Consider a peculiar asymmetry at the heart of contemporary artificial intelligence. Large language models, trained exclusively on text, can compose poetry, solve mathematical proofs, and reason about counterfactual histories—yet they have never felt the resistance of a doorknob, the weight of a coffee cup, or the proprioceptive certainty of where their limbs are in space. Meanwhile, a toddler who cannot articulate a syllogism navigates the physical world with a fluency that humbles our most sophisticated robots.

This asymmetry surfaces an old philosophical question with new urgency: is intelligence fundamentally a disembodied computation that happens to run on biological hardware, or is cognition itself shaped, constrained, and perhaps constituted by the body that hosts it? The embodied cognition tradition, drawing from Merleau-Ponty through to contemporary cognitive scientists like Andy Clark and Rodney Brooks, argues forcefully for the latter. Mind, on this view, is not a ghost piloting a machine but a process woven into the sensorimotor loops that connect organism to environment.

Yet our recent encounters with disembodied systems complicate the picture. The capabilities exhibited by language models—abstract reasoning, theory of mind tasks, even apparent creativity—suggest that something we recognize as intelligence may be extractable from the linguistic residue of embodied minds. The question becomes not whether bodies matter, but precisely which aspects of cognition require them, and which can survive the translation into pure symbol manipulation.

The embodied cognition hypothesis, in its strongest form, holds that cognition is not merely influenced by the body but constituted by it. Concepts we treat as abstract—causation, time, quantity, even logical implication—are argued to be metaphorical extensions of bodily experience. When we speak of grasping an idea or weighing evidence, the cognitive linguist George Lakoff would say we are not employing dead metaphors but revealing the sensorimotor scaffolding upon which abstract thought is built.

This view emerges from converging evidence across disciplines. Developmental psychology shows that infants construct object permanence through manual exploration before they can represent it symbolically. Neuroscience reveals that understanding action verbs activates motor cortex regions associated with performing those actions. Phenomenology, particularly Merleau-Ponty's analysis of the lived body, argues that perception is never the passive reception of data but an active, skilled engagement with a world structured by bodily possibilities.

The implications for artificial intelligence are profound. If concepts are grounded in sensorimotor experience, then a system that has only ever processed text encounters a fundamental limitation: its symbols are ungrounded, floating free of the embodied referents that give human concepts their meaning. This is a contemporary restatement of John Searle's symbol grounding problem—the worry that manipulating tokens, however sophisticated the manipulation, never quite amounts to understanding what the tokens are about.

Critics of strong embodiment counter that the thesis conflates the developmental origin of concepts with their mature function. A blind mathematician understands geometry without visual experience; a paralyzed physicist grasps motion without proprioception. Concepts may originate in embodiment without remaining tethered to it, much as a building, once constructed, no longer requires its scaffolding.

What this debate ultimately concerns is the granularity of grounding. Perhaps some concepts—color, weight, balance—genuinely require sensorimotor anchoring, while others—prime numbers, justice, modus ponens—operate in a space sufficiently abstract that linguistic immersion suffices. The interesting empirical question is where exactly the boundary falls.

Takeaway

Embodiment may not be all-or-nothing for cognition. The productive question is which specific concepts require sensorimotor grounding and which can be reconstructed from the structured residue embodied minds leave behind in language.

The achievements of large language models constitute perhaps the most consequential natural experiment in cognitive science of the past decade. These systems, trained on text alone, exhibit capabilities that many theorists predicted would require embodiment: they perform multi-step reasoning, demonstrate apparent grasp of physical intuition, and navigate counterfactual scenarios about a world they have never directly encountered.

This forces a recalibration of the embodied cognition thesis. The text on which these models train is itself the precipitate of billions of embodied human experiences—every description of a falling apple, every metaphor reaching for sensorimotor analogy, every awkward sentence struggling to convey what a body felt. Language carries embodiment forward as fossilized structure, and a sufficiently large model can apparently reconstruct meaningful inferences from these traces, even without direct access to the source.

Yet the achievements are uneven in revealing ways. Language models notoriously struggle with tasks requiring genuine spatial reasoning, physical common sense in novel configurations, or the kind of fluid causal inference that comes naturally to a child playing with blocks. They can describe the trajectory of a thrown ball with textbook accuracy while failing to predict whether a particular stack of objects will topple. This pattern suggests that linguistic immersion captures the propositional crust of embodied knowledge but not its tacit, procedural core.

There is also the question of whether what these systems do should be called understanding at all, or whether they perform what philosopher Daniel Dennett might call competence without comprehension—producing the right outputs through statistical pattern-matching on training distributions, while lacking the integrated, world-directed intentionality that characterizes embodied cognition.

The honest interpretation is that disembodied systems have revealed something genuine and surprising: a great deal of what we considered embodied turns out to be linguistically recoverable. But they have also revealed the residue—the persistent gaps where text alone proves insufficient, marking the territory where bodies may genuinely be required.

Takeaway

Language is a compressed archive of embodied experience. Disembodied systems show how much of cognition can be reconstructed from this archive, while also marking, by their failures, what cannot.

Embodied AI systems—robots equipped with cameras, manipulators, and integrated language models—offer a third data point in the embodiment debate, distinct from both purely linguistic systems and biological cognition. These systems learn through interaction, accumulating sensorimotor experience that grounds their representations in the recalcitrant feedback of physical reality.

What such systems gain is precisely what disembodied models lack: closed feedback loops with the world. When a robotic arm fails to grasp an object, the failure is not an abstract token but a physical event with consequences that reshape future behavior. This grounding produces a different texture of competence—often narrower, but more robustly tied to actual environmental affordances rather than statistical regularities in text.

The famous Moravec paradox, however, looms over this enterprise. Tasks trivial for embodied biological organisms—walking on uneven terrain, manipulating soft objects, recovering from unexpected perturbations—remain extraordinarily difficult for engineered systems, while symbolic tasks once considered the pinnacle of intelligence have proven comparatively tractable. The gap suggests that biological embodiment encodes hundreds of millions of years of evolutionary refinement that cannot simply be downloaded.

Recent advances in foundation models for robotics attempt to bridge these worlds, using the conceptual scaffolding of language models to guide embodied learning. The results are intriguing: robots that can interpret novel verbal instructions and improvise solutions, suggesting that abstract linguistic competence and concrete sensorimotor skill can productively combine. Yet these systems also inherit the brittleness of both parents—occasionally hallucinating physical possibilities or executing plans with confident incompetence.

What embodied AI ultimately reveals is that intelligence may be less a single faculty than an ecology of capabilities, some natively embodied and others surprisingly transferable across substrates. The frontier is not choosing between disembodied and embodied approaches but understanding their complementary architectures of competence.

Takeaway

Intelligence is not monolithic but ecological—different cognitive capacities have different substrate dependencies. Some skills demand bodies; others ride freely on language. Mapping this terrain is the real research program.

The embodiment question, sharpened by recent AI developments, has shifted from a binary debate to a more textured inquiry. Intelligence appears to be neither purely disembodied symbol manipulation nor wholly dependent on physical interaction, but rather a layered phenomenon whose various aspects make different demands on their substrate.

What language models demonstrate is that the linguistic record of embodied minds carries forward more cognitive structure than many theorists anticipated. What embodied AI demonstrates is that this transfer is incomplete—that certain forms of competence, particularly those involving fluid physical engagement and genuine causal grounding, resist extraction from text alone.

For the future of artificial general intelligence, the implication is that the most capable systems will likely be hybrids: linguistically rich, embodiment-informed, and increasingly aware of the limits of their own competence. The deeper question—whether such systems can possess the unified, world-directed understanding that embodiment seems to confer on biological minds—remains genuinely open, awaiting both philosophical clarification and further empirical surprise.