Executive Summary
Cao’s central claim is that multiple realizability (MR) — the idea that the same mental state could be built out of very different physical stuff — is a much worse bet than philosophers have assumed. The paper argues that because of how deeply intertwined biological functions are in the brain, the physical material matters far more than functionalists typically acknowledge. You can’t just swap out neurons for silicon chips and expect to preserve the mind, because the brain’s functional organization is inseparable from the messy biological details that produce it.
This isn’t an argument against functionalism outright. It’s an argument that functionalism, taken seriously, pushes us toward biology rather than away from it — that the “spirit” of functionalism is actually more constraining than its proponents have realized.
The Setup: What’s at Stake
Multiple realizability (MR) is the thesis that the same kind of mental state — say, pain — could in principle be “realized” (that is, physically built or instantiated) by very different physical systems. A human brain, an octopus brain, a silicon computer — if they have the right functional organization (the right pattern of causal roles and relationships among their parts), they could all be in pain, regardless of what they’re made of.
This idea has been hugely influential. It’s the backbone of functionalism — the view that what makes something a mental state isn’t what it’s made of, but what it does, how its parts relate to each other, to inputs from the world, and to behavioral outputs. Functionalism is also what underwrites a lot of the optimism in AI and cognitive science: if the mind is defined by its functional organization rather than its biological substrate, then in principle we could recreate it in silicon.
Cao thinks this optimism is empirically unwarranted. Not because functionalism is wrong in principle, but because the actual functional organization of the brain turns out to be far more entangled with its specific biological properties than the standard picture suggests.
The Core Argument, in Pieces
1. The Neural Replacement Thought Experiment
Cao starts with a famous scenario from Chalmers: imagine replacing your neurons one by one with silicon chips that have the same input-output function (that is, they take the same inputs and produce the same outputs as the original neuron). Eventually your whole brain is silicon. Functionalists say you’d still have the same mind.
Cao’s response: actually doing this — even in principle — is way harder than it sounds, for reasons that come from real neuroscience rather than armchair philosophy.
2. Why Neurons Are Hard to Replace
This is where Cao gets into the empirical weeds, but the upshot is accessible. Neurons aren’t simple on/off switches. They’re spatially extended, history-sensitive, chemically complex analog devices. To actually duplicate a neuron’s function, your silicon chip would need to handle:
Spatiotemporal complexity — the where and when of a neuron’s activity matters. Signals arrive at different points on the neuron’s branching structure, and their spatial distribution affects the output. A silicon replacement would need to replicate this geometry.
Chemical signaling — neurons don’t just communicate via electrical spikes. They use a huge variety of chemical signals, including molecules like nitric oxide that literally diffuse through tissue in all directions. A silicon chip has no natural way to detect or produce these signals.
Biophysical sensitivities — neurons respond to temperature, blood flow, and other biological conditions. These aren’t noise; they’re part of how the system functions.
Self-modification — neurons change themselves over time (this is basically what learning is). They grow new connections, alter their sensitivity, change which proteins they express. The biological neuron doesn’t store this “record” in an explicit database — it’s implicit in the cell’s biochemistry.
Non-neuronal cells — glia (support cells in the brain), blood vessels, and immune cells all play functional roles in cognition. You can’t just replace the neurons and call it a day.
3. Generative Entrenchment: Why Biology Makes Things Worse
This is a key concept in the paper. Generative entrenchment (a term from Wimsatt) refers to how, in evolved systems, components that were originally there for one purpose get co-opted to serve many purposes over evolutionary time. Think of it like a city where the same roads serve commuters, delivery trucks, emergency vehicles, and parades — you can’t redesign the roads for just one of those functions without disrupting all the others.
In the brain, this means individual molecules and cells play many different functional roles simultaneously. Cao calls this multiplexing — one component, many jobs. ATP (the cell’s energy molecule) is a great example: it powers the cell, but it also serves as a signaling molecule between cells, modulates neural activity, and regulates sleep pressure. You can’t replace it with something that only does one of those jobs.
The upshot: because biological components are multiply employed, the constraints on what material can play their role multiply too. The more jobs a part does, the fewer alternative materials can do all of them at once.
4. Can We Loosen the Constraints?
Cao acknowledges that defining the “behavioral profile” at maximum fine-grained detail would make nothing multiply realizable — which proves too much. So the question becomes: is there a useful intermediate level of description (not neuron-by-neuron detail, but not just “behaves roughly human-like” either) at which we can specify the functional organization, and at which it could be realized in different stuff?
Cao’s answer is: maybe, but we don’t have good evidence for it yet, and the burden of proof is on the functionalist. Standard functionalism in cognitive science assumes there’s a nice “functional level” sitting between biology and behavior (what Marr called the “algorithmic level”). But there’s no guarantee such a level exists, and the neuroscience Cao surveys gives us reason to worry.
The Payoff: What This Means for Functionalism
Cao doesn’t reject functionalism. Instead she argues for what she calls a “chauvinist” functionalism — one that takes the biological details seriously and accepts that the kind of causal structure needed for a mind like ours can probably only be satisfied by systems made of biologically similar stuff.
The irony she highlights: functionalism was supposed to liberate the study of mind from biological details, but if you take it seriously — if you really follow through on what it means for functional roles to constrain their realizers — it points you right back at biology. In one direction it pushes toward something like behaviorism (just match the behavior, who cares how), and in the other toward something like the identity theory (mental states just are brain states). The clean middle ground that made functionalism so appealing may be harder to occupy than philosophers assumed.
That said, Cao ends on a concessive note: for narrowly defined cognitive capacities in specific domains (think: a neural network that replicates one particular perceptual ability), multiple realizability might well hold. The skepticism is about the whole package — the idea that an entire mind, with all its integrated capacities, could be ripped out of biology and implemented in something fundamentally different.
Key Terms to Keep in Your Head
Multiple realizability (MR): Same mental state, different physical stuff. The core claim Cao is challenging.
Functionalism: Mental states are defined by their functional roles (what they do), not what they’re made of.
Neuron doctrine: The assumption that the only functionally relevant brain activity is neuronal electrical signaling. Cao thinks this is wrong.
Generative entrenchment: In evolved systems, components become deeply relied-upon for many purposes over time, making them very hard to swap out.
Multiplexing: One physical component serving many functional roles simultaneously.
Nomological possibility: What’s possible given the actual laws of physics (as opposed to what’s logically or metaphysically conceivable). Cao cares about this rather than abstract conceivability.