Every large language model you interact with has processed billions of words about coffee — its bitterness, its warmth, the way it smells on a cold morning. Yet no AI has ever experienced coffee. This isn't a limitation that better training data or more parameters will solve. It's a structural absence that reveals something important about what these systems actually are.
The gap between describing and experiencing has troubled philosophers for centuries. Thomas Nagel's famous question — what is it like to be a bat? — gets at something similar. But with AI, the question isn't about accessing another creature's subjective experience. It's about whether there's any subjective experience at all. Language models don't have sensory organs. They don't have bodies. They exist as statistical patterns over text, predicting the next likely token based on everything they've processed before.
The grounding problem
Linguists and cognitive scientists call this the "symbol grounding problem." Words get their meaning, at least partly, from being anchored to real-world experiences. When you read "the lemon was sour," your brain activates not just linguistic associations but traces of actual lemon-tasting experiences. Language models have only the words, never the lemons.
This creates predictable blind spots. Ask an AI to describe the difference between silk and velvet, and it will produce plausible text drawn from countless descriptions. But it has no tactile memory to draw upon. Its knowledge is entirely derivative — patterns extracted from what embodied humans have written. The AI is an extraordinarily sophisticated parrot, but a parrot with a library instead of a cage.
Where disembodiment matters
For many tasks, this limitation is irrelevant. Summarizing documents, writing code, answering factual questions — none of these require having tasted coffee. Language models excel precisely because language itself is their native domain.
But the limitation surfaces in unexpected places. AI systems struggle with physical reasoning — understanding that you can't fit a basketball through a coffee cup, or that wet surfaces are slippery. They falter on questions requiring intuitive physics, spatial relationships, or the kind of common sense that comes from navigating a physical world. When an AI confidently explains how to perform a delicate manual task, it's synthesizing instructions without any motor memory of what "delicate" actually feels like.
The implications extend to emotional understanding. Language models can discuss grief eloquently, drawing on humanity's vast literature of loss. But they have never lost anything. Their engagement with emotion is entirely textual — pattern-matching on how humans describe feelings, without the physiological substrate that makes feelings what they are.
Our take
None of this means AI systems aren't useful, even transformatively so. But the hype cycle tends to obscure what these systems fundamentally are: extraordinarily capable text-prediction engines that have never stubbed a toe, smelled rain, or felt afraid. Understanding this distinction matters as we decide what to trust them with. An AI can help you write a eulogy, but it cannot mourn. It can describe a sunset in a thousand ways, but it has never watched one. These aren't failures to be engineered away. They're features of what it means to be a pattern over words rather than a creature in the world.
