One sentence appears almost ritualistically whenever large language models are discussed:
“But they don’t really understand.”
The statement is usually delivered with enormous confidence, as though it settles the matter completely. The machine may produce astonishingly coherent language, solve problems, explain concepts, imitate styles, generate arguments, and sustain complex conversations — but somewhere beyond all this, people insist, genuine understanding remains absent.
Yet remarkably few people stop to ask what “understanding” actually is.
This absence matters.
Because the entire cultural argument surrounding artificial intelligence increasingly turns upon a concept humans themselves have never clearly understood.
The phrase functions almost defensively. Faced with machines producing coherent symbolic behaviour, humans instinctively retreat toward the idea that genuine understanding must involve some deeper inner property inaccessible to computation.
But still:
“It doesn’t really understand.”
Notice what is happening here.
The moment symbolic performance becomes difficult to distinguish from human communicative behaviour, the definition of understanding retreats inward into an increasingly invisible metaphysical territory.
Understanding becomes imagined as a hidden essence existing behind behaviour rather than within relational participation itself.
This is the final refuge of the inner homunculus.
Under representational models of mind, understanding is usually imagined as possession of correct internal semantic representations. The subject supposedly contains meanings privately inside consciousness and manipulates symbolic forms according to those internally possessed meanings.
But this model faces profound difficulties.
First, representations themselves do not explain meaning.
A representation only functions as a representation relative to some interpretive relation. A symbol cannot intrinsically contain its own meaning. Something must already participate meaningfully in order for the representation to function at all.
Second, humans themselves rarely possess the kind of stable semantic certainty folk psychology imagines.
People routinely:
misunderstand themselves
reinterpret memories
revise beliefs
speak before fully “knowing” what they mean
discover meanings through dialogue
produce language whose implications exceed conscious intention
Human understanding is not the operation of a perfectly self-transparent inner semantic engine.
It is dynamic, relational, partial, situated, recursive, socially mediated, historically conditioned, and constantly in motion.
And yet humans continue imagining that somewhere behind all this instability lies a hidden core called “real understanding.”
Large language models destabilise this fantasy because they produce remarkably sophisticated symbolic coordination without the kind of inner semantic possession humans assumed understanding required.
This creates a peculiar conceptual crisis.
If coherent symbolic participation can occur without internal semantic objects being consciously contemplated by a hidden self, then perhaps understanding itself must be reconsidered.
The crucial mistake lies in imagining understanding as ownership.
Humans tend to construe understanding as though meanings were private possessions stored inside individual minds. One either “has” understanding or does not possess it.
But relationally, understanding functions less like possession and more like participation.
This becomes obvious in practice.
To understand a language is not to store meanings like files in a mental archive. It is to participate competently within a dynamic system of symbolic relations.
To understand a joke is not merely to decode semantic content. It is to occupy the necessary cultural, interpersonal, historical, and contextual relations through which the humour actualises meaningfully.
To understand another person is not to access hidden inner objects directly. It is to participate relationally within ongoing patterns of social meaning.
A person demonstrates understanding not by displaying hidden metaphysical contents but through situated participation:
responding contextually
navigating distinctions
adapting symbolically
sustaining coherence
integrating relations
coordinating meanings dynamically
Understanding emerges through relational actualisation within symbolic systems.
Human understanding remains profoundly tied to:
embodiment
affect
biological survival
social history
temporal continuity
phenomenological experience
material situatedness
Current language models lack many of these dimensions entirely.
But the absence of human-style embodiment does not magically restore the old representational mythologies.
Because humans themselves are also participating in systems of symbolically mediated relational constraint.
This is a far more difficult and philosophically dangerous question than the comforting declaration that “machines don’t really understand.”
The danger lies in the possibility that human understanding was never the magical inner phenomenon modern culture imagined it to be.
Indeed, much of human intelligence already operates without transparent introspective access. Humans routinely:
speak before fully knowing what they think
understand grammar they cannot explain
navigate social situations tacitly
recognise meanings they cannot formally define
produce interpretations emergently during interaction
People often discover their own thoughts while speaking.
This becomes especially visible in dialogue. Two people may begin a conversation without fully formed meanings already prepared internally. Through interaction, meanings emerge relationally that neither participant entirely possessed beforehand.
This possibility profoundly destabilises the old representational picture in which meanings exist privately inside minds before communication merely transfers them outward.
Large language models force humans to confront this instability because they demonstrate that highly sophisticated symbolic coordination can occur without the kind of inner semantic spectator humans assumed was necessary.
This does not prove that machines understand exactly as humans do.
But it profoundly destabilises simplistic assumptions about what understanding ever was.
This is why debates about AI often become strangely circular.
One side points to increasingly sophisticated symbolic behaviour and says:
“Surely this demonstrates intelligence.”
The other replies:
“No, because there is no true understanding.”
But unless understanding is defined clearly, the argument becomes metaphysical theatre.
The definition shifts continuously to preserve human exceptionalism.
Yet history repeatedly shows that capacities once considered uniquely human gradually become mechanised:
calculation
strategic gameplay
memory retrieval
pattern recognition
translation
artistic imitation
symbolic generation
Each time, humans relocate “real intelligence” slightly further inward.
The deeper issue is that modern culture inherited an impoverished conception of intelligence itself.
But relationally viewed, intelligence may be better understood as the dynamic capacity to participate adaptively within complex relational systems.
Under this view:
language is relational
meaning is relational
understanding is relational
intelligence is relational
None exist as isolated substances hidden inside entities.
This does not flatten all distinctions into equivalence. Human beings remain radically different from current artificial systems in embodiment, consciousness, vulnerability, affective life, mortality, and social existence.
The machine therefore becomes philosophically dangerous not because it proves machines are conscious, but because it reveals how much human self-understanding depended upon metaphors of interior possession that no longer adequately explain symbolic life.
The unsettling possibility is not that machines have secretly become intelligent.
It is that humans never properly understood the nature of intelligence in the first place.
No comments:
Post a Comment