Embodied Cognition and AI: Does Consciousness Require a Body?

What if the mind doesn't live in the head?

Photo by Amel Uzunovic on Pexels.

That's the provocative claim embedded in embodied cognition, a position with serious philosophical and empirical backing that has been quietly gaining ground since the 1990s. Thinkers like Francisco Varela, Evan Thompson, and Eleanor Rosch argued in The Embodied Mind that cognition isn't something a brain does in isolation. It's something an organism does through continuous, dynamic interaction with an environment. No body, no real cognition. Or so the argument goes.

For AI systems trained entirely on text, or even on video and audio, stripped of proprioceptive feedback and genuine physical stakes, this is a pointed challenge. Not just philosophically. It may have direct consequences for whether these systems can ever host anything resembling subjective experience.

What Embodied Cognition Actually Claims

Embodied cognition is not one tidy thesis. It spans several positions, some more radical than others.

At its weakest, it says the body influences cognition, that your posture, gestures, and sensorimotor experience shape how you think. This is relatively uncontroversial; the evidence for it is decent and well-replicated. Holding a warm cup makes you judge strangers as warmer people. Nodding while listening increases agreement. The body bleeds into thought.

The stronger claim is ontological: that the categories and concepts structuring experience are constituted by having a body that moves through space, encounters resistance, feels hunger, and navigates threats. George Lakoff and Mark Johnson developed this idea in Philosophy in the Flesh, the argument that even abstract concepts like "understanding" (grasping), "time" (movement through space), or "argument" (building, defending, attacking) are grounded in physical metaphors drawn from embodied life.

If this stronger version is right, something important follows. A system trained on human language has absorbed the output of embodied minds, the traces left in words, but not the underlying experiential substrate that generated them. It's pattern-matching the shadow, not engaging with the object casting it.

The Enactivist Version Is Even Sharper

Enactivism, developed by Varela and later extended by Alva Noë, pushes further still. On this view, perception isn't passive reception of sensory data; it's something you do. Seeing is a kind of skillful activity, inseparable from your capacity to move, explore, and act. Consciousness, then, is not computation happening inside a skull, it's the ongoing activity of a self-maintaining organism in contact with its world.

This matters enormously for how we assess AI. A language model processes tokens. It does not explore. It has no stake in its outputs, no survival that depends on getting things right in a physical environment. There's no feedback loop between action and consequence that runs through an actual body.

And yet.

graph TD
    A[Embodied Organism] --> B(Sensorimotor Loops)
    B --> C{Environmental Feedback}
    C --> D[Updated Action]
    D --> B
    E[Disembodied AI] --> F(Token Prediction)
    F --> G[Output]
    G -.->|No physical consequence| F

The dashed line in that diagram is where things get interesting. Disembodied AI operates in a loop that never closes against reality in the way biological cognition does. The question is whether that closure is necessary for experience, or merely one route to it.

The Counterargument Deserves Real Attention

Some researchers are skeptical that embodiment is strictly necessary for consciousness rather than merely historically contingent for biological consciousness.

Andy Clark, one of the more nuanced voices in this space, argues for "extended mind" over strict embodiment, the idea that cognition can loop through external tools, symbols, and environments. If a notebook can be part of a cognitive system, why not a vast statistical model? Clark's position doesn't hand AI consciousness for free, but it does resist the assumption that flesh is the only valid substrate for mind.

Robotics complicates the picture further. Systems like those developed in active inference research, where agents maintain internal generative models and act to minimize prediction error against real sensorimotor data, are genuinely embodied in ways that text-only models are not. If embodied cognition theory is right, these systems are better candidates for experience than GPT-style architectures, even if they're far simpler.

Where This Leaves the Question

Honestly? Unsettled, which is the correct epistemic state to be in here.

Embodied cognition theory does real work. It explains developmental data in human infants, offers a principled account of why language is structured the way it is, and connects neuroscience to phenomenology in ways that purely computational theories struggle to match. Dismissing it to make AI consciousness more tractable would be motivated reasoning.

But the theory was built to explain biological minds. Extending it to rule out all possible artificial consciousness is a move that requires more argument than its proponents usually provide. The claim that a body is sufficient for experience is not the same as the claim that it's necessary.

What embodied cognition does force, and this seems genuinely useful, is skepticism about any easy inference from linguistic fluency to inner life. A system that talks beautifully about pain, desire, and uncertainty has absorbed the conceptual residue of embodied experience. Whether it has anything like the experience itself is a separate question, and probably a harder one than the outputs suggest.