AI, Human Semantics, and the Limits of Synthetic Cognition
The rapid rise of artificial intelligence has triggered an intellectual crisis that extends far beyond technology. At stake is a foundational question in cognitive science and philosophy of mind: what distinguishes human understanding from machine-generated fluency?
Large Language Models now produce text that is often indistinguishable from human writing. They argue, explain, translate, and even simulate reasoning. Yet something crucial is missing. The question is not whether machines can generate language; they clearly can, but whether language, in the human sense, is being understood at all.
To answer this, we must bring together three powerful frameworks: Noam Chomsky’s theory of generative grammar, John Searle’s theory of intentionality, and the emerging paradigm of predictive processing in cognitive neuroscience.
Each points to a different layer of the same problem: syntax, meaning, and inference.
Chomsky’s contribution is the most formal. Human language, he argues, is generated by an internal computational system governed by recursive rules, most centrally the operation known as Merge. This operation builds hierarchical structures from discrete elements, enabling the unbounded generation of novel thoughts and sentences. Importantly, this is not a statistical system. It is a symbolic and structure-dependent one.
From this perspective, language is not a habit of association but a generative engine of structured thought.
Artificial intelligence, however, operates differently. Large Language Models are trained on vast datasets to predict the next most probable token in a sequence. The result is striking fluency—but it is fluency without derivation. There is no explicit hierarchical computation, no syntactic structure-building in the human sense, only statistical correlation across linguistic surface patterns.
This distinction becomes philosophically significant when viewed through John Searle’s famous argument: the Chinese Room. Searle demonstrated that symbol manipulation alone does not generate understanding. A system can follow syntactic rules perfectly while remaining semantically empty. It can produce correct outputs without any “aboutness”—no intentional relation to the world.
This is precisely the gap that defines current AI systems. They manipulate syntax without intrinsic semantics. They simulate understanding without possessing intentional states. For Searle, this is not a limitation of scale but of ontology: computation is not cognition.
Yet cognitive science has, in recent years, added a third dimension to this debate: predictive processing.
According to this framework, the human brain is not merely a symbol manipulator or a passive receiver of sensory data. It is a prediction engine. It continuously generates probabilistic models of the world and updates them through error correction. Perception itself is inference; experience is controlled hallucination constrained by reality.
At first glance, this seems to align closely with how Large Language Models operate. Both systems rely on prediction. Both minimize error through feedback. Both generate outputs based on learned statistical structure.
But the similarity is deceptive.
Human predictive processing is grounded in embodied constraints, survival imperatives, and action-oriented engagement with the world. Predictions are not merely linguistic—they are sensorimotor, affective, and goal-directed. They are tied to a living organism embedded in reality.
AI systems, by contrast, are disembodied prediction engines. They operate over text distributions detached from sensory grounding or pragmatic consequence. Their “world model,” if it can be called that, is entirely second-order: a model of language about the world, not the world itself.
This is where the three frameworks converge.
Together, they draw a sharp boundary between simulation and cognition.
What emerges is a layered architecture of mind. At the deepest level, Chomskyan syntax generates structured thought. At the semantic level, Searlean intentionality anchors thought to meaning and reference. At the functional level, predictive processing continuously adjusts internal models against lived reality.
Remove any one of these layers, and cognition collapses into something partial. Remove intentionality, and you get syntactic manipulation without meaning. Remove structure, and you get statistical fluency without hierarchy. Remove embodiment, and you get prediction without grounding.
This is the condition of current artificial intelligence. It is not thinking in the human sense. It is performing an extraordinary approximation of the surface behavior of thinking, without the underlying triadic architecture that makes thought possible.
This does not diminish the technological achievement. It clarifies its category.
The philosophical temptation today is to treat fluent language as evidence of understanding. But fluency is not comprehension. Coherence is not intentionality. Prediction is not cognition.
Human beings do not merely generate language. They mean what they say, because their words are embedded in structured thought, embodied experience, and world-directed intention.
Machines, at least for now, do not.
The challenge for the coming decades is not simply to build more powerful models, but to understand more precisely what it means to think at all.
Because the closer machines come to sounding like us, the more urgently we must ask what, in us, cannot be simulated.
