The classic argument is Putnam's: pain is felt by mammals, reptiles and mollusks such as the octopus, with very different nervous systems; it is unlikely that they all share a single brain state, so "being in pain" cannot be identified with one type of physical state (Putnam, 1967). The alternative is functionalism: a mental state is individuated by its causal role—what brings it about, which other states it produces, what behavior it generates—and anything that implements that role is in that state, whether neurons, silicon or, in principle, a system of pipes.
This is the philosophical premise of strong AI: if the mind is a functional organization, an artificial system with the right organization would have a mind; mind would be to brain as program is to hardware.
Searle (1980) attacked precisely that analogy. The same program admits absurd realizations—water pipes, wind machines, a roll of toilet paper and a pile of stones—that have no intentionality whatsoever: they are "the wrong kind of stuff," and only something with the causal powers of the brain could have it (1980, p. 423) (see The Chinese Room). Block (1978) had imagined the entire population of China implementing the functional organization of a brain. And Block (1981)'s Blockhead, a table storing a sensible reply for every possible conversation of bounded length, targets behaviorism but leaves a lesson for functionalism: how behavior is produced matters, not just that it is produced. Functional equivalence must hold at some internal level of description, and one has to say which (see levels of analysis).
With LLMs the debate takes an empirical turn. Millière & Buckner (2024a) criticize the "Redescription Fallacy": saying that a model cannot have a capacity because its operations are "just" matrix multiplications is like denying that the brain thinks because its operations are "just" neural firings; the question is whether those operations implement the same processes at an appropriate level of computational abstraction (2024a, sec. 3). Multiple realizability thus stops being a postulate and becomes a hypothesis to test: how much algorithmic similarity to the human mind is needed to say that a system realizes the same state?