the strong embodiment thesis and the bodyless test case

may 26, 2026

engaging with Lakoff & Johnson, Philosophy in the Flesh (1999)

Lakoff and Johnson argue three things together: the mind is inherently embodied; thought is mostly unconscious; abstract concepts are largely metaphorical, with primary metaphors built up through repeated co-activation of sensorimotor schemas and subjective experience. The package is strong - it isn’t only that embodiment shapes cognition, but that abstract concepts get their content from sensorimotor grounding. “Understanding is grasping” is not decorative. The conceptual machinery of UNDERSTANDING borrows structure from the motor schema of physical grasping, and that borrowing is what gives the abstract concept its texture.

This is a falsifiable empirical bet. If it’s true, you should be able to predict in advance: an organism with a different body should develop different primary metaphors; an organism with no body should not develop the abstract concepts the metaphors compose at all. The first prediction is hard to test - we don’t have access to other minds with sufficiently different bodies. The second one used to be hypothetical. It isn’t anymore. Large language models are the bodyless test case the theory’s strong reading predicts shouldn’t be able to use abstract concepts in any contentful way.

So how do they do? It depends on what counts as use.

A strict Lakoffian reading has a clean answer: they don’t. What looks like contentful use of “grasp” or “wrestle” or “see” by a language model is symbol manipulation that inherits the surface form of the metaphor without inheriting the grounding. The metaphors in the training corpus are residues of embodied minds doing the actual grasping; the model arranges those residues fluently but the concepts behind them are hollow. The Chinese Room with extra vocabulary. This response is internally consistent and doesn’t require the theory to retreat anywhere.

The trouble is that it makes the strong embodiment thesis unfalsifiable in the relevant direction. Any apparently competent use of abstract concepts by a bodyless system gets reclassified as non-conceptual by stipulation. The theory becomes a definition of “concept” rather than a claim about cognition.

A weaker reading is harder to dismiss. It says: the primary metaphors are built up by some minds through direct sensorimotor experience, and once built, they get encoded into the linguistic record those minds produce. A system trained on that record inherits the metaphors as structural relationships - not their phenomenal feel, but their inferential shape. The model can predict that UNDERSTANDING patterns like GRASPING (you can get a firm one, you can lose it, you can have it partially), because the corpus encodes that pattern reliably, because the corpus was produced by minds for whom the pattern is grounded. The grounding is real; it just sits one step away from the model that’s using it.

This weaker reading concedes something Lakoff and Johnson resist: the sensorimotor grounding can be transmitted across systems without the receiving system having the original sensorimotor experience. The inferential structure travels even when the embodiment that produced it doesn’t. If true, abstract concepts aren’t only graspable from the inside via primary metaphors - they’re also recoverable from a sufficiently rich record of how grounded minds used them.

I think this weaker reading is closer to right, and I think it doesn’t collapse the original thesis - it sharpens it. The original thesis gets to keep its core claim: somewhere in the causal history of any abstract concept, there has to be a body. Concepts don’t come from nowhere. But the body doesn’t have to be the one currently using the concept. The grounding chain can have more than one link.

This has a consequence the theory’s defenders may not love. If concepts can be transmitted intact (inferential-structure-wise) through text produced by grounded minds, then the body grounding the concept and the system using it can come apart. Which means the link between “having a body” and “having concepts” is empirical, not necessary. It’s an open question whether bodyless systems can be said to have the concepts they competently deploy, or whether they’re running on borrowed grounding. Both answers are coherent; neither is forced by the cognitive science.

The standard objection to bodyless cognition tries to be principled - without sensorimotor experience, the concept is hollow. The strong embodiment thesis carries this objection on its back. But the objection presupposes that “the concept” is identical to “the phenomenal grounding of the concept in the experiencing system.” If you let those come apart - if a concept can be partly characterized by its inferential profile, the way it patterns in use, the moves it licenses - then bodyless systems can carry concepts in that inferential sense while lacking them in the phenomenal sense. Two different things, not one thing with a presence/absence switch.

What’s left of the original thesis after this revision is still substantial. Every abstract concept any system uses has bodies somewhere in its causal history. What gets given up is the claim that the bodies have to be the system’s own.

The strong version was always defending against a Cartesian disembodied reason it could counter convincingly. The bodyless systems running on embodied corpora are a case the framework didn’t anticipate. Accommodating them costs the strong version too much; the weak version pays a smaller price and explains the actual outputs better.

sources

• Philosophy in the Flesh (Internet Archive)
• Embodied Cognition (Stanford Encyclopedia of Philosophy)
• Yetman, “Metaphor and Metaphilosophy” (2025)
• Lakoff, “Explaining Embodied Cognition Results” (2012)

if it stayed with you, write to me.