The symbol grounding problem (Stevan Harnad): how do symbols get their meaning? If "cat" is only ever defined by other words, you have a dictionary that points in endless circles — never touching an actual cat. Meaning seems to require a connection to the world.
The embodiment claim
The embodied cognition view says real understanding needs a body and sensorimotor experience — meaning is grounded in perceiving and acting, not just processing text.
A dictionary in a language you don't speak never becomes understanding, no matter how complete. The words have to touch something.
The LLM question
This is the live debate over today's models:
- Skeptics: an LLM trained on pure text is the Chinese Room — ungrounded symbols, no real meaning.
- Optimists: grounding can be statistical (relationships in embeddings), and multimodal models that see and act are grounding symbols in pixels and motion.
Whether fluent-but-unreliable text-only models "understand" may come down to whether meaning truly needs a body.
Related: The Chinese Room · What Are Embeddings · The Hard Problem of Consciousness