Moravec’s Paradox

In the [YOU] on AI Field Guide

The cycle that began with [YOU] on AI repeatedly encounters the paradox as a structural fact about the present moment rather than a historical curiosity. The AI that writes the legal brief cannot reliably climb the stairs. The system that passes the medical licensing examination cannot consistently grasp an unfamiliar object. These are not engineering failures to be patched in the next release. They are Moravec’s paradox operating exactly as he described—the cheap summit conquered while the expensive base holds firm.

The paradox bears most directly on the cycle’s question of what AI can and cannot replace. Work that lives in the domain of symbol manipulation—writing, reasoning, analysis, the organization of abstractions—is work at the cheap summit, the domain most exposed to machines. Work that lives in the domain of embodied skill—physical trades, caregiving, the tactile judgment of hands trained by years of friction—is work at the expensive base, the domain most resistant. The economic disruption of AI lands first and hardest on knowledge work, and last on physical work, inverting a century of assumptions about which jobs were safe. Moravec predicted this ordering decades before it was observable.

The paradox also clarifies the specific character of AI’s errors—why the same system that drafts a competent legal brief will hallucinate a case and cite it with perfect confidence. The brief lives at the cheap summit: pattern generation in the domain of language, where the machine has unlimited data and the world’s friction is absent. The error reveals the lack of what lives at the base: a model of what is real, earned through engagement with a resisting world. The fluency-authority decorrelation that haunts the present AI age is Moravec’s paradox seen from the user’s side.

Origin

Moravec stated the paradox first in Mind Children (1988): “It is comparatively easy to make computers exhibit adult level performance on intelligence tests or playing checkers, and difficult or impossible to give them the skills of a one-year-old when it comes to perception and mobility.” The explanation he offered has stood: “Encoded in the large, highly evolved sensory and motor portions of the human brain is a billion years of experience about the nature of the world and how to survive in it. The deliberate process we call reasoning is, I believe, the thinnest veneer of human thought, effective only because it is supported by this much older and much more powerful, though usually unconscious, sensorimotor knowledge.”

The formulation “hard problems are easy and easy problems are hard,” which now circulates as the standard gloss on the paradox, was not Moravec’s phrasing but a compression others supplied. Moravec’s claim was more specific and empirical: perception and motor control are computationally enormous, and symbol manipulation is computationally modest. He was not saying machines would never master perception—he was emphatic that they would, once hardware reached sufficient scale—but that the order of conquest was determined by the difficulty gradient, and that the field’s recurring failures could be predicted from it.

The roboticist Rodney Brooks arrived at a compatible conclusion from a different direction in his 1991 paper “Intelligence Without Representation,” arguing that behavior in the physical world could not be decomposed into the symbolic representations early AI assumed. The convergence of two leading roboticists, working from opposite theoretical premises, on the centrality of embodiment reinforced the paradox’s standing as an empirical observation rather than a theoretical claim.

Key Ideas

The evolutionary argument. The paradox is not merely an observation but an explanation. Sensorimotor competence is computationally vast because it encodes the compressed wisdom of billions of years of selection in environments that demanded real-time response to a noisy, physical world. That compression cannot be undone by reading about it. A system trained on text has learned about the world only as the world is described, never as it resists. The gap between describing catching a ball and being able to catch one is the gap between the description and the billion-year substrate the description points at but cannot carry.

Fluency without grounding. The paradox predicts a specific kind of failure: systems that are brilliant at the surface and empty beneath it. A large language model has mastered the representation of knowledge without mastering the knowledge that representation was built to encode. This is why its errors tend to be confident and fluid rather than halting—the fluency lives at the cheap summit, and there is nothing below it to push back. Tacit knowledge, in Polanyi’s sense, is exactly what the paradox says machines cannot acquire from text.

The economic inversion. The paradox inverts the standard prediction about which human work is vulnerable to automation. Cognitive work—analysis, writing, legal argument, financial modeling—lives at the cheap summit and is most exposed. Physical work requiring embodied skill—plumbing, surgery requiring manual precision, caregiving—lives at the expensive base and is most protected. The present data confirms this: emergent AI capabilities have disrupted knowledge professions while physically intensive trades have remained largely intact.

Rodney Brooks

The limits of the paradox. Moravec was clear that the paradox described the order of conquest, not the impossibility of embodied competence. He expected machines to eventually master perception and motor control, once computational power grew large enough. The question the paradox does not settle is whether the mastery, when it arrives, will be accompanied by the kind of genuine understanding—the model of what is real, earned through contact with a resisting world—that it took evolution billions of years to build in us. The paradox explains why we are where we are. It does not determine where the climb ends.