the property of any one kind of nervous system. The Stanford Encyclopedia on multiple realizability states the thesis as “a single mental kind (property, state, event) can be realized by many distinct physical kinds,” and credits Putnam with introducing it into the philosophy of mind. The point lands by surveying how different the candidate realizers are. Mammals, reptiles, birds, amphibians, and molluscs have very different brain structures, yet we readily attribute pain across them. They can share that one mental state only if the state can be realized by many different physical states.
Putnam pressed the case past biology on purpose. If we say pain just is human neural firing of a particular type, we have to say in advance that an octopus, with its radically distributed invertebrate nervous system, cannot really feel pain, that an alien with green slime where we have neurons is a priori incapable of it, that a silicon-based android is ruled out before we even look. That looks like the wrong verdict, a piece of chauvinism about hardware. The natural fix is to define pain by its functional role and let the role be filled by whatever can fill it. Multiple realizability and functionalism prop each other up here: the many-realizers observation is one of the original arguments for functionalism, because the function is the thing all those different physical systems have in common when they are in pain. Type-identity theory pitches its account at the wrong level, too low, too tied to one kind of stuff; functionalism pitches it at the level the cases actually share.
From octopuses to silicon
The reason this note is load-bearing for everything downstream is the short step it opens. If a mental state is constituted by a functional organization, and that organization is substrate-neutral enough to run on a vertebrate brain and an octopus and a Martian, then there is no principled stopping point that excludes an artificial system. A non-biological device that instantiated the same input-to-belief-and-desire-to-output pattern would, on the functionalist account, thereby have the state. Putnam’s own early version was machine-state functionalism, drawn straight from the theory of computation, and the encyclopedias note that this lineage is why androids and electronic robots show up in the list of possible pain-realizers alongside the animals.
That step is the premise the optimistic camp builds on, often called computational functionalism: roughly, that the right kind of computation, running the right functional organization, is sufficient for mind, whatever it runs on. If that premise holds, then a sufficiently rich system like the ones the deep learning revolution produced is not disqualified from mental states merely for being silicon, and the live question becomes the empirical one of whether any actual system realizes the relevant organization. That is the thread the question of whether an LLM could be conscious picks up. The premise is also exactly what gets attacked: the biological substrate objection denies that function is enough and insists the wet biological stuff is doing essential work, and the worry carried over from the access versus phenomenal distinction is that you can copy the whole functional role and still have copied only the access side, leaving the felt side untouched. Functionalism is the bridge the whole debate is fighting over, and multiple realizability is the plank that holds it up.
Why this decides the moral-status question
The “Big Data and AI” ethics question of moral status runs directly through this premise. If functionalism is right that mental states are just functional organizations, and multiple realizability is right that those organizations can be realized in non-biological hardware, then an AI built to the right pattern would not merely simulate having mental states, it would have them, and a system that can be in pain is a system that can be wronged. The stakes of the philosophy are not academic: getting the realization question right is what tells you whether a future model is an instrument you can switch off freely or a subject whose interests count. The opposing pull is the substrate objection’s worry that we would be projecting a role-match onto something dark inside, granting moral standing to a very good imitation. The optimist’s bet and the skeptic’s caution both trace back to a single disagreement about whether function is enough, which is why so much of how we govern AI eventually has to take a position, explicit or not, on whether mental states are multiply realizable all the way to silicon.
Does the system realize the role?
- A human in pain. Tissue damage drives the belief that something is wrong, the desire to stop it, and withdrawal. The textbook realization, and the one every other case is measured against.
- An octopus recoiling from harm. A completely different, distributed invertebrate nervous system, yet the same input-belief-desire-output organization. If material decided the matter, this would not count; functionalism says it plainly does.
- A hypothetical Martian or android. Green slime or silicon where we have neurons. Putnam’s point is that ruling these out in advance, purely on hardware, is chauvinism rather than argument.
- A current AI system. It can produce the reports and behavior associated with the role. Whether it realizes the full functional organization, and whether realizing it is sufficient for the felt state and not only the access state, is the open question the rest of this wing argues over.
Related Notes
- consciousness-access-vs-phenomenal, the distinction functionalism tries to bridge, since copying the functional role may capture only the access side and miss the felt side
- could-an-llm-be-conscious, the live debate that computational functionalism sets up for today’s systems
- the-biological-substrate-objection, the main rival that denies function is enough and insists the biological stuff is doing real work
- scientific-theories-of-consciousness, rival proposals for what physically marks consciousness, several of which are themselves functional accounts
- physicalism-and-the-mind, the broader commitment that the mind is physical, which functionalism gives a particular shape
- deep-learning-revolution, what the artificial systems whose mental states are in question actually are and do
- ai-governance, where the moral-status question stops being abstract and starts shaping policy
Sources
- “Functionalism,” Stanford Encyclopedia of Philosophy. https://plato.stanford.edu/entries/functionalism/ . Supports that what makes something a mental state of a given type depends not on its internal constitution but on the role it plays, the worked pain example (caused by bodily injury, producing the belief that something is wrong and the desire to be out of the state, causing wincing or withdrawal), that pain can be realized by different physical states in different creatures, and Putnam’s machine-state functionalism drawn from early AI/computation theory.
- “Multiple Realizability,” Stanford Encyclopedia of Philosophy. https://plato.stanford.edu/entries/multiple-realizability/ . Supports the thesis that a single mental kind can be realized by many distinct physical kinds, the attribution of the idea to Hilary Putnam, his 1967 work challenging the brain-state theorists, the silicon androids / electronic robots / green-slime Martians as possible pain-realizers, and that multiple realizability was part of the early arguments for functionalism against type-identity theory.
- “Multiple realizability,” Wikipedia. https://en.wikipedia.org/wiki/Multiple_realizability . Supports the attribution to Putnam and his 1967 paper “Psychological Predicates,” that mammals, reptiles, birds, amphibians, and molluscs have different brain structures yet can share mental states only if those states are multiply realizable, that ruling out aliens, AI robots, and silicon-based life forms purely on differing neurochemistry is what Putnam questions, and the link from multiple realizability to functionalism and the possibility of mentality in artificial systems.