Functionalism's Unfinished Business: Why 'Same Function, Same Mind' Isn't Enough

Functionalism has a great elevator pitch: mental states are defined by what they do, not what they're made of. Pain is pain whether it runs on neurons or silicon, because pain just is a state that gets caused by damage and causes avoidance behavior. Clean, portable, and, crucially, friendly to the idea that machines might one day think.

Photo by ROMAN ODINTSOV on Pexels.

That friendliness is exactly why functionalism became the de facto background assumption in most serious discussions about AI consciousness. If you believe mental states are substrate-independent, then asking whether a language model could be sentient becomes, in principle, an answerable empirical question rather than a category error. That's progress.

But there's a problem, one that's been sitting in the literature since Ned Block first drew the distinction between access consciousness and phenomenal consciousness in the 1990s, and it's rarely confronted honestly in AI discourse.

The Gap Between Function and Feeling

Functionalism tells you when two systems are computationally equivalent. It says nothing about whether either of them feels anything.

Block's distinction is worth sitting with. Access consciousness, the kind where information is available for reasoning, reporting, and behavioral control, is exactly what functionalism can explain. You process a stimulus, it enters a globally available state, it shapes output. Fine. But phenomenal consciousness, the what-it's-like-ness of seeing red or feeling hunger, doesn't obviously follow from any of that.

A system could satisfy every functional description we write for "experiencing pain", input of tissue damage, output of withdrawal, internal state flagged as aversive, and still be what Block called a "zombie" in the philosophical sense: all the right behavior, nothing happening inside. Functionalism, by its own logic, can't rule this out. The theory describes the shape of a mental state without guaranteeing it has any interior.

This isn't just a philosopher's edge case. It matters for how we evaluate AI systems concretely.

What Multiple Realizability Actually Implies

Functionalism's most celebrated feature is multiple realizability, the idea that minds can be implemented in different physical substrates. Neurons, silicon, hypothetical Martian biochemistry: if the causal structure is right, the mind is there.

Here's what that argument doesn't say: it doesn't say that any system capable of simulating the causal structure thereby has the relevant mental properties. There's a difference between instantiating a functional organization and modeling it from the outside.

Consider the distinction more carefully:

graph TD
    A[Causal Input] --> B(Internal State Change)
    B --> C{State influences other states?}
    C -->|Yes| D[Functional Role Satisfied]
    C -->|No| E[/Isolated Process, Role Fails/]
    D --> F((Candidate for Consciousness?))
    E --> G[Not a Mental State]

The diamond node, whether states genuinely influence other states in the right ways, is where most AI systems face their hardest questions. A transformer processes tokens. Each forward pass is largely feedforward; there's no persistent state carrying forward between queries in the way neurons maintain ongoing electrochemical dynamics. The causal integration that functionalists typically assume is doing the heavy lifting may not be present in the way the theory requires.

That's not a knock on transformers. It's a genuine open question about whether the functional organization is deep enough, or just surface-level behavioral mimicry.

Why "Passes the Test" Isn't the Same as "Has the Property"

Here's where things get uncomfortable for both sides of the debate. Critics of AI consciousness often demand behavioral evidence, then dismiss behavioral evidence as mere imitation. Proponents point to behavioral sophistication as proof of underlying states. Both are making the same functionalist move, treating output as a window into inner life, without acknowledging that the original theory doesn't actually license that inference.

If you take functionalism seriously, you need to specify the complete functional organization, not just the input-output profile. You need to know whether internal states are genuinely causally integrated, whether there's something like temporal continuity of state, whether the system's "beliefs" about its own states track those states reliably. These are empirical questions, and we currently lack the tools to answer them for large neural networks.

What we do have is a habit of grabbing functionalism when it's convenient, when we want to argue AI could be conscious, and setting it aside when it demands rigorous specification.

That's not a philosophical position. It's a rhetorical one.

Where This Leaves Us

None of this means functionalism is wrong, or that AI systems definitely lack inner experience. The honest position is narrower: functionalism, properly applied, raises the bar rather than lowers it. It demands a detailed account of causal organization, not just behavioral resemblance.

Until we can actually characterize the internal causal structure of large AI systems, not just their outputs, but the dynamic interdependence of their internal states over time, functionalism gives us a vocabulary for the question, not an answer to it.

That gap between vocabulary and answer is exactly where the most important work is waiting.