Consider what happens when you step outside after a summer storm. The petrichor hits you before you consciously register it—a complex cocktail of plant oils, bacterial spores, and ozone that your brain instantly files under "rain" along with a cascade of associated memories, moods, and predictions about whether you need an umbrella. This entire experience, from nostril to neural pathway, takes perhaps 200 milliseconds. No large language model will ever have it.
This is not a limitation that more training data or larger parameter counts will solve. It is architectural. Today's most capable AI systems are, at their core, extraordinarily sophisticated pattern-matchers trained on text—billions of words describing human experience without any mechanism for having experiences themselves. They can discuss petrichor eloquently, cite its chemical composition, even generate poetry about it. What they cannot do is know what rain smells like.
The symbol grounding problem, revisited
Philosophers have worried about this since the 1980s, when Stevan Harnad posed what he called the "symbol grounding problem": how do abstract symbols acquire meaning? For humans, the answer involves bodies. The word "heavy" connects to memories of lifting things, of tired muscles, of gravity's insistent pull. For a language model, "heavy" connects only to other words—contexts where "heavy" appeared, statistical relationships with "weight" and "burden" and "dense."
This creates a peculiar kind of competence. Language models can pass medical licensing exams while having no conception of pain. They can write restaurant reviews without ever having tasted food. They can discuss heartbreak with apparent sensitivity while possessing nothing that could break. The performance is remarkable; the understanding is, in a meaningful sense, absent.
Where disembodiment shows
The gaps emerge in unexpected places. Ask a language model to estimate how long it takes to walk across a room and it may give you a reasonable answer—not because it understands walking, but because it has encountered similar questions in training text. Ask it something slightly novel about physical experience, and the facade cracks. Models routinely struggle with questions about spatial relationships, physical causation, and the felt qualities of sensation that philosophers call qualia.
More subtly, disembodiment shapes what models find interesting, important, or worth mentioning. Human cognition is fundamentally organized around a body that needs food, fears injury, seeks comfort, and will eventually die. These facts structure our attention, our metaphors, our entire relationship with meaning. A system without survival instincts or mortality has no such organizing principle—just statistical regularities in text.
Our take
None of this makes language models useless; they are genuinely transformative tools for tasks involving text manipulation, information synthesis, and pattern recognition. But the breathless discourse around artificial general intelligence tends to skip past a foundational question: what kind of intelligence can exist without embodiment? The answer may be "a very useful kind" rather than "the human kind." Understanding this distinction is not pessimism about AI—it is clarity about what we are building and what we are not.
